Axis Spatial

Geospatial Intelligence Index: details

Geospatial Agent Index

Which AI model is best at geospatial (GIS and Earth observation) work? The Geospatial Intelligence Index combines agent tasks, coding, Earth science and perception into one score for each model, shown at its highest reasoning level.

14 models ranked (22 counting each reasoning level) · 14 in the text-only view · 15 with partial coverage · 9 developers · Geospatial Agent Index · Perception · Methodology

Ongoing evaluation: results update as runs complete. Last updated 13 October 2026.

Geospatial Intelligence Index

How the Geospatial Intelligence Index is built

A weighted average of four groups, scored for every model and reasoning level and shown by default at each model's highest level:

Each model is scored in its developer's own harness where we run one, otherwise in Terminus-2. Claude models are scored in Claude Code, OpenAI models in Codex, Kimi models in Kimi Code once their runs there are complete, and all other models in Terminus-2. This applies to the agents and coding groups; Earth science and perception use one model call per question, with no agent harness. Part of a difference between two models can therefore come from the harness; the Geospatial Agent Index is the place to compare harnesses. The structure is a geospatial counterpart of Artificial Analysis's Intelligence Index. The weights are its weights for the groups that apply to geospatial work (agents 30, coding 20, scientific reasoning 20), with perception added at 20, the weight of its smaller groups, all scaled to 100. Its general-knowledge group is left out, because all the knowledge we test is Earth-science knowledge. Methodology 0.31.

A text-only view (the same index without perception, weights 3/7, 2/7, 2/7) is shown next to the headline.

Geospatial Intelligence Index

Geospatial Intelligence Index

Weighted average of agents, coding, Earth science and perception, 0 to 100 · Higher is better

Geospatial Intelligence IndexClaude Opus 5.5 (xhigh): 81; GPT-6.1 Sol (xhigh): 79; Claude Sonnet 5.5 (xhigh): 78; Claude Opus 5 (high): 77; Claude Haiku 5.5 (xhigh): 76; Claude Opus 4.8 (high): 76; GPT-6 Luna (xhigh): 75; Claude Opus 4.7 (high): 75; Claude Sonnet 5 (high): 74; GLM-5.3-Flash (max): 67; DeepSeek V4 Pro (max): 62; Claude Haiku 4.5 (thinking 32k): 61; GLM-5.2 (none): 57; gpt-oss-120b (low): 41025507510081Claude CodeClaude Opus5.5 (xhigh)79CodexGPT-6.1 Sol(xhigh)78Claude CodeClaude Sonnet5.5 (xhigh)77Claude CodeClaude Opus 5(high)76Claude CodeClaude Haiku5.5 (xhigh)76Claude CodeClaude Opus4.8 (high)75CodexGPT-6 Luna(xhigh)75Claude CodeClaude Opus4.7 (high)74Claude CodeClaude Sonnet5 (high)67Terminus-2GLM-5.3-Flash(max)62Terminus-2DeepSeek V4Pro (max)61Claude CodeClaude Haiku4.5 (thinking32k)57Terminus-2GLM-5.2(none)41Terminus-2gpt-oss-120b(low)
How to read this chart

Rows. Each model once, at its highest reasoning level, labelled with that level, as Artificial Analysis does; when that level is not ranked yet, the highest ranked level, said in its tooltip. Other levels are under 'Show every reasoning level'.

What this metric means. The weighted average of the group scores, each 0 to 100: agents 3/9; coding 2/9; earth science 2/9; perception 2/9. The structure is a geospatial counterpart of Artificial Analysis's Intelligence Index. The weights are its weights for the groups that apply to geospatial work (agents 30, coding 20, scientific reasoning 20), with perception added at 20, the weight of its smaller groups, all scaled to 100. Its general-knowledge group is left out, because all the knowledge we test is Earth-science knowledge.

Which harness. Each model is scored in its developer's own harness where we run one, otherwise in Terminus-2. Claude models are scored in Claude Code, OpenAI models in Codex, Kimi models in Kimi Code once their runs there are complete, and all other models in Terminus-2. This applies to the agents and coding groups; Earth science and perception use one model call per question, with no agent harness. Part of a difference between two models can therefore come from the harness; the Geospatial Agent Index is the place to compare harnesses. The bar shows the harness the agents and coding groups are scored in.

Cannot take images. A model that cannot take images at all, according to its developer's documentation, scores 0 in perception, labelled "cannot take images". The text-only view leaves perception out.

Not shown: partial coverage. 15 models are not yet scored in full in every group. They are listed under partial coverage in the leaderboard, with what is still to come.

Data table: Geospatial Intelligence Index
Geospatial Intelligence Index
RankModelDeveloperGeospatial Intelligence IndexText-only viewAgents (3/9)Coding (2/9)Earth science (2/9)Perception (2/9)Scored in
1Claude Opus 5.5 (xhigh)Anthropic80.887.174999459Claude Code
2GPT-6.1 Sol (xhigh)OpenAI78.785.172979356Codex
3Claude Sonnet 5.5 (xhigh)Anthropic78.184.069969457Claude Code
4Claude Opus 5 (high)Anthropic77.083.870959353Claude Code
5Claude Haiku 5.5 (xhigh)Anthropic76.082.867979252Claude Code
6Claude Opus 4.8 (high)Anthropic75.883.266989350Claude Code
7GPT-6 Luna (xhigh)OpenAI74.781.664989150Codex
8Claude Opus 4.7 (high)Anthropic74.682.365989347Claude Code
9Claude Sonnet 5 (high)Anthropic73.981.665969147Claude Code
10GLM-5.3-Flash (max)Z.ai66.873.454849244Terminus-2
11DeepSeek V4 Pro (max)DeepSeek61.879.55998920 (cannot take images)Terminus-2
12Claude Haiku 4.5 (thinking 32k)Anthropic61.268.951768934Claude Code
13GLM-5.2 (none) (Highest ranked level; GLM-5.2 (max) is not ranked yet)Z.ai56.772.85583890 (cannot take images)Terminus-2
14gpt-oss-120b (low) (Highest ranked level; gpt-oss-120b (high) is not ranked yet)OpenAI40.652.22854880 (cannot take images)Terminus-2
Show every reasoning level

Geospatial Intelligence Index, every reasoning level

Weighted average of agents, coding, Earth science and perception, 0 to 100 · Higher is better

Geospatial Intelligence Index, every reasoning levelClaude Opus 5.5 (xhigh): 81; Claude Opus 5.5 (high): 81; GPT-6.1 Sol (medium): 79; Claude Sonnet 5.5 (high): 79; GPT-6.1 Sol (xhigh): 79; Claude Opus 5.5 (medium): 79; Claude Sonnet 5.5 (xhigh): 78; GPT-6.1 Sol (high): 78; Claude Opus 5 (high): 77; Claude Haiku 5.5 (xhigh): 76; Claude Opus 4.8 (high): 76; Claude Sonnet 5.5 (medium): 75; GPT-6 Luna (xhigh): 75; Claude Opus 4.7 (high): 75; Claude Sonnet 5 (high): 74; Claude Haiku 5.5 (high): 73; Claude Haiku 5.5 (medium): 72; GLM-5.3-Flash (max): 67; DeepSeek V4 Pro (max): 62; Claude Haiku 4.5 (thinking 32k): 61; GLM-5.2 (none): 57; gpt-oss-120b (low): 41025507510081Claude CodeClaude Opus5.5 (xhigh)81Claude CodeClaude Opus5.5 (high)79CodexGPT-6.1 Sol(medium)79Claude CodeClaude Sonnet5.5 (high)79CodexGPT-6.1 Sol(xhigh)79Claude CodeClaude Opus5.5 (medium)78Claude CodeClaude Sonnet5.5 (xhigh)78CodexGPT-6.1 Sol(high)77Claude CodeClaude Opus 5(high)76Claude CodeClaude Haiku5.5 (xhigh)76Claude CodeClaude Opus4.8 (high)75Claude CodeClaude Sonnet5.5 (medium)75CodexGPT-6 Luna(xhigh)75Claude CodeClaude Opus4.7 (high)74Claude CodeClaude Sonnet5 (high)73Claude CodeClaude Haiku5.5 (high)72Claude CodeClaude Haiku5.5 (medium)67Terminus-2GLM-5.3-Flash(max)62Terminus-2DeepSeek V4Pro (max)61Claude CodeClaude Haiku4.5 (thinking32k)57Terminus-2GLM-5.2(none)41Terminus-2gpt-oss-120b(low)
How to read this chart

Rows. One bar per model and reasoning level.

What this metric means. The weighted average of the group scores, each 0 to 100: agents 3/9; coding 2/9; earth science 2/9; perception 2/9. The structure is a geospatial counterpart of Artificial Analysis's Intelligence Index. The weights are its weights for the groups that apply to geospatial work (agents 30, coding 20, scientific reasoning 20), with perception added at 20, the weight of its smaller groups, all scaled to 100. Its general-knowledge group is left out, because all the knowledge we test is Earth-science knowledge.

Which harness. Each model is scored in its developer's own harness where we run one, otherwise in Terminus-2. Claude models are scored in Claude Code, OpenAI models in Codex, Kimi models in Kimi Code once their runs there are complete, and all other models in Terminus-2. This applies to the agents and coding groups; Earth science and perception use one model call per question, with no agent harness. Part of a difference between two models can therefore come from the harness; the Geospatial Agent Index is the place to compare harnesses. The bar shows the harness the agents and coding groups are scored in.

Cannot take images. A model that cannot take images at all, according to its developer's documentation, scores 0 in perception, labelled "cannot take images". The text-only view leaves perception out.

Not shown: partial coverage. 26 models are not yet scored in full in every group. They are listed under partial coverage in the leaderboard, with what is still to come.

Data table: Geospatial Intelligence Index, every reasoning level
Geospatial Intelligence Index, every reasoning level
RankModelDeveloperGeospatial Intelligence IndexText-only viewAgents (3/9)Coding (2/9)Earth science (2/9)Perception (2/9)Scored in
1Claude Opus 5.5 (xhigh)Anthropic80.887.174999459Claude Code
2Claude Opus 5.5 (high)Anthropic80.686.8731009559Claude Code
3GPT-6.1 Sol (medium)OpenAI78.885.8721009255Codex
4Claude Sonnet 5.5 (high)Anthropic78.784.368999459Claude Code
5GPT-6.1 Sol (xhigh)OpenAI78.785.172979356Codex
6Claude Opus 5.5 (medium)Anthropic78.784.871969457Claude Code
7Claude Sonnet 5.5 (xhigh)Anthropic78.184.069969457Claude Code
8GPT-6.1 Sol (high)OpenAI77.885.1701009353Codex
9Claude Opus 5 (high)Anthropic77.083.870959353Claude Code
10Claude Haiku 5.5 (xhigh)Anthropic76.082.867979252Claude Code
11Claude Opus 4.8 (high)Anthropic75.883.266989350Claude Code
12Claude Sonnet 5.5 (medium)Anthropic75.481.063969356Claude Code
13GPT-6 Luna (xhigh)OpenAI74.781.664989150Codex
14Claude Opus 4.7 (high)Anthropic74.682.365989347Claude Code
15Claude Sonnet 5 (high)Anthropic73.981.665969147Claude Code
16Claude Haiku 5.5 (high)Anthropic73.280.566919247Claude Code
17Claude Haiku 5.5 (medium)Anthropic72.480.163949245Claude Code
18GLM-5.3-Flash (max)Z.ai66.873.454849244Terminus-2
19DeepSeek V4 Pro (max)DeepSeek61.879.55998920 (cannot take images)Terminus-2
20Claude Haiku 4.5 (thinking 32k)Anthropic61.268.951768934Claude Code
21GLM-5.2 (none)Z.ai56.772.85583890 (cannot take images)Terminus-2
22gpt-oss-120b (low)OpenAI40.652.22854880 (cannot take images)Terminus-2

Open weights vs proprietary

Geospatial Intelligence Index: open weights vs proprietary

Each model at its highest reasoning level, 0 to 100 · Higher is better

  • Open weights
  • Proprietary
Geospatial Intelligence Index: open weights vs proprietaryClaude Opus 5.5 (xhigh): 81; GPT-6.1 Sol (xhigh): 79; Claude Sonnet 5.5 (xhigh): 78; Claude Opus 5 (high): 77; Claude Haiku 5.5 (xhigh): 76; Claude Opus 4.8 (high): 76; GPT-6 Luna (xhigh): 75; Claude Opus 4.7 (high): 75; Claude Sonnet 5 (high): 74; GLM-5.3-Flash (max): 67; DeepSeek V4 Pro (max): 62; Claude Haiku 4.5 (thinking 32k): 61; GLM-5.2 (none): 57; gpt-oss-120b (low): 41025507510081Claude CodeClaude Opus5.5 (xhigh)79CodexGPT-6.1 Sol(xhigh)78Claude CodeClaude Sonnet5.5 (xhigh)77Claude CodeClaude Opus 5(high)76Claude CodeClaude Haiku5.5 (xhigh)76Claude CodeClaude Opus4.8 (high)75CodexGPT-6 Luna(xhigh)75Claude CodeClaude Opus4.7 (high)74Claude CodeClaude Sonnet5 (high)67Terminus-2GLM-5.3-Flash(max)62Terminus-2DeepSeek V4Pro (max)61Claude CodeClaude Haiku4.5 (thinking32k)57Terminus-2GLM-5.2(none)41Terminus-2gpt-oss-120b(low)
How to read this chart

What this chart shows. The default view of the Geospatial Intelligence Index, coloured by whether the developer publishes the model's weights for download (open weights) or not (proprietary). Highest open weights model: GLM-5.3-Flash (max) (66.8); highest proprietary model: Claude Opus 5.5 (xhigh) (80.8).

Rows. Each model is shown at its highest reasoning level, as Artificial Analysis does; other levels are under 'Show every reasoning level'.

Data table: Geospatial Intelligence Index: open weights vs proprietary
Geospatial Intelligence Index: open weights vs proprietary
ModelDeveloperWeightsGeospatial Intelligence Index
Claude Opus 5.5 (xhigh)AnthropicProprietary80.8
GPT-6.1 Sol (xhigh)OpenAIProprietary78.7
Claude Sonnet 5.5 (xhigh)AnthropicProprietary78.1
Claude Opus 5 (high)AnthropicProprietary77.0
Claude Haiku 5.5 (xhigh)AnthropicProprietary76.0
Claude Opus 4.8 (high)AnthropicProprietary75.8
GPT-6 Luna (xhigh)OpenAIProprietary74.7
Claude Opus 4.7 (high)AnthropicProprietary74.6
Claude Sonnet 5 (high)AnthropicProprietary73.9
GLM-5.3-Flash (max)Z.aiOpen weights66.8
DeepSeek V4 Pro (max)DeepSeekOpen weights61.8
Claude Haiku 4.5 (thinking 32k)AnthropicProprietary61.2
GLM-5.2 (none)Z.aiOpen weights56.7
gpt-oss-120b (low)OpenAIOpen weights40.6

Text-only view

Geospatial Intelligence Index: text-only view

Weighted average of agents, coding and Earth science, 0 to 100 · Higher is better

Geospatial Intelligence Index: text-only viewClaude Opus 5.5 (xhigh): 87; GPT-6.1 Sol (xhigh): 85; Claude Sonnet 5.5 (xhigh): 84; Claude Opus 5 (high): 84; Claude Opus 4.8 (high): 83; Claude Haiku 5.5 (xhigh): 83; Claude Opus 4.7 (high): 82; Claude Sonnet 5 (high): 82; GPT-6 Luna (xhigh): 82; DeepSeek V4 Pro (max): 79; GLM-5.3-Flash (max): 73; GLM-5.2 (none): 73; Claude Haiku 4.5 (thinking 32k): 69; gpt-oss-120b (low): 52025507510087Claude CodeClaude Opus5.5 (xhigh)85CodexGPT-6.1 Sol(xhigh)84Claude CodeClaude Sonnet5.5 (xhigh)84Claude CodeClaude Opus 5(high)83Claude CodeClaude Opus4.8 (high)83Claude CodeClaude Haiku5.5 (xhigh)82Claude CodeClaude Opus4.7 (high)82Claude CodeClaude Sonnet5 (high)82CodexGPT-6 Luna(xhigh)79Terminus-2DeepSeek V4Pro (max)73Terminus-2GLM-5.3-Flash(max)73Terminus-2GLM-5.2(none)69Claude CodeClaude Haiku4.5 (thinking32k)52Terminus-2gpt-oss-120b(low)
How to read this chart

Rows. Each model once, at its highest reasoning level, labelled with that level, as Artificial Analysis does; when that level is not ranked yet, the highest ranked level, said in its tooltip. Other levels are under 'Show every reasoning level'.

What this metric means. The weighted average of the group scores, each 0 to 100: agents 3/7; coding 2/7; earth science 2/7. The same index without perception, so models that cannot take images compare fairly.

Which harness. Each model is scored in its developer's own harness where we run one, otherwise in Terminus-2. Claude models are scored in Claude Code, OpenAI models in Codex, Kimi models in Kimi Code once their runs there are complete, and all other models in Terminus-2. This applies to the agents and coding groups; Earth science and perception use one model call per question, with no agent harness. Part of a difference between two models can therefore come from the harness; the Geospatial Agent Index is the place to compare harnesses. The bar shows the harness the agents and coding groups are scored in.

Not shown: partial coverage. 15 models are not yet scored in full in every group of this view. They are listed under partial coverage in the leaderboard, with what is still to come.

Data table: Geospatial Intelligence Index: text-only view
Geospatial Intelligence Index: text-only view
RankModelDeveloperText-only viewAgents (3/7)Coding (2/7)Earth science (2/7)Scored in
1Claude Opus 5.5 (xhigh)Anthropic87.1749994Claude Code
2GPT-6.1 Sol (xhigh)OpenAI85.1729793Codex
3Claude Sonnet 5.5 (xhigh)Anthropic84.0699694Claude Code
4Claude Opus 5 (high)Anthropic83.8709593Claude Code
5Claude Opus 4.8 (high)Anthropic83.2669893Claude Code
6Claude Haiku 5.5 (xhigh)Anthropic82.8679792Claude Code
7Claude Opus 4.7 (high)Anthropic82.3659893Claude Code
8Claude Sonnet 5 (high)Anthropic81.6659691Claude Code
9GPT-6 Luna (xhigh)OpenAI81.6649891Codex
10DeepSeek V4 Pro (max)DeepSeek79.5599892Terminus-2
11GLM-5.3-Flash (max)Z.ai73.4548492Terminus-2
12GLM-5.2 (none) (Highest ranked level; GLM-5.2 (max) is not ranked yet)Z.ai72.8558389Terminus-2
13Claude Haiku 4.5 (thinking 32k)Anthropic68.9517689Claude Code
14gpt-oss-120b (low) (Highest ranked level; gpt-oss-120b (high) is not ranked yet)OpenAI52.2285488Terminus-2
Show every reasoning level

Geospatial Intelligence Index: text-only view, every reasoning level

Weighted average of agents, coding and Earth science, 0 to 100 · Higher is better

Geospatial Intelligence Index: text-only view, every reasoning levelClaude Opus 5.5 (xhigh): 87; Claude Opus 5.5 (high): 87; GPT-6.1 Sol (medium): 86; GPT-6.1 Sol (xhigh): 85; GPT-6.1 Sol (high): 85; Claude Opus 5.5 (medium): 85; Claude Sonnet 5.5 (high): 84; Claude Sonnet 5.5 (xhigh): 84; Claude Opus 5 (high): 84; Claude Opus 4.8 (high): 83; Claude Haiku 5.5 (xhigh): 83; Claude Opus 4.7 (high): 82; Claude Sonnet 5 (high): 82; GPT-6 Luna (xhigh): 82; Claude Sonnet 5.5 (medium): 81; Claude Haiku 5.5 (high): 81; Claude Haiku 5.5 (medium): 80; DeepSeek V4 Pro (max): 79; GLM-5.3-Flash (max): 73; GLM-5.2 (none): 73; Claude Haiku 4.5 (thinking 32k): 69; gpt-oss-120b (low): 52025507510087Claude CodeClaude Opus5.5 (xhigh)87Claude CodeClaude Opus5.5 (high)86CodexGPT-6.1 Sol(medium)85CodexGPT-6.1 Sol(xhigh)85CodexGPT-6.1 Sol(high)85Claude CodeClaude Opus5.5 (medium)84Claude CodeClaude Sonnet5.5 (high)84Claude CodeClaude Sonnet5.5 (xhigh)84Claude CodeClaude Opus 5(high)83Claude CodeClaude Opus4.8 (high)83Claude CodeClaude Haiku5.5 (xhigh)82Claude CodeClaude Opus4.7 (high)82Claude CodeClaude Sonnet5 (high)82CodexGPT-6 Luna(xhigh)81Claude CodeClaude Sonnet5.5 (medium)81Claude CodeClaude Haiku5.5 (high)80Claude CodeClaude Haiku5.5 (medium)79Terminus-2DeepSeek V4Pro (max)73Terminus-2GLM-5.3-Flash(max)73Terminus-2GLM-5.2(none)69Claude CodeClaude Haiku4.5 (thinking32k)52Terminus-2gpt-oss-120b(low)
How to read this chart

Rows. One bar per model and reasoning level.

What this metric means. The weighted average of the group scores, each 0 to 100: agents 3/7; coding 2/7; earth science 2/7. The same index without perception, so models that cannot take images compare fairly.

Which harness. Each model is scored in its developer's own harness where we run one, otherwise in Terminus-2. Claude models are scored in Claude Code, OpenAI models in Codex, Kimi models in Kimi Code once their runs there are complete, and all other models in Terminus-2. This applies to the agents and coding groups; Earth science and perception use one model call per question, with no agent harness. Part of a difference between two models can therefore come from the harness; the Geospatial Agent Index is the place to compare harnesses. The bar shows the harness the agents and coding groups are scored in.

Not shown: partial coverage. 26 models are not yet scored in full in every group of this view. They are listed under partial coverage in the leaderboard, with what is still to come.

Data table: Geospatial Intelligence Index: text-only view, every reasoning level
Geospatial Intelligence Index: text-only view, every reasoning level
RankModelDeveloperText-only viewAgents (3/7)Coding (2/7)Earth science (2/7)Scored in
1Claude Opus 5.5 (xhigh)Anthropic87.1749994Claude Code
2Claude Opus 5.5 (high)Anthropic86.87310095Claude Code
3GPT-6.1 Sol (medium)OpenAI85.87210092Codex
4GPT-6.1 Sol (xhigh)OpenAI85.1729793Codex
5GPT-6.1 Sol (high)OpenAI85.17010093Codex
6Claude Opus 5.5 (medium)Anthropic84.8719694Claude Code
7Claude Sonnet 5.5 (high)Anthropic84.3689994Claude Code
8Claude Sonnet 5.5 (xhigh)Anthropic84.0699694Claude Code
9Claude Opus 5 (high)Anthropic83.8709593Claude Code
10Claude Opus 4.8 (high)Anthropic83.2669893Claude Code
11Claude Haiku 5.5 (xhigh)Anthropic82.8679792Claude Code
12Claude Opus 4.7 (high)Anthropic82.3659893Claude Code
13Claude Sonnet 5 (high)Anthropic81.6659691Claude Code
14GPT-6 Luna (xhigh)OpenAI81.6649891Codex
15Claude Sonnet 5.5 (medium)Anthropic81.0639693Claude Code
16Claude Haiku 5.5 (high)Anthropic80.5669192Claude Code
17Claude Haiku 5.5 (medium)Anthropic80.1639492Claude Code
18DeepSeek V4 Pro (max)DeepSeek79.5599892Terminus-2
19GLM-5.3-Flash (max)Z.ai73.4548492Terminus-2
20GLM-5.2 (none)Z.ai72.8558389Terminus-2
21Claude Haiku 4.5 (thinking 32k)Anthropic68.9517689Claude Code
22gpt-oss-120b (low)OpenAI52.2285488Terminus-2

Cost per index task

The average cost of one Agent Index attempt in the scored harness (USD) · Lower is better

  • Planned attempts still to come
Cost per index taskTerminus-2 – gpt-oss-120b (low): $0.0048 (partial coverage); Codex – GPT-6 Luna (xhigh): $0.0076 (partial coverage); Claude Code – Claude Haiku 5.5 (xhigh): $0.018 (partial coverage); Codex – GPT-6.1 Sol (xhigh): $0.091 (partial coverage); Claude Code – Claude Haiku 4.5 (thinking 32k): $0.100 (partial coverage); Terminus-2 – GLM-5.3-Flash (max): $0.106 (partial coverage); Terminus-2 – GLM-5.2 (none): $0.112 (partial coverage); Claude Code – Claude Sonnet 5.5 (xhigh): $0.123 (partial coverage); Claude Code – Claude Sonnet 5 (high): $0.229 (partial coverage); Claude Code – Claude Opus 5.5 (xhigh): $0.284 (partial coverage); Claude Code – Claude Opus 4.8 (high): $0.296 (partial coverage); Terminus-2 – DeepSeek V4 Pro (max): $0.341 (partial coverage); Claude Code – Claude Opus 5 (high): $0.357 (partial coverage); Claude Code – Claude Opus 4.7 (high): $0.382 (partial coverage)$0$0.100$0.200$0.300$0.400$0.0048Terminus-2gpt-oss-120b(low)$0.0076CodexGPT-6 Luna(xhigh)$0.018Claude CodeClaude Haiku5.5 (xhigh)$0.091CodexGPT-6.1 Sol(xhigh)$0.100Claude CodeClaude Haiku4.5 (thinking32k)$0.106Terminus-2GLM-5.3-Flash(max)$0.112Terminus-2GLM-5.2(none)$0.123Claude CodeClaude Sonnet5.5 (xhigh)$0.229Claude CodeClaude Sonnet5 (high)$0.284Claude CodeClaude Opus5.5 (xhigh)$0.296Claude CodeClaude Opus4.8 (high)$0.341Terminus-2DeepSeek V4Pro (max)$0.357Claude CodeClaude Opus 5(high)$0.382Claude CodeClaude Opus4.7 (high)
How to read this chart

Cost per index task. The average cost of one Agent Index attempt in the scored harness: one attempt at a Geospatial Agent Index task, at public API list prices, in the harness the model is scored in (the agents group). Quiz and image questions are one model call each and are not included.

Rows. Each model is shown at its highest reasoning level, as Artificial Analysis does; other levels are under 'Show every reasoning level'.

Data table: Cost per index task
Cost per index task
ModelCreatorCost per index taskCoverage (attempts)
Claude Code – Claude Opus 5.5 (xhigh)Anthropic$0.284565 of 859 planned attempts
Codex – GPT-6.1 Sol (xhigh)OpenAI$0.091565 of 859 planned attempts
Claude Code – Claude Sonnet 5.5 (xhigh)Anthropic$0.123565 of 859 planned attempts
Claude Code – Claude Opus 5 (high)Anthropic$0.357565 of 859 planned attempts
Claude Code – Claude Haiku 5.5 (xhigh)Anthropic$0.018565 of 859 planned attempts
Claude Code – Claude Opus 4.8 (high)Anthropic$0.296565 of 859 planned attempts
Codex – GPT-6 Luna (xhigh)OpenAI$0.0076565 of 859 planned attempts
Claude Code – Claude Opus 4.7 (high)Anthropic$0.382565 of 859 planned attempts
Claude Code – Claude Sonnet 5 (high)Anthropic$0.229565 of 859 planned attempts
Terminus-2 – GLM-5.3-Flash (max)Z.ai$0.106565 of 859 planned attempts
Terminus-2 – DeepSeek V4 Pro (max)DeepSeek$0.341565 of 859 planned attempts
Claude Code – Claude Haiku 4.5 (thinking 32k)Anthropic$0.100565 of 859 planned attempts
Terminus-2 – GLM-5.2 (none)Z.ai$0.112565 of 859 planned attempts
Terminus-2 – gpt-oss-120b (low)OpenAI$0.0048565 of 859 planned attempts

Geospatial Intelligence Index vs. cost per index task

Higher and further left is better

  • Most attractive quadrant
  • Pareto line
Geospatial Intelligence Index vs. cost per index taskClaude Opus 5.5 (xhigh): Geospatial Intelligence Index 80.81, cost per index task (usd) $0.284; GPT-6.1 Sol (xhigh): Geospatial Intelligence Index 78.68, cost per index task (usd) $0.091; Claude Sonnet 5.5 (xhigh): Geospatial Intelligence Index 78.08, cost per index task (usd) $0.123; Claude Opus 5 (high): Geospatial Intelligence Index 76.98, cost per index task (usd) $0.357; Claude Haiku 5.5 (xhigh): Geospatial Intelligence Index 76.05, cost per index task (usd) $0.018; Claude Opus 4.8 (high): Geospatial Intelligence Index 75.76, cost per index task (usd) $0.296; GPT-6 Luna (xhigh): Geospatial Intelligence Index 74.65, cost per index task (usd) $0.0076; Claude Opus 4.7 (high): Geospatial Intelligence Index 74.58, cost per index task (usd) $0.382; Claude Sonnet 5 (high): Geospatial Intelligence Index 73.92, cost per index task (usd) $0.229; GLM-5.3-Flash (max): Geospatial Intelligence Index 66.82, cost per index task (usd) $0.106; DeepSeek V4 Pro (max): Geospatial Intelligence Index 61.81, cost per index task (usd) $0.341; Claude Haiku 4.5 (thinking 32k): Geospatial Intelligence Index 61.2, cost per index task (usd) $0.100; GLM-5.2 (none): Geospatial Intelligence Index 56.66, cost per index task (usd) $0.112; gpt-oss-120b (low): Geospatial Intelligence Index 40.57, cost per index task (usd) $0.0048Most attractive quadrant3540455055606570758085$0.002$0.005$0.01$0.02$0.05$0.10$0.20$0.50Cost per index task (USD) (log scale)Geospatial Intelligence IndexClaude Opus 5.5 (xhigh)GPT-6.1 Sol (xhigh)Claude Sonnet 5.5 (xhigh)Claude Opus 5 (high)Claude Haiku 5.5 (xhigh)Claude Opus 4.8 (high)GPT-6 Luna (xhigh)Claude Opus 4.7 (high)Claude Sonnet 5 (high)GLM-5.3-Flash (max)DeepSeek V4 Pro (max)Claude Haiku 4.5 (thinking 32k)GLM-5.2 (none)gpt-oss-120b (low)
How to read this chart

How to read this chart. Each point is one model at its highest reasoning level. Further left means a lower average cost per index task (log scale); higher means a higher Geospatial Intelligence Index. Both axes fit the models shown, so the score axis does not start at zero. The shaded most attractive quadrant is the top-left quarter of the plot, split at the middle of each axis: a reading aid, not a fixed bar. The dotted Pareto line joins the models that no cheaper model beats on the index.

Cost per index task. The average cost of one Agent Index attempt in the scored harness: one attempt at a Geospatial Agent Index task, at public API list prices, in the harness the model is scored in (the agents group). Quiz and image questions are one model call each and are not included.

Data table: Geospatial Intelligence Index vs. cost per index task
Geospatial Intelligence Index vs. cost per index task
ModelGeospatial Intelligence IndexCost per index task (USD)On the Pareto line
Claude Opus 5.5 (xhigh)80.8$0.284Yes
GPT-6.1 Sol (xhigh)78.7$0.091Yes
Claude Sonnet 5.5 (xhigh)78.1$0.123No
Claude Opus 5 (high)77.0$0.357No
Claude Haiku 5.5 (xhigh)76.0$0.018Yes
Claude Opus 4.8 (high)75.8$0.296No
GPT-6 Luna (xhigh)74.7$0.0076Yes
Claude Opus 4.7 (high)74.6$0.382No
Claude Sonnet 5 (high)73.9$0.229No
GLM-5.3-Flash (max)66.8$0.106No
DeepSeek V4 Pro (max)61.8$0.341No
Claude Haiku 4.5 (thinking 32k)61.2$0.100No
GLM-5.2 (none)56.7$0.112No
gpt-oss-120b (low)40.6$0.0048Yes

Leaderboard

Show the leaderboard tables

Every group score is shown next to the index, 0 to 100. An asterisk marks a group still to be completed.

Geospatial Intelligence Index

Geospatial Intelligence Index: ranked models. Group scores 0 to 100; the index is their weighted average.
RankModelDeveloperGeospatial Intelligence IndexText-only viewAgents (3/9)Coding (2/9)Earth science (2/9)Perception (2/9)Scored in
1Claude Opus 5.5 (xhigh)Anthropic80.887.174999459Claude Code
2GPT-6.1 Sol (xhigh)OpenAI78.785.172979356Codex
3Claude Sonnet 5.5 (xhigh)Anthropic78.184.069969457Claude Code
4Claude Opus 5 (high)Anthropic77.083.870959353Claude Code
5Claude Haiku 5.5 (xhigh)Anthropic76.082.867979252Claude Code
6Claude Opus 4.8 (high)Anthropic75.883.266989350Claude Code
7GPT-6 Luna (xhigh)OpenAI74.781.664989150Codex
8Claude Opus 4.7 (high)Anthropic74.682.365989347Claude Code
9Claude Sonnet 5 (high)Anthropic73.981.665969147Claude Code
10GLM-5.3-Flash (max)Z.ai66.873.454849244Terminus-2
11DeepSeek V4 Pro (max)DeepSeek61.879.55998920 (cannot take images)Terminus-2
12Claude Haiku 4.5 (thinking 32k)Anthropic61.268.951768934Claude Code
13GLM-5.2 (none) (Highest ranked level; GLM-5.2 (max) is not ranked yet)Z.ai56.772.85583890 (cannot take images)Terminus-2
14gpt-oss-120b (low) (Highest ranked level; gpt-oss-120b (high) is not ranked yet)OpenAI40.652.22854880 (cannot take images)Terminus-2

Each model is shown at its highest reasoning level, as Artificial Analysis does; other levels are under 'Show every reasoning level'.

Partial coverage

A model is ranked only when every group is scored in full. These models are not ranked yet and have no index; their group scores so far carry an asterisk, and Coverage says what is still to come.

Geospatial Intelligence Index: partial coverage, not ranked
CoverageModelDeveloperAgents (3/9)Coding (2/9)Earth science (2/9)Perception (2/9)Scored in
Coding: attempt 1 on 0 of 50 GeoAgentBench, 0 of 19 GeoAnalystBench tasksKimi K2.7 Code (always on)Moonshot AI64No data9244Kimi Code
Agents: attempt 1 on 242 of 245 GS-QA2 tasksDeepSeek V4 Flash (max)DeepSeek61*95910 (cannot take images)Terminus-2
Agents: attempt 1 on 240 of 245 GS-QA2 tasksGLM-5.3 (max, default)Z.ai63*89920 (cannot take images)Terminus-2
Coding: attempt 1 on 0 of 50 GeoAgentBench, 0 of 19 GeoAnalystBench tasksKimi K2.6 (always on)Moonshot AI61No data9244Kimi Code
Agents: attempt 1 on 210 of 245 GS-QA2 tasksGemma 4 26B A4B (thinking on)Google51*758943Terminus-2
Agents: attempt 1 on 235 of 245 GS-QA2 tasksGLM-4.7-Flash (thinking on)Z.ai39*39860 (cannot take images)Terminus-2
Earth science: quiz not runLlama 3.3 70B (Non-reasoning)Meta84No data0 (cannot take images)Terminus-2
Earth science: quiz 1,580 of 1,685 questions so far; Perception: 134 of 134 image questions scored so farKimi K3 (always on)Moonshot AI649293*48*Terminus-2
Agents: attempt 1 on 172 of 173 GeoBenchX, 230 of 245 GS-QA2 tasks; Coding: attempt 1 on 49 of 50 GeoAgentBench tasksgpt-oss-20b (high)OpenAI29*45*850 (cannot take images)Terminus-2
Agents and coding: no agent entry publishedMistral Small 3.1 24B (Non-reasoning)Mistral AINo dataNo data8233No agent entry
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.8 Flash (high)Google39*53No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.1 Pro (preview) (high)Google54*90No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz stopped part-way (840 of 1,685 questions); Perception: image questions not run yetClaude Sonnet 4.6 (adaptive, effort high)Anthropic47*9898*No dataTerminus-2
Agents: attempt 1 on 35 of 147 Earth-Bench V2, 170 of 173 GeoBenchX, 0 of 245 GS-QA2 tasks; Coding: attempt 1 on 48 of 50 GeoAgentBench tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.1 Flash-Lite (high)Google46*63*No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 173 GeoBenchX, 0 of 245 GS-QA2 tasks; Coding: attempt 1 on 48 of 50 GeoAgentBench, 5 of 19 GeoAnalystBench tasks; Perception: 128 of 136 image questions scored so farQwen 3.8 27B (xhigh)Alibaba (Qwen)28*80*9153*Terminus-2
Show every reasoning level
Geospatial Intelligence Index: ranked models (every reasoning level). Group scores 0 to 100; the index is their weighted average.
RankModelDeveloperGeospatial Intelligence IndexText-only viewAgents (3/9)Coding (2/9)Earth science (2/9)Perception (2/9)Scored in
1Claude Opus 5.5 (xhigh)Anthropic80.887.174999459Claude Code
2Claude Opus 5.5 (high)Anthropic80.686.8731009559Claude Code
3GPT-6.1 Sol (medium)OpenAI78.885.8721009255Codex
4Claude Sonnet 5.5 (high)Anthropic78.784.368999459Claude Code
5GPT-6.1 Sol (xhigh)OpenAI78.785.172979356Codex
6Claude Opus 5.5 (medium)Anthropic78.784.871969457Claude Code
7Claude Sonnet 5.5 (xhigh)Anthropic78.184.069969457Claude Code
8GPT-6.1 Sol (high)OpenAI77.885.1701009353Codex
9Claude Opus 5 (high)Anthropic77.083.870959353Claude Code
10Claude Haiku 5.5 (xhigh)Anthropic76.082.867979252Claude Code
11Claude Opus 4.8 (high)Anthropic75.883.266989350Claude Code
12Claude Sonnet 5.5 (medium)Anthropic75.481.063969356Claude Code
13GPT-6 Luna (xhigh)OpenAI74.781.664989150Codex
14Claude Opus 4.7 (high)Anthropic74.682.365989347Claude Code
15Claude Sonnet 5 (high)Anthropic73.981.665969147Claude Code
16Claude Haiku 5.5 (high)Anthropic73.280.566919247Claude Code
17Claude Haiku 5.5 (medium)Anthropic72.480.163949245Claude Code
18GLM-5.3-Flash (max)Z.ai66.873.454849244Terminus-2
19DeepSeek V4 Pro (max)DeepSeek61.879.55998920 (cannot take images)Terminus-2
20Claude Haiku 4.5 (thinking 32k)Anthropic61.268.951768934Claude Code
21GLM-5.2 (none)Z.ai56.772.85583890 (cannot take images)Terminus-2
22gpt-oss-120b (low)OpenAI40.652.22854880 (cannot take images)Terminus-2

Partial coverage (every reasoning level)

A model is ranked only when every group is scored in full. These models are not ranked yet and have no index; their group scores so far carry an asterisk, and Coverage says what is still to come.

Geospatial Intelligence Index: partial coverage, not ranked (every reasoning level)
CoverageModelDeveloperAgents (3/9)Coding (2/9)Earth science (2/9)Perception (2/9)Scored in
Coding: attempt 1 on 0 of 50 GeoAgentBench, 0 of 19 GeoAnalystBench tasksKimi K2.7 Code (always on)Moonshot AI64No data9244Kimi Code
Agents: attempt 1 on 238 of 245 GS-QA2 tasksGLM-5.2 (max)Z.ai64*94890 (cannot take images)Terminus-2
Agents: attempt 1 on 236 of 245 GS-QA2 tasksDeepSeek V4 Flash (high)DeepSeek64*92910 (cannot take images)Terminus-2
Agents: attempt 1 on 239 of 245 GS-QA2 tasksDeepSeek V4 Pro (high)DeepSeek63*98920 (cannot take images)Terminus-2
Agents: attempt 1 on 240 of 245 GS-QA2 tasksGLM-5.3 (max, default)Z.ai63*89920 (cannot take images)Terminus-2
Agents: attempt 1 on 244 of 245 GS-QA2 tasksGLM-5.2 (high)Z.ai62*97910 (cannot take images)Terminus-2
Agents: attempt 1 on 242 of 245 GS-QA2 tasksDeepSeek V4 Flash (max)DeepSeek61*95910 (cannot take images)Terminus-2
Coding: attempt 1 on 0 of 50 GeoAgentBench, 0 of 19 GeoAnalystBench tasksKimi K2.6 (always on)Moonshot AI61No data9244Kimi Code
Agents: attempt 1 on 241 of 245 GS-QA2 tasksGLM-5.3 (low)Z.ai58*93910 (cannot take images)Terminus-2
Agents: attempt 1 on 210 of 245 GS-QA2 tasksGemma 4 26B A4B (thinking on)Google51*758943Terminus-2
Agents: attempt 1 on 241 of 245 GS-QA2 tasksgpt-oss-120b (high)OpenAI40*61900 (cannot take images)Terminus-2
Agents: attempt 1 on 235 of 245 GS-QA2 tasksGLM-4.7-Flash (thinking on)Z.ai39*39860 (cannot take images)Terminus-2
Earth science: quiz not runLlama 3.3 70B (Non-reasoning)Meta84No data0 (cannot take images)Terminus-2
Earth science: quiz 1,580 of 1,685 questions so far; Perception: 134 of 134 image questions scored so farKimi K3 (always on)Moonshot AI649293*48*Terminus-2
Agents: attempt 1 on 244 of 245 GS-QA2 tasks; Perception: 122 of 136 image questions scored so farGLM-5.3-Flash (low)Z.ai52*948546*Terminus-2
Agents: attempt 1 on 172 of 173 GeoBenchX, 230 of 245 GS-QA2 tasks; Coding: attempt 1 on 49 of 50 GeoAgentBench tasksgpt-oss-20b (high)OpenAI29*45*850 (cannot take images)Terminus-2
Agents and coding: no agent entry publishedMistral Small 3.1 24B (Non-reasoning)Mistral AINo dataNo data8233No agent entry
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.8 Flash (low)Google57*94No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.1 Pro (preview) (high)Google54*90No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.1 Pro (preview) (low)Google48*81No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz stopped part-way (840 of 1,685 questions); Perception: image questions not run yetClaude Sonnet 4.6 (adaptive, effort high)Anthropic47*9898*No dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.8 Flash (high)Google39*53No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.1 Flash-Lite (low)Google39*58No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 245 GS-QA2 tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.1 Flash-Lite (minimal)Google35*42No dataNo dataTerminus-2
Agents: attempt 1 on 36 of 147 Earth-Bench V2, 0 of 173 GeoBenchX, 0 of 245 GS-QA2 tasks; Coding: attempt 1 on 48 of 50 GeoAgentBench, 5 of 19 GeoAnalystBench tasks; Perception: 128 of 136 image questions scored so farQwen 3.8 27B (xhigh)Alibaba (Qwen)28*80*9153*Terminus-2
Agents: attempt 1 on 35 of 147 Earth-Bench V2, 170 of 173 GeoBenchX, 0 of 245 GS-QA2 tasks; Coding: attempt 1 on 48 of 50 GeoAgentBench tasks; Earth science: quiz not run; Perception: image questions not run yetGemini 3.1 Flash-Lite (high)Google46*63*No dataNo dataTerminus-2

Text-only view

The same index without perception, with weights agents 3/7, coding 2/7, earth science 2/7. It compares models that cannot take images fairly, and shows how much perception changes the order. A model is ranked here when its agents, coding and Earth science groups are scored in full.

Text-only view: ranked models
RankModelDeveloperText-only viewAgents (3/7)Coding (2/7)Earth science (2/7)Scored in
1Claude Opus 5.5 (xhigh)Anthropic87.1749994Claude Code
2GPT-6.1 Sol (xhigh)OpenAI85.1729793Codex
3Claude Sonnet 5.5 (xhigh)Anthropic84.0699694Claude Code
4Claude Opus 5 (high)Anthropic83.8709593Claude Code
5Claude Opus 4.8 (high)Anthropic83.2669893Claude Code
6Claude Haiku 5.5 (xhigh)Anthropic82.8679792Claude Code
7Claude Opus 4.7 (high)Anthropic82.3659893Claude Code
8Claude Sonnet 5 (high)Anthropic81.6659691Claude Code
9GPT-6 Luna (xhigh)OpenAI81.6649891Codex
10DeepSeek V4 Pro (max)DeepSeek79.5599892Terminus-2
11GLM-5.3-Flash (max)Z.ai73.4548492Terminus-2
12GLM-5.2 (none) (Highest ranked level; GLM-5.2 (max) is not ranked yet)Z.ai72.8558389Terminus-2
13Claude Haiku 4.5 (thinking 32k)Anthropic68.9517689Claude Code
14gpt-oss-120b (low) (Highest ranked level; gpt-oss-120b (high) is not ranked yet)OpenAI52.2285488Terminus-2

Each model is shown at its highest reasoning level, as Artificial Analysis does; other levels are under 'Show every reasoning level'.

Show every reasoning level
Text-only view: ranked models (every reasoning level)
RankModelDeveloperText-only viewAgents (3/7)Coding (2/7)Earth science (2/7)Scored in
1Claude Opus 5.5 (xhigh)Anthropic87.1749994Claude Code
2Claude Opus 5.5 (high)Anthropic86.87310095Claude Code
3GPT-6.1 Sol (medium)OpenAI85.87210092Codex
4GPT-6.1 Sol (xhigh)OpenAI85.1729793Codex
5GPT-6.1 Sol (high)OpenAI85.17010093Codex
6Claude Opus 5.5 (medium)Anthropic84.8719694Claude Code
7Claude Sonnet 5.5 (high)Anthropic84.3689994Claude Code
8Claude Sonnet 5.5 (xhigh)Anthropic84.0699694Claude Code
9Claude Opus 5 (high)Anthropic83.8709593Claude Code
10Claude Opus 4.8 (high)Anthropic83.2669893Claude Code
11Claude Haiku 5.5 (xhigh)Anthropic82.8679792Claude Code
12Claude Opus 4.7 (high)Anthropic82.3659893Claude Code
13Claude Sonnet 5 (high)Anthropic81.6659691Claude Code
14GPT-6 Luna (xhigh)OpenAI81.6649891Codex
15Claude Sonnet 5.5 (medium)Anthropic81.0639693Claude Code
16Claude Haiku 5.5 (high)Anthropic80.5669192Claude Code
17Claude Haiku 5.5 (medium)Anthropic80.1639492Claude Code
18DeepSeek V4 Pro (max)DeepSeek79.5599892Terminus-2
19GLM-5.3-Flash (max)Z.ai73.4548492Terminus-2
20GLM-5.2 (none)Z.ai72.8558389Terminus-2
21Claude Haiku 4.5 (thinking 32k)Anthropic68.9517689Claude Code
22gpt-oss-120b (low)OpenAI52.2285488Terminus-2

Geospatial Agent Index

Which agent setup to use: each row is a model in one agent harness, at its highest reasoning level here. The Geospatial Agent Index is the agents group of the Geospatial Intelligence Index. See the full Geospatial Agent Index, with every harness and reasoning level.

Geospatial Agent Index

Equal-weight average of the evaluation scores, 0 to 100 · Higher is better

  • Planned attempts still to come
Geospatial Agent IndexClaude Code – Claude Opus 5.5 (xhigh): 74* (partial coverage); Codex – GPT-6.1 Sol (xhigh): 72* (partial coverage); Claude Code – Claude Opus 5 (high): 70* (partial coverage); Claude Code – Claude Sonnet 5.5 (xhigh): 69* (partial coverage); Claude Code – Claude Haiku 5.5 (xhigh): 67* (partial coverage); Claude Code – Claude Opus 4.8 (high): 66* (partial coverage); Claude Code – Claude Sonnet 5 (high): 65* (partial coverage); Claude Code – Claude Opus 4.7 (high): 65* (partial coverage); Codex – GPT-6 Luna (xhigh): 64* (partial coverage); Kimi Code – Kimi K2.7 Code (always on): 64* (partial coverage); Terminus-2 – Kimi K3 (always on): 64* (partial coverage); Kimi Code – Kimi K2.6 (always on): 61* (partial coverage); Terminus-2 – DeepSeek V4 Pro (max): 59* (partial coverage); Terminus-2 – GLM-5.2 (none): 55* (partial coverage); Terminus-2 – GLM-5.3-Flash (max): 54* (partial coverage); Claude Code – Claude Haiku 4.5 (thinking 32k): 51* (partial coverage); Terminus-2 – gpt-oss-120b (low): 28* (partial coverage); Terminus-2 – Llama 3.3 70B (Non-reasoning): 8* (partial coverage)025507510074*Claude CodeClaude Opus5.5 (xhigh)72*CodexGPT-6.1 Sol(xhigh)70*Claude CodeClaude Opus 5(high)69*Claude CodeClaude Sonnet5.5 (xhigh)67*Claude CodeClaude Haiku5.5 (xhigh)66*Claude CodeClaude Opus4.8 (high)65*Claude CodeClaude Sonnet5 (high)65*Claude CodeClaude Opus4.7 (high)64*CodexGPT-6 Luna(xhigh)64*Kimi CodeKimi K2.7Code (alwayson)64*Terminus-2Kimi K3(always on)61*Kimi CodeKimi K2.6(always on)59*Terminus-2DeepSeek V4Pro (max)55*Terminus-2GLM-5.2(none)54*Terminus-2GLM-5.3-Flash(max)51*Claude CodeClaude Haiku4.5 (thinking32k)28*Terminus-2gpt-oss-120b(low)8*Terminus-2Llama 3.3 70B(Non-reasoning)
How to read this chart

What this metric means. The equal-weight average of the 3 evaluation scores. Each evaluation score is the average pass@1 over that benchmark's tasks, from 0 to 100.

Partial coverage entries. An entry is ranked only when it has a first attempt on every task of every index component. Entries without that are marked partial coverage: they are listed apart with their coverage, for example "36 of 147 Earth-Bench V2 tasks", and are left out of the ranking, the index charts and the Pareto lines. An entry becomes ranked as soon as its first attempts are complete.

Coverage hatch. A hatched bar, and an asterisk in tables, marks an entry that does not yet have results for every planned attempt. While attempts await review, the table gives the range the score could still reach.

First attempt only. Scores currently use each entry's first attempt on every task. The full protocol is 3 attempts; averages over 3 will replace these once every entry has them. An entry without a first attempt on every task yet is marked "partial: attempt 1 on n of N tasks".

Data table: Geospatial Agent Index
Geospatial Agent Index
ModelCreatorGeospatial Agent IndexRange while attempts are pendingBenchmarks coveredCoverage (attempts)
Claude Code – Claude Opus 5.5 (xhigh)Anthropic74*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts
Codex – GPT-6.1 Sol (xhigh)OpenAI72*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts
Claude Code – Claude Opus 5 (high)Anthropic70*No pending judgements; coverage incomplete; 205 results awaiting review3 of 3565 of 859 planned attempts
Claude Code – Claude Sonnet 5.5 (xhigh)Anthropic69*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts
Claude Code – Claude Haiku 5.5 (xhigh)Anthropic67*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts
Claude Code – Claude Opus 4.8 (high)Anthropic66*No pending judgements; coverage incomplete; 188 results awaiting review3 of 3565 of 859 planned attempts
Claude Code – Claude Sonnet 5 (high)Anthropic65*No pending judgements; coverage incomplete; 174 results awaiting review3 of 3565 of 859 planned attempts
Claude Code – Claude Opus 4.7 (high)Anthropic65*No pending judgements; coverage incomplete; 178 results awaiting review3 of 3565 of 859 planned attempts
Codex – GPT-6 Luna (xhigh)OpenAI64*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts
Kimi Code – Kimi K2.7 Code (always on)Moonshot AI64*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts (3 excluded)
Terminus-2 – Kimi K3 (always on)Moonshot AI64*No pending judgements; coverage incomplete; 183 results awaiting review; 1 trace review declined by the judge3 of 3565 of 859 planned attempts
Kimi Code – Kimi K2.6 (always on)Moonshot AI61*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts (16 excluded)
Terminus-2 – DeepSeek V4 Pro (max)DeepSeek59*No pending judgements; coverage incomplete; 172 results awaiting review; 1 trace review declined by the judge3 of 3565 of 859 planned attempts
Terminus-2 – GLM-5.2 (none)Z.ai55*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts
Terminus-2 – GLM-5.3-Flash (max)Z.ai54*No pending judgements; coverage incomplete; 175 results awaiting review3 of 3565 of 859 planned attempts
Claude Code – Claude Haiku 4.5 (thinking 32k)Anthropic51*No pending judgements; coverage incomplete; 159 results awaiting review3 of 3565 of 859 planned attempts
Terminus-2 – gpt-oss-120b (low)OpenAI28*No pending judgements; coverage incomplete3 of 3565 of 859 planned attempts
Terminus-2 – Llama 3.3 70B (Non-reasoning)Meta8*No pending judgements; coverage incomplete; 24 results awaiting review3 of 3565 of 859 planned attempts

Geospatial Agent Index vs. cost per task

Higher and further left is better

  • Planned attempts still to come
  • Most attractive quadrant
  • Pareto line
Geospatial Agent Index vs. cost per taskTerminus-2 – DeepSeek V4 Pro (max): Geospatial Agent Index 58.8, cost per task (usd) $0.341; Terminus-2 – GLM-5.2 (none): Geospatial Agent Index 55.34, cost per task (usd) $0.112; Terminus-2 – GLM-5.3-Flash (max): Geospatial Agent Index 54.02, cost per task (usd) $0.106; Terminus-2 – gpt-oss-120b (low): Geospatial Agent Index 27.57, cost per task (usd) $0.0048; Terminus-2 – Kimi K3 (always on): Geospatial Agent Index 63.57, cost per task (usd) $0.252; Terminus-2 – Llama 3.3 70B (Non-reasoning): Geospatial Agent Index 7.88, cost per task (usd) $0.022; Claude Code – Claude Haiku 4.5 (thinking 32k): Geospatial Agent Index 50.88, cost per task (usd) $0.100; Kimi Code – Kimi K2.6 (always on): Geospatial Agent Index 60.99, cost per task (usd) $0.302; Kimi Code – Kimi K2.7 Code (always on): Geospatial Agent Index 64.12, cost per task (usd) $0.215; Claude Code – Claude Opus 5.5 (xhigh): Geospatial Agent Index 74.33, cost per task (usd) $0.284; Claude Code – Claude Sonnet 5.5 (xhigh): Geospatial Agent Index 69.08, cost per task (usd) $0.123; Claude Code – Claude Haiku 5.5 (xhigh): Geospatial Agent Index 67.08, cost per task (usd) $0.018; Claude Code – Claude Opus 5 (high): Geospatial Agent Index 70.3, cost per task (usd) $0.357; Claude Code – Claude Sonnet 5 (high): Geospatial Agent Index 65.3, cost per task (usd) $0.229; Claude Code – Claude Opus 4.8 (high): Geospatial Agent Index 66.5, cost per task (usd) $0.296; Claude Code – Claude Opus 4.7 (high): Geospatial Agent Index 64.95, cost per task (usd) $0.382; Codex – GPT-6.1 Sol (xhigh): Geospatial Agent Index 71.87, cost per task (usd) $0.091; Codex – GPT-6 Luna (xhigh): Geospatial Agent Index 64.15, cost per task (usd) $0.0076Most attractive quadrant0102030405060708090$0.002$0.005$0.01$0.02$0.05$0.10$0.20$0.50Cost per task (USD) (log scale)Geospatial Agent IndexClaude Code – Claude Opus 5.5 (xhigh)Codex – GPT-6.1 Sol (xhigh)Claude Code – Claude Opus 5 (high)Claude Code – Claude Sonnet 5.5 (xhigh)Claude Code – Claude Haiku 5.5 (xhigh)Claude Code – Claude Opus 4.8 (high)Claude Code – Claude Sonnet 5 (high)Claude Code – Claude Opus 4.7 (high)Codex – GPT-6 Luna (xhigh)Kimi Code – Kimi K2.7 Code (always on)Terminus-2 – Kimi K3 (always on)Kimi Code – Kimi K2.6 (always on)Terminus-2 – DeepSeek V4 Pro (max)Terminus-2 – GLM-5.2 (none)Terminus-2 – GLM-5.3-Flash (max)Claude Code – Claude Haiku 4.5 (thinking 32k)Terminus-2 – gpt-oss-120b (low)Terminus-2 – Llama 3.3 70B (Non-reasoning)
How to read this chart

Rows. Each model and harness is shown at its highest reasoning level, as Artificial Analysis does; other levels are under 'Show every reasoning level'.

Partial coverage. Entries without a first attempt on every index task are not plotted; the table lists them. What is still running is on the models page.

What cost is measuring. What one attempt at a task would cost if paid token by token at the provider's published API prices, averaged over attempts: ordinary input tokens, input tokens served from the provider's cache at the lower cached price, tokens written to the cache where the provider charges for that, and output tokens (thinking included). Cost assumes each harness uses the model's pay-per-token API, priced at public list rates; sandboxes and our own engineering time are not included. Failed attempts count too.

How to read this chart. Each point is one model configuration. Further left means a lower average cost per task (log scale); higher means a higher Geospatial Agent Index. Both axes fit the models shown, so the score axis does not start at zero. The shaded most attractive quadrant is the top-left quarter of the plot, split at the middle of each axis: a reading aid that moves with the axes when you change the models shown, not a fixed bar. The dotted Pareto line joins the shown models that no cheaper shown model beats on score.

Data table: Geospatial Agent Index vs. cost per task
Geospatial Agent Index vs. cost per task
ModelGeospatial Agent IndexCost per task (USD)On the Pareto line
Terminus-2 – Claude Sonnet 4.6 (adaptive, effort high)47* (partial: attempt 1 on 209 of 565 tasks)$0.291Not plotted (partial coverage)
Terminus-2 – DeepSeek V4 Flash (max)61* (partial: attempt 1 on 562 of 565 tasks)$0.069Not plotted (partial coverage)
Terminus-2 – DeepSeek V4 Pro (max)59*$0.341No
Terminus-2 – Gemini 3.1 Flash-Lite (high)46* (partial: attempt 1 on 205 of 565 tasks)$0.021Not plotted (partial coverage)
Terminus-2 – Gemini 3.1 Pro (preview) (high)54* (partial: attempt 1 on 209 of 565 tasks)$0.275Not plotted (partial coverage)
Terminus-2 – Gemini 3.8 Flash (high)39* (partial: attempt 1 on 209 of 565 tasks)$0.336Not plotted (partial coverage)
Terminus-2 – Gemma 4 26B A4B (thinking on)51* (partial: attempt 1 on 530 of 565 tasks)$0.021Not plotted (partial coverage)
Terminus-2 – GLM-4.7-Flash (thinking on)39* (partial: attempt 1 on 555 of 565 tasks)$0.038Not plotted (partial coverage)
Terminus-2 – GLM-5.2 (none)55*$0.112No
Terminus-2 – GLM-5.3-Flash (max)54*$0.106No
Terminus-2 – GLM-5.3 (max, default)63* (partial: attempt 1 on 560 of 565 tasks)$0.481Not plotted (partial coverage)
Terminus-2 – gpt-oss-120b (low)28*$0.0048Yes
Terminus-2 – gpt-oss-20b (high)29* (partial: attempt 1 on 549 of 565 tasks)$0.017Not plotted (partial coverage)
Terminus-2 – Kimi K2.6 (always on)61* (partial: attempt 1 on 562 of 565 tasks)$0.108Not plotted (partial coverage)
Terminus-2 – Kimi K2.7 Code (always on)63* (partial: attempt 1 on 560 of 565 tasks)$0.131Not plotted (partial coverage)
Terminus-2 – Kimi K3 (always on)64*$0.252No
Terminus-2 – Llama 3.3 70B (Non-reasoning)8*$0.022No
Terminus-2 – Qwen 3.8 27B (xhigh)28* (partial: attempt 1 on 36 of 565 tasks)$0.051Not plotted (partial coverage)
Claude Code – Claude Haiku 4.5 (thinking 32k)51*$0.100No
Claude Code – GLM-4.7-Flash (thinking on)37*$0.244Not plotted (partial coverage)
Claude Code – GLM-5.2 (max)66* (partial: attempt 1 on 550 of 565 tasks)$0.373Not plotted (partial coverage)
Claude Code – GLM-5.3-Flash (max)57* (partial: attempt 1 on 557 of 565 tasks)$0.132Not plotted (partial coverage)
Claude Code – GLM-5.3 (max, default)61* (partial: attempt 1 on 294 of 565 tasks)$0.761Not plotted (partial coverage)
Codex – DeepSeek V4 Flash (high)57* (partial: attempt 1 on 412 of 565 tasks)$0.046Not plotted (partial coverage)
Codex – DeepSeek V4 Pro (high)55* (partial: attempt 1 on 129 of 565 tasks)$0.497Not plotted (partial coverage)
Codex – gpt-oss-120b (high)40* (partial: attempt 1 on 401 of 565 tasks)$0.376Not plotted (partial coverage)
Codex – gpt-oss-20b (high)8* (partial: attempt 1 on 540 of 565 tasks)$0.098Not plotted (partial coverage)
Kimi Code – Kimi K2.6 (always on)61*$0.302No
Kimi Code – Kimi K2.7 Code (always on)64*$0.215No
Claude Code – Claude Opus 5.5 (xhigh)74*$0.284Yes
Claude Code – Claude Sonnet 5.5 (xhigh)69*$0.123No
Claude Code – Claude Haiku 5.5 (xhigh)67*$0.018Yes
Claude Code – Claude Opus 5 (high)70*$0.357No
Claude Code – Claude Sonnet 5 (high)65*$0.229No
Claude Code – Claude Opus 4.8 (high)66*$0.296No
Claude Code – Claude Opus 4.7 (high)65*$0.382No
Codex – GPT-6.1 Sol (xhigh)72*$0.091Yes
Codex – GPT-6 Luna (xhigh)64*$0.0076Yes

Benchmarks

The evaluations behind each group of the Geospatial Intelligence Index. Each links to its own page.

Agents

Weight 3/9, about 33.3% of the Geospatial Intelligence Index.

  • Earth-Bench V2 Earth observation data analysis.
  • GeoBenchX Multi-step spatial execution, including tasks to reject.
  • GS-QA2 Spatial data reasoning over map and elevation data.

Agents

Weight 3/9 of the Geospatial Intelligence Index, 0 to 100 · Higher is better

AgentsClaude Opus 5.5 (xhigh): 74; GPT-6.1 Sol (xhigh): 72; Claude Opus 5 (high): 70; Claude Sonnet 5.5 (xhigh): 69; Claude Haiku 5.5 (xhigh): 67; Claude Opus 4.8 (high): 66; Claude Sonnet 5 (high): 65; Claude Opus 4.7 (high): 65; GPT-6 Luna (xhigh): 64; Kimi K2.7 Code (always on): 64; Kimi K3 (always on): 64; Kimi K2.6 (always on): 61; DeepSeek V4 Pro (max): 59; GLM-5.2 (none): 55; GLM-5.3-Flash (max): 54; Claude Haiku 4.5 (thinking 32k): 51; gpt-oss-120b (low): 28; Llama 3.3 70B (Non-reasoning): 8025507510074Claude CodeClaude Opus5.5 (xhigh)72CodexGPT-6.1 Sol(xhigh)70Claude CodeClaude Opus 5(high)69Claude CodeClaude Sonnet5.5 (xhigh)67Claude CodeClaude Haiku5.5 (xhigh)66Claude CodeClaude Opus4.8 (high)65Claude CodeClaude Sonnet5 (high)65Claude CodeClaude Opus4.7 (high)64CodexGPT-6 Luna(xhigh)64Kimi CodeKimi K2.7Code (alwayson)64Terminus-2Kimi K3(always on)61Kimi CodeKimi K2.6(always on)59Terminus-2DeepSeek V4Pro (max)55Terminus-2GLM-5.2(none)54Terminus-2GLM-5.3-Flash(max)51Claude CodeClaude Haiku4.5 (thinking32k)28Terminus-2gpt-oss-120b(low)8Terminus-2Llama 3.3 70B(Non-reasoning)
How to read this chart

Rows. Each model once, at its highest reasoning level with this group scored in full, so models not yet ranked in the Geospatial Intelligence Index can appear here.

Data table: Agents
Agents: score per model
ModelDeveloperAgentsScored in
Claude Opus 5.5 (xhigh)Anthropic74Claude Code
GPT-6.1 Sol (xhigh)OpenAI72Codex
Claude Opus 5 (high)Anthropic70Claude Code
Claude Sonnet 5.5 (xhigh)Anthropic69Claude Code
Claude Haiku 5.5 (xhigh)Anthropic67Claude Code
Claude Opus 4.8 (high)Anthropic66Claude Code
Claude Sonnet 5 (high)Anthropic65Claude Code
Claude Opus 4.7 (high)Anthropic65Claude Code
GPT-6 Luna (xhigh)OpenAI64Codex
Kimi K2.7 Code (always on)Moonshot AI64Kimi Code
Kimi K3 (always on)Moonshot AI64Terminus-2
Kimi K2.6 (always on)Moonshot AI61Kimi Code
DeepSeek V4 Pro (max)DeepSeek59Terminus-2
GLM-5.2 (none)Z.ai55Terminus-2
GLM-5.3-Flash (max)Z.ai54Terminus-2
Claude Haiku 4.5 (thinking 32k)Anthropic51Claude Code
gpt-oss-120b (low)OpenAI28Terminus-2
Llama 3.3 70B (Non-reasoning)Meta8Terminus-2

Coding

Weight 2/9, about 22.2% of the Geospatial Intelligence Index.

Coding

Weight 2/9 of the Geospatial Intelligence Index, 0 to 100 · Higher is better

CodingClaude Opus 5.5 (xhigh): 99; Claude Opus 4.8 (high): 98; GPT-6 Luna (xhigh): 98; Claude Opus 4.7 (high): 98; DeepSeek V4 Pro (max): 98; Claude Sonnet 4.6 (adaptive, effort high): 98; GPT-6.1 Sol (xhigh): 97; Claude Haiku 5.5 (xhigh): 97; Claude Sonnet 5.5 (xhigh): 96; Claude Sonnet 5 (high): 96; DeepSeek V4 Flash (max): 95; Claude Opus 5 (high): 95; GLM-5.2 (max): 94; Kimi K3 (always on): 92; Gemini 3.1 Pro (preview) (high): 90; GLM-5.3 (max, default): 89; GLM-5.3-Flash (max): 84; Claude Haiku 4.5 (thinking 32k): 76; Gemma 4 26B A4B (thinking on): 75; gpt-oss-120b (high): 61; Gemini 3.1 Flash-Lite (low): 58; Gemini 3.8 Flash (high): 53; GLM-4.7-Flash (thinking on): 39; Llama 3.3 70B (Non-reasoning): 4025507510099Claude CodeClaude Opus5.5 (xhigh)98Claude CodeClaude Opus4.8 (high)98CodexGPT-6 Luna(xhigh)98Claude CodeClaude Opus4.7 (high)98Terminus-2DeepSeek V4Pro (max)98Terminus-2Claude Sonnet4.6(adaptive,97CodexGPT-6.1 Sol(xhigh)97Claude CodeClaude Haiku5.5 (xhigh)96Claude CodeClaude Sonnet5.5 (xhigh)96Claude CodeClaude Sonnet5 (high)95Terminus-2DeepSeek V4Flash (max)95Claude CodeClaude Opus 5(high)94Terminus-2GLM-5.2 (max)92Terminus-2Kimi K3(always on)90Terminus-2Gemini 3.1Pro (preview)(high)89Terminus-2GLM-5.3 (max,default)84Terminus-2GLM-5.3-Flash(max)76Claude CodeClaude Haiku4.5 (thinking32k)75Terminus-2Gemma 4 26BA4B (thinkingon)61Terminus-2gpt-oss-120b(high)58Terminus-2Gemini 3.1Flash-Lite(low)53Terminus-2Gemini 3.8Flash (high)39Terminus-2GLM-4.7-Flash(thinking on)4Terminus-2Llama 3.3 70B(Non-reasoning)
How to read this chart

Rows. Each model once, at its highest reasoning level with this group scored in full, so models not yet ranked in the Geospatial Intelligence Index can appear here.

Data table: Coding
Coding: score per model
ModelDeveloperCodingScored in
Claude Opus 5.5 (xhigh)Anthropic99Claude Code
Claude Opus 4.8 (high)Anthropic98Claude Code
GPT-6 Luna (xhigh)OpenAI98Codex
Claude Opus 4.7 (high)Anthropic98Claude Code
DeepSeek V4 Pro (max)DeepSeek98Terminus-2
Claude Sonnet 4.6 (adaptive, effort high)Anthropic98Terminus-2
GPT-6.1 Sol (xhigh)OpenAI97Codex
Claude Haiku 5.5 (xhigh)Anthropic97Claude Code
Claude Sonnet 5.5 (xhigh)Anthropic96Claude Code
Claude Sonnet 5 (high)Anthropic96Claude Code
DeepSeek V4 Flash (max)DeepSeek95Terminus-2
Claude Opus 5 (high)Anthropic95Claude Code
GLM-5.2 (max)Z.ai94Terminus-2
Kimi K3 (always on)Moonshot AI92Terminus-2
Gemini 3.1 Pro (preview) (high)Google90Terminus-2
GLM-5.3 (max, default)Z.ai89Terminus-2
GLM-5.3-Flash (max)Z.ai84Terminus-2
Claude Haiku 4.5 (thinking 32k)Anthropic76Claude Code
Gemma 4 26B A4B (thinking on)Google75Terminus-2
gpt-oss-120b (high)OpenAI61Terminus-2
Gemini 3.1 Flash-Lite (low)Google58Terminus-2
Gemini 3.8 Flash (high)Google53Terminus-2
GLM-4.7-Flash (thinking on)Z.ai39Terminus-2
Llama 3.3 70B (Non-reasoning)Meta4Terminus-2

Earth science

Weight 2/9, about 22.2% of the Geospatial Intelligence Index.

Earth science

Weight 2/9 of the Geospatial Intelligence Index, 0 to 100 · Higher is better

Earth scienceClaude Opus 5.5 (xhigh): 94; Claude Sonnet 5.5 (xhigh): 94; Claude Opus 4.8 (high): 93; Claude Opus 5 (high): 93; Claude Opus 4.7 (high): 93; GPT-6.1 Sol (xhigh): 93; Kimi K2.7 Code (always on): 92; DeepSeek V4 Pro (max): 92; Claude Haiku 5.5 (xhigh): 92; GLM-5.3 (max, default): 92; Kimi K2.6 (always on): 92; GLM-5.3-Flash (max): 92; DeepSeek V4 Flash (max): 91; Claude Sonnet 5 (high): 91; GPT-6 Luna (xhigh): 91; Qwen 3.8 27B (xhigh): 91; gpt-oss-120b (high): 90; Claude Haiku 4.5 (thinking 32k): 89; GLM-5.2 (max): 89; Gemma 4 26B A4B (thinking on): 89; GLM-4.7-Flash (thinking on): 86; gpt-oss-20b (high): 85; Mistral Small 3.1 24B (Non-reasoning): 82025507510094Claude CodeClaude Opus5.5 (xhigh)94Claude CodeClaude Sonnet5.5 (xhigh)93Claude CodeClaude Opus4.8 (high)93Claude CodeClaude Opus 5(high)93Claude CodeClaude Opus4.7 (high)93CodexGPT-6.1 Sol(xhigh)92Kimi CodeKimi K2.7Code (alwayson)92Terminus-2DeepSeek V4Pro (max)92Claude CodeClaude Haiku5.5 (xhigh)92Terminus-2GLM-5.3 (max,default)92Kimi CodeKimi K2.6(always on)92Terminus-2GLM-5.3-Flash(max)91Terminus-2DeepSeek V4Flash (max)91Claude CodeClaude Sonnet5 (high)91CodexGPT-6 Luna(xhigh)91Terminus-2Qwen 3.8 27B(xhigh)90Terminus-2gpt-oss-120b(high)89Claude CodeClaude Haiku4.5 (thinking32k)89Terminus-2GLM-5.2 (max)89Terminus-2Gemma 4 26BA4B (thinkingon)86Terminus-2GLM-4.7-Flash(thinking on)85Terminus-2gpt-oss-20b(high)82No agent entryMistral Small3.1 24B(Non-reasoning)
How to read this chart

Rows. Each model once, at its highest reasoning level with this group scored in full, so models not yet ranked in the Geospatial Intelligence Index can appear here.

Data table: Earth science
Earth science: score per model
ModelDeveloperEarth scienceScored in
Claude Opus 5.5 (xhigh)Anthropic94Claude Code
Claude Sonnet 5.5 (xhigh)Anthropic94Claude Code
Claude Opus 4.8 (high)Anthropic93Claude Code
Claude Opus 5 (high)Anthropic93Claude Code
Claude Opus 4.7 (high)Anthropic93Claude Code
GPT-6.1 Sol (xhigh)OpenAI93Codex
Kimi K2.7 Code (always on)Moonshot AI92Kimi Code
DeepSeek V4 Pro (max)DeepSeek92Terminus-2
Claude Haiku 5.5 (xhigh)Anthropic92Claude Code
GLM-5.3 (max, default)Z.ai92Terminus-2
Kimi K2.6 (always on)Moonshot AI92Kimi Code
GLM-5.3-Flash (max)Z.ai92Terminus-2
DeepSeek V4 Flash (max)DeepSeek91Terminus-2
Claude Sonnet 5 (high)Anthropic91Claude Code
GPT-6 Luna (xhigh)OpenAI91Codex
Qwen 3.8 27B (xhigh)Alibaba (Qwen)91Terminus-2
gpt-oss-120b (high)OpenAI90Terminus-2
Claude Haiku 4.5 (thinking 32k)Anthropic89Claude Code
GLM-5.2 (max)Z.ai89Terminus-2
Gemma 4 26B A4B (thinking on)Google89Terminus-2
GLM-4.7-Flash (thinking on)Z.ai86Terminus-2
gpt-oss-20b (high)OpenAI85Terminus-2
Mistral Small 3.1 24B (Non-reasoning)Mistral AI82No agent entry

Perception

Weight 2/9, about 22.2% of the Geospatial Intelligence Index.

  • Perception Questions answered from satellite and aerial images, look and answer, the same for every model.
  • Geospatial image understanding The image track in full, with the look-and-work protocol reported apart.

Perception

Weight 2/9 of the Geospatial Intelligence Index, 0 to 100 · Higher is better

PerceptionClaude Opus 5.5 (xhigh): 59; Claude Sonnet 5.5 (xhigh): 57; GPT-6.1 Sol (xhigh): 56; Claude Opus 5 (high): 53; Claude Haiku 5.5 (xhigh): 52; GPT-6 Luna (xhigh): 50; Claude Opus 4.8 (high): 50; Claude Opus 4.7 (high): 47; Claude Sonnet 5 (high): 47; Kimi K2.6 (always on): 44; GLM-5.3-Flash (max): 44; Kimi K2.7 Code (always on): 44; Gemma 4 26B A4B (thinking on): 43; Claude Haiku 4.5 (thinking 32k): 34; Mistral Small 3.1 24B (Non-reasoning): 33; DeepSeek V4 Pro (max): 0; GLM-5.2 (max): 0; gpt-oss-120b (high): 0; DeepSeek V4 Flash (max): 0; GLM-5.3 (max, default): 0; GLM-4.7-Flash (thinking on): 0; Llama 3.3 70B (Non-reasoning): 0; gpt-oss-20b (high): 0025507510059Claude CodeClaude Opus5.5 (xhigh)57Claude CodeClaude Sonnet5.5 (xhigh)56CodexGPT-6.1 Sol(xhigh)53Claude CodeClaude Opus 5(high)52Claude CodeClaude Haiku5.5 (xhigh)50CodexGPT-6 Luna(xhigh)50Claude CodeClaude Opus4.8 (high)47Claude CodeClaude Opus4.7 (high)47Claude CodeClaude Sonnet5 (high)44Kimi CodeKimi K2.6(always on)44Terminus-2GLM-5.3-Flash(max)44Kimi CodeKimi K2.7Code (alwayson)43Terminus-2Gemma 4 26BA4B (thinkingon)34Claude CodeClaude Haiku4.5 (thinking32k)33No agent entryMistral Small3.1 24B(Non-reasoning)0Terminus-2DeepSeek V4Pro (max)0Terminus-2GLM-5.2 (max)0Terminus-2gpt-oss-120b(high)0Terminus-2DeepSeek V4Flash (max)0Terminus-2GLM-5.3 (max,default)0Terminus-2GLM-4.7-Flash(thinking on)0Terminus-2Llama 3.3 70B(Non-reasoning)0Terminus-2gpt-oss-20b(high)
How to read this chart

Rows. Each model once, at its highest reasoning level with this group scored in full, so models not yet ranked in the Geospatial Intelligence Index can appear here.

Data table: Perception
Perception: score per model
ModelDeveloperPerceptionScored in
Claude Opus 5.5 (xhigh)Anthropic59Claude Code
Claude Sonnet 5.5 (xhigh)Anthropic57Claude Code
GPT-6.1 Sol (xhigh)OpenAI56Codex
Claude Opus 5 (high)Anthropic53Claude Code
Claude Haiku 5.5 (xhigh)Anthropic52Claude Code
GPT-6 Luna (xhigh)OpenAI50Codex
Claude Opus 4.8 (high)Anthropic50Claude Code
Claude Opus 4.7 (high)Anthropic47Claude Code
Claude Sonnet 5 (high)Anthropic47Claude Code
Kimi K2.6 (always on)Moonshot AI44Kimi Code
GLM-5.3-Flash (max)Z.ai44Terminus-2
Kimi K2.7 Code (always on)Moonshot AI44Kimi Code
Gemma 4 26B A4B (thinking on)Google43Terminus-2
Claude Haiku 4.5 (thinking 32k)Anthropic34Claude Code
Mistral Small 3.1 24B (Non-reasoning)Mistral AI33No agent entry
DeepSeek V4 Pro (max)DeepSeek0 (cannot take images)Terminus-2
GLM-5.2 (max)Z.ai0 (cannot take images)Terminus-2
gpt-oss-120b (high)OpenAI0 (cannot take images)Terminus-2
DeepSeek V4 Flash (max)DeepSeek0 (cannot take images)Terminus-2
GLM-5.3 (max, default)Z.ai0 (cannot take images)Terminus-2
GLM-4.7-Flash (thinking on)Z.ai0 (cannot take images)Terminus-2
Llama 3.3 70B (Non-reasoning)Meta0 (cannot take images)Terminus-2
gpt-oss-20b (high)OpenAI0 (cannot take images)Terminus-2

Tokens, cost and time

For the agent index tasks, in the harness each model is scored in; each model at its highest reasoning level. Time is compared through output tokens, as Artificial Analysis does; measured time per task is on each model page.

Cost per index task

Cost per index task

The average cost of one Agent Index attempt in the scored harness (USD) · Lower is better

  • Planned attempts still to come
Cost per index taskTerminus-2 – gpt-oss-120b (low): $0.0048 (partial coverage); Codex – GPT-6 Luna (xhigh): $0.0076 (partial coverage); Claude Code – Claude Haiku 5.5 (xhigh): $0.018 (partial coverage); Codex – GPT-6.1 Sol (xhigh): $0.091 (partial coverage); Claude Code – Claude Haiku 4.5 (thinking 32k): $0.100 (partial coverage); Terminus-2 – GLM-5.3-Flash (max): $0.106 (partial coverage); Terminus-2 – GLM-5.2 (none): $0.112 (partial coverage); Claude Code – Claude Sonnet 5.5 (xhigh): $0.123 (partial coverage); Claude Code – Claude Sonnet 5 (high): $0.229 (partial coverage); Claude Code – Claude Opus 5.5 (xhigh): $0.284 (partial coverage); Claude Code – Claude Opus 4.8 (high): $0.296 (partial coverage); Terminus-2 – DeepSeek V4 Pro (max): $0.341 (partial coverage); Claude Code – Claude Opus 5 (high): $0.357 (partial coverage); Claude Code – Claude Opus 4.7 (high): $0.382 (partial coverage)$0$0.100$0.200$0.300$0.400$0.0048Terminus-2gpt-oss-120b(low)$0.0076CodexGPT-6 Luna(xhigh)$0.018Claude CodeClaude Haiku5.5 (xhigh)$0.091CodexGPT-6.1 Sol(xhigh)$0.100Claude CodeClaude Haiku4.5 (thinking32k)$0.106Terminus-2GLM-5.3-Flash(max)$0.112Terminus-2GLM-5.2(none)$0.123Claude CodeClaude Sonnet5.5 (xhigh)$0.229Claude CodeClaude Sonnet5 (high)$0.284Claude CodeClaude Opus5.5 (xhigh)$0.296Claude CodeClaude Opus4.8 (high)$0.341Terminus-2DeepSeek V4Pro (max)$0.357Claude CodeClaude Opus 5(high)$0.382Claude CodeClaude Opus4.7 (high)
How to read this chart

Cost per index task. The average cost of one Agent Index attempt in the scored harness: one attempt at a Geospatial Agent Index task, at public API list prices, in the harness the model is scored in (the agents group). Quiz and image questions are one model call each and are not included.

Rows. Each model is shown at its highest reasoning level, as Artificial Analysis does; other levels are under 'Show every reasoning level'.

Data table: Cost per index task
Cost per index task
ModelCreatorCost per index taskCoverage (attempts)
Claude Code – Claude Opus 5.5 (xhigh)Anthropic$0.284565 of 859 planned attempts
Codex – GPT-6.1 Sol (xhigh)OpenAI$0.091565 of 859 planned attempts
Claude Code – Claude Sonnet 5.5 (xhigh)Anthropic$0.123565 of 859 planned attempts
Claude Code – Claude Opus 5 (high)Anthropic$0.357565 of 859 planned attempts
Claude Code – Claude Haiku 5.5 (xhigh)Anthropic$0.018565 of 859 planned attempts
Claude Code – Claude Opus 4.8 (high)Anthropic$0.296565 of 859 planned attempts
Codex – GPT-6 Luna (xhigh)OpenAI$0.0076565 of 859 planned attempts
Claude Code – Claude Opus 4.7 (high)Anthropic$0.382565 of 859 planned attempts
Claude Code – Claude Sonnet 5 (high)Anthropic$0.229565 of 859 planned attempts
Terminus-2 – GLM-5.3-Flash (max)Z.ai$0.106565 of 859 planned attempts
Terminus-2 – DeepSeek V4 Pro (max)DeepSeek$0.341565 of 859 planned attempts
Claude Code – Claude Haiku 4.5 (thinking 32k)Anthropic$0.100565 of 859 planned attempts
Terminus-2 – GLM-5.2 (none)Z.ai$0.112565 of 859 planned attempts
Terminus-2 – gpt-oss-120b (low)OpenAI$0.0048565 of 859 planned attempts

Output tokens per index task

Output tokens per index task

The average output tokens of one Agent Index attempt in the scored harness, reasoning included · Lower is better

  • Planned attempts still to come
Output tokens per index taskTerminus-2 – gpt-oss-120b (low): 2k (partial coverage); Codex – GPT-6.1 Sol (xhigh): 4k (partial coverage); Claude Code – Claude Opus 4.8 (high): 5k (partial coverage); Claude Code – Claude Sonnet 5.5 (xhigh): 5k (partial coverage); Claude Code – Claude Opus 4.7 (high): 5k (partial coverage); Claude Code – Claude Opus 5 (high): 5k (partial coverage); Claude Code – Claude Opus 5.5 (xhigh): 6k (partial coverage); Terminus-2 – GLM-5.2 (none): 6k (partial coverage); Codex – GPT-6 Luna (xhigh): 7k (partial coverage); Claude Code – Claude Haiku 4.5 (thinking 32k): 8k (partial coverage); Claude Code – Claude Sonnet 5 (high): 8k (partial coverage); Claude Code – Claude Haiku 5.5 (xhigh): 14k (partial coverage); Terminus-2 – DeepSeek V4 Pro (max): 24k (partial coverage); Terminus-2 – GLM-5.3-Flash (max): 34k (partial coverage)010k20k30k40k2kTerminus-2gpt-oss-120b(low)4kCodexGPT-6.1 Sol(xhigh)5kClaude CodeClaude Opus4.8 (high)5kClaude CodeClaude Sonnet5.5 (xhigh)5kClaude CodeClaude Opus4.7 (high)5kClaude CodeClaude Opus 5(high)6kClaude CodeClaude Opus5.5 (xhigh)6kTerminus-2GLM-5.2(none)7kCodexGPT-6 Luna(xhigh)8kClaude CodeClaude Haiku4.5 (thinking32k)8kClaude CodeClaude Sonnet5 (high)14kClaude CodeClaude Haiku5.5 (xhigh)24kTerminus-2DeepSeek V4Pro (max)34kTerminus-2GLM-5.3-Flash(max)
How to read this chart

Output tokens per task. The average number of tokens the model wrote per attempt, its reasoning included, as reported by the provider.

Why output tokens, not time. As Artificial Analysis does, the charts compare output tokens per task rather than time: an attempt's time depends on how fast the model is served, which should not decide a comparison, and how long the model waited for its replies is not recorded for every agent harness. Measured time per task stays on each model page and in the tables.

Compare with care. Models use different tokenisers, so the same text is a different number of tokens for different models. Use token counts to see how much work each model did, and cost to compare spending.

Rows. Each model is shown at its highest reasoning level, as Artificial Analysis does; other levels are under 'Show every reasoning level'.

Data table: Output tokens per index task
Output tokens per index task
ModelCreatorOutput tokens per index taskCoverage (attempts)
Claude Code – Claude Opus 5.5 (xhigh)Anthropic6k565 of 859 planned attempts
Codex – GPT-6.1 Sol (xhigh)OpenAI4k565 of 859 planned attempts
Claude Code – Claude Sonnet 5.5 (xhigh)Anthropic5k565 of 859 planned attempts
Claude Code – Claude Opus 5 (high)Anthropic5k565 of 859 planned attempts
Claude Code – Claude Haiku 5.5 (xhigh)Anthropic14k565 of 859 planned attempts
Claude Code – Claude Opus 4.8 (high)Anthropic5k565 of 859 planned attempts
Codex – GPT-6 Luna (xhigh)OpenAI7k565 of 859 planned attempts
Claude Code – Claude Opus 4.7 (high)Anthropic5k565 of 859 planned attempts
Claude Code – Claude Sonnet 5 (high)Anthropic8k565 of 859 planned attempts
Terminus-2 – GLM-5.3-Flash (max)Z.ai34k565 of 859 planned attempts
Terminus-2 – DeepSeek V4 Pro (max)DeepSeek24k565 of 859 planned attempts
Claude Code – Claude Haiku 4.5 (thinking 32k)Anthropic8k565 of 859 planned attempts
Terminus-2 – GLM-5.2 (none)Z.ai6k565 of 859 planned attempts
Terminus-2 – gpt-oss-120b (low)OpenAI2k565 of 859 planned attempts

Geospatial Intelligence Index vs. output tokens per index task

Higher and further left is better

  • Most attractive quadrant
  • Pareto line
Geospatial Intelligence Index vs. output tokens per index taskClaude Opus 5.5 (xhigh): Geospatial Intelligence Index 80.81, output tokens per index task 6k; GPT-6.1 Sol (xhigh): Geospatial Intelligence Index 78.68, output tokens per index task 4k; Claude Sonnet 5.5 (xhigh): Geospatial Intelligence Index 78.08, output tokens per index task 5k; Claude Opus 5 (high): Geospatial Intelligence Index 76.98, output tokens per index task 5k; Claude Haiku 5.5 (xhigh): Geospatial Intelligence Index 76.05, output tokens per index task 14k; Claude Opus 4.8 (high): Geospatial Intelligence Index 75.76, output tokens per index task 5k; GPT-6 Luna (xhigh): Geospatial Intelligence Index 74.65, output tokens per index task 7k; Claude Opus 4.7 (high): Geospatial Intelligence Index 74.58, output tokens per index task 5k; Claude Sonnet 5 (high): Geospatial Intelligence Index 73.92, output tokens per index task 8k; GLM-5.3-Flash (max): Geospatial Intelligence Index 66.82, output tokens per index task 34k; DeepSeek V4 Pro (max): Geospatial Intelligence Index 61.81, output tokens per index task 24k; Claude Haiku 4.5 (thinking 32k): Geospatial Intelligence Index 61.2, output tokens per index task 8k; GLM-5.2 (none): Geospatial Intelligence Index 56.66, output tokens per index task 6k; gpt-oss-120b (low): Geospatial Intelligence Index 40.57, output tokens per index task 2kMost attractive quadrant35404550556065707580851k2k5k10k20k50kOutput tokens per index task (log scale)Geospatial Intelligence IndexClaude Opus 5.5 (xhigh)GPT-6.1 Sol (xhigh)Claude Sonnet 5.5 (xhigh)Claude Opus 5 (high)Claude Haiku 5.5 (xhigh)Claude Opus 4.8 (high)GPT-6 Luna (xhigh)Claude Opus 4.7 (high)Claude Sonnet 5 (high)GLM-5.3-Flash (max)DeepSeek V4 Pro (max)Claude Haiku 4.5 (thinking 32k)GLM-5.2 (none)gpt-oss-120b (low)
How to read this chart

How to read this chart. Each point is one model at its highest reasoning level. Further left means a lower average output tokens per index task (log scale); higher means a higher Geospatial Intelligence Index. Both axes fit the models shown, so the score axis does not start at zero. The shaded most attractive quadrant is the top-left quarter of the plot, split at the middle of each axis: a reading aid, not a fixed bar. The dotted Pareto line joins the models that no more economical model beats on the index.

Output tokens per index task. The average number of tokens the model wrote per attempt, its reasoning included, as reported by the provider. For Agent Index attempts in the harness the model is scored in.

Why output tokens, not time. As Artificial Analysis does, the charts compare output tokens per task rather than time: an attempt's time depends on how fast the model is served, which should not decide a comparison, and how long the model waited for its replies is not recorded for every agent harness. Measured time per task stays on each model page and in the tables.

Compare with care. Models use different tokenisers, so the same text is a different number of tokens for different models. Use token counts to see how much work each model did, and cost to compare spending.

Data table: Geospatial Intelligence Index vs. output tokens per index task
Geospatial Intelligence Index vs. output tokens per index task
ModelGeospatial Intelligence IndexOutput tokens per index taskOn the Pareto line
Claude Opus 5.5 (xhigh)80.86kYes
GPT-6.1 Sol (xhigh)78.74kYes
Claude Sonnet 5.5 (xhigh)78.15kNo
Claude Opus 5 (high)77.05kNo
Claude Haiku 5.5 (xhigh)76.014kNo
Claude Opus 4.8 (high)75.85kNo
GPT-6 Luna (xhigh)74.77kNo
Claude Opus 4.7 (high)74.65kNo
Claude Sonnet 5 (high)73.98kNo
GLM-5.3-Flash (max)66.834kNo
DeepSeek V4 Pro (max)61.824kNo
Claude Haiku 4.5 (thinking 32k)61.28kNo
GLM-5.2 (none)56.76kNo
gpt-oss-120b (low)40.62kYes

Explore

Frequently asked questions

What is the Geospatial Intelligence Index?
It asks which AI model is best at geospatial (GIS and Earth observation) work overall. The default view shows each model once, at its highest reasoning level, labelled with that level, as Artificial Analysis does; "Show every reasoning level" lists every level with its own rank. The score is the weighted average of four groups: agents (3/9), coding (2/9), earth science (2/9), perception (2/9). The structure is a geospatial counterpart of Artificial Analysis's Intelligence Index. The weights are its weights for the groups that apply to geospatial work (agents 30, coding 20, scientific reasoning 20), with perception added at 20, the weight of its smaller groups, all scaled to 100. Its general-knowledge group is left out, because all the knowledge we test is Earth-science knowledge.
How is it different from the Geospatial Agent Index?
The Geospatial Agent Index asks which agent setup to use: each row is a model, an agent harness and a reasoning level, and it compares harnesses. The Intelligence Index asks which model is best: each model is scored once, in its developer's own harness where we run one, otherwise in Terminus-2, and the Geospatial Agent Index score is one of its groups. See the Geospatial Agent Index.
What does the Earth science group measure?
Earth science: knowledge now, reasoning coming. Knowledge is a closed-book quiz of 1,685 multiple-choice questions built on ESA's published question sets: one model call per question, no tools and no internet. Reasoning (spatial, and Earth-science derivations) will join the group as an equal average with knowledge, as a new version; the group weights do not change. See the knowledge results.
How is perception scored?
Earth observation is imagery, so image understanding is in the headline. Every model answers the same image questions in one call per question (look and answer). A model that cannot take images at all, according to its developer's documentation, scores 0, labelled "cannot take images"; a model that can take images but has not been run yet is not ranked. Each model has one run of the questions for now; the run count is shown. With 136 questions a score can move by about 4 points between runs. More runs and questions will follow. See the perception results.
What is the text-only view?
The same index without perception, with weights agents 3/7, coding 2/7, earth science 2/7. It compares models that cannot take images fairly, and shows how much perception changes the order.
Why are some models not ranked?
A model is ranked only when every group is scored in full: a first attempt on every agent and coding task, a quiz with every question asked and at most 1% left unscored by provider errors, and a full look-and-answer image run or a labelled "cannot take images" 0. Until then it is listed under partial coverage, with its group scores so far and what is still to come, and has no index.