Google's Gemini 3.8 Flash, at low reasoning, scores 86 out of 100 on GS-QA2, our benchmark of AI agents answering questions about places and terrain. Claude Opus 5.5 at the same low level also scores 86. Claude Sonnet 5.5 at low scores 76, ten points behind.
Gemini 3.8 Flash runs in Terminus-2, Terminal-Bench's open harness, our default. The Claude models run in Anthropic's own coding agent, Claude Code. Results as of 16 November 2026.
At low reasoning, Gemini 3.8 Flash is level with Claude Opus 5.5 on spatial questions
- Claude Opus 5.5 (low)Claude Code86
- Gemini 3.8 Flash (low)Terminus-286
- GLM-5.3 (low)Terminus-280
- Claude Sonnet 5.5 (low)Claude Code76
- GPT-5.6 Luna (low)Codex76
- GLM-5.3-Flash (low)Terminus-276
- Claude Haiku 5.5 (low)Claude Code75
- GPT-6 Luna (low)Codex59
- Gemini 3.1 Flash-Lite (low)Terminus-249
- gpt-oss-120b (low)Terminus-229
Which models we ran, and on what
We benchmark five Google models. Three have closed weights: Gemini 3.8 Flash, Gemini 3.1 Flash-Lite and Gemini 3.1 Pro (preview). Two are open weights: Gemma 4 26B A4B and Gemma 4 31B. Every Google entry runs in Terminus-2, with one scored attempt per task in a sandbox with no internet.
Gemini 3.1 Pro (preview), and Gemini 3.8 Flash at high reasoning, are still being benchmarked. Their results are provisional and not ranked, so this article leaves them out.
The headline score is the Geospatial Intelligence Index, a score out of 100 across three groups of work:
- Agent tasks (4 of 6 parts): Earth-Bench V2 (satellite and Earth observation data), GeoBenchX (GIS tasks that end in a map, a number or a list) and GS-QA2 (questions about places and terrain, answered by querying a database of OpenStreetMap features and elevation).
- Guided workflows (1 of 6): GeoAgentBench and GeoAnalystBench, GIS workflows where the task gives the steps and the model writes and runs the code.
- Geospatial knowledge (1 of 6): questions about GIS and Earth observation, answered with no tools.
41 models are ranked on the index. Gemini 3.8 Flash at low reasoning scores 79.6, the highest of the Google models.
Geospatial Intelligence Index, Text only (0 to 100, higher is better)
- Claude Sonnet 5.5 (xhigh)84.4
- Claude Opus 4.7 (high)80.8
- Gemini 3.8 Flash (low)Closed weights79.6
- DeepSeek V4 Flash (high)78.7
- Qwen 3.8 27B (xhigh)75.8
- Claude Haiku 4.5 (Reasoning)65.8
- Step 5 Preview (Reasoning)64.6
- Gemma 4 26B A4B (Reasoning)Open weights64.4
- Gemini 3.1 Flash-Lite (high)Closed weights63.2
- MiniMax M3.1 Flash (max)61.3
- gpt-oss-120b (high)51.9
Where Gemini 3.8 Flash is strong
Spatial questions over a database. On GS-QA2 the agent answers questions such as the nearest park to a landmark or the slope at a point, by writing queries against a PostGIS database. At low reasoning, Gemini 3.8 Flash is level with Claude Opus 5.5 and ahead of every other entry at that level. Among all entries in Terminus-2 at any level, none scores higher than its 86.
Following a given workflow. On GeoAgentBench it scores 98, against 96 for Claude Opus 5.5 at low. On GeoAnalystBench it scores 95, the same as Claude Sonnet 5.5 at low. Its guided workflow score is 96.
Knowledge. It scores 94 on geospatial knowledge. The best score of any model is 95.
Gemini 3.8 Flash against Claude Opus 5.5, both at low reasoning (0 to 100)
Gemini 3.8 Flash (low), Terminus-2Claude Opus 5.5 (low), Claude Code
- Earth-Bench V25459
- GeoBenchX7677
- GS-QA28686
- GeoAgentBench9896
- GeoAnalystBench95100
Where it falls behind, and why
Earth observation data. On Earth-Bench V2 it scores 54, against 59 for Claude Opus 5.5 and 56 for Claude Sonnet 5.5, both at low. Earth-Bench V2 has the agent work through satellite rasters: vegetation indices, land surface temperature, land cover. It is the one of the three agent benchmarks where Gemini 3.8 Flash at low is furthest from Claude Opus 5.5.
It refuses too often. GeoBenchX mixes solvable tasks with tasks that cannot be done with the data given, and a correct answer to those is to say so. Gemini 3.8 Flash says so on every impossible task: it scores 100 on them and answers none. Claude Opus 5.5 at low answers 23% of them anyway.
The cost of that caution shows on the solvable tasks. Gemini 3.8 Flash refuses 20% of them, where Claude Opus 5.5 refuses 1%. It scores 60 on solvable tasks against 77, so its total of 76 lands one point behind. If your work has many tasks that cannot be done as asked, that caution is useful. If most of your tasks can be done, it costs you answers.
Open and closed: Gemma and Flash-Lite
Google's open weights model Gemma 4 26B A4B scores 64.4 on the index. Gemini 3.1 Flash-Lite at high reasoning, a closed model, scores 63.2. They tie on agent tasks at 56 each.
They differ elsewhere. Gemma 4 26B A4B scores 86 on GeoAgentBench against 70 for Flash-Lite, and 75 on guided workflows against 64. Flash-Lite scores higher on Earth-Bench V2 (53 against 46). Gemma 4 31B, the larger open model, scores 55 on agent tasks; it is not on the index.
The fair rival for Gemma 4 26B A4B is Claude Haiku 4.5, which scores 65.8 on the index, 1.4 points more. Gemma 4 26B A4B costs under a quarter of Claude Haiku 4.5 per agent task, and as an open weights model you can run it on your own servers.
What it costs
Cost per task is what one attempt would cost at the provider's list prices, failed attempts included.
- Gemini 3.8 Flash at low costs about a third of Claude Opus 4.7 at high per agent task, 1.2 points behind it on the index.
- Against Claude Opus 5.5 at low, on agent tasks, it costs about a third more, for 2 points less (72 against 74).
- On GS-QA2, where they are level, it costs about a tenth more than Claude Opus 5.5 at low.
- Gemma 4 26B A4B costs about a fifth of Gemini 3.8 Flash per agent task, and Flash-Lite at high just over a quarter.
What this means if you are choosing a model
- Questions about places, terrain and spatial databases: Gemini 3.8 Flash at low reasoning is level with Claude Opus 5.5 at the same level. Put it on your shortlist.
- Satellite and Earth observation analysis: other models score higher on Earth-Bench V2. Compare them on your own data first.
- Tasks that must never invent an answer: Gemini 3.8 Flash refused every impossible GeoBenchX task. Check that it does not also refuse your solvable ones.
- An open model on your own hardware: Gemma 4 26B A4B comes within 1.4 points of Claude Haiku 4.5 on the index.
Every score for every model is on the Geospatial Intelligence Index. If you want to know how these models do on a task your team runs today, send it to us and we will benchmark it.
