Terminus-2: geospatial agent results

Harbor's stock terminal agent, open source. The model works in a terminal session: it reads the screen and types commands. It is the harness for models whose developer has no agent of its own. An agent is a harness plus a model: this page ranks every model we benchmark in Terminus-2 on GIS and Earth observation tasks, and shows how the same model does in other harnesses.

Made by Harbor, open source · Version 2.0.0, pinned · 27 models from 12 labs · 38 entries, 34 ranked

Overview

Best on the Geospatial Agent Index

Geospatial Agent Index, Text only

Lowest cost

#1gpt-oss-120b (low)$0.0046
#2Gemini 3.1 Flash-Lite (minimal)$0.0066
#3Nemotron 3.5 Super VL 120B (Reasoning)$0.0096
#4Nemotron 3.5 Lightning 30B (Reasoning)$0.013
#5gpt-oss-20b (high)$0.017

Cost per agent task, at list prices

Fastest

#1gpt-oss-120b (low)0.5 min
#2Gemini 3.1 Flash-Lite (minimal)0.6 min
#3Mistral Large 3 (Non-reasoning)1.1 min
#4GLM-5.2 (Non-reasoning)1.5 min
#5Gemini 3.1 Flash-Lite (low)1.7 min

Time per agent task, at 100 output tokens/s

In Terminus-2, the best entry is Kimi K3 (Reasoning), 25th of 83 entries on the Geospatial Agent Index, at 73.1.
11 of the models in Terminus-2 also run in another harness at the same reasoning level. Against the best other harness, Terminus-2 scores higher for 2, lower for 1 and within noise for 8.
Still being benchmarked (provisional): Claude Sonnet 4.6 (high), Gemini 3.1 Pro (preview) (high), Gemini 3.1 Pro (preview) (low), Gemini 3.8 Flash (high).

Models

One per model

Geospatial Agent Index: every model in Terminus-2

Text only, 0 to 100 · Higher is better · one bar per model, at its best ranked reasoning level

How to read this chart

Bars. Every entry in Terminus-2 with a result, highlighted. Provisional: not yet ranked.

Data table: Geospatial Agent Index: every model in Terminus-2

Geospatial Agent Index: every model in Terminus-2
Model	Developer	Geospatial Agent Index: every model in Terminus-2
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	73
GLM-5.3 (max), Terminus-2	Z.ai	72
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	72
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	72
Gemini 3.8 Flash (low), Terminus-2	Google	72
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	72
GLM-5.2 (max), Terminus-2	Z.ai	72
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	70
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	68
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	67
Muse Glimmer 30B (high), Terminus-2	Meta	64
GLM-5.3-Flash (low), Terminus-2	Z.ai	63
Mistral Large 4 (high), Terminus-2	Mistral AI	63
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	56
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	56
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	53
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	52
Step 5 Preview (Reasoning), Terminus-2	StepFun	52
gpt-oss-120b (high), Terminus-2	OpenAI	46
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	43
gpt-oss-20b (high), Terminus-2	OpenAI	33
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	33
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	21
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	13
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	8

Every reasoning level

Geospatial Agent Index: every entry in Terminus-2

Text only, 0 to 100 · Higher is better · every reasoning level

How to read this chart

Bars. Every entry in Terminus-2 with a result, highlighted. Provisional: not yet ranked.

Data table: Geospatial Agent Index: every entry in Terminus-2

Geospatial Agent Index: every entry in Terminus-2
Model	Developer	Geospatial Agent Index: every entry in Terminus-2
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	73
GLM-5.3 (max), Terminus-2	Z.ai	72
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	72
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	72
Gemini 3.8 Flash (low), Terminus-2	Google	72
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	72
GLM-5.2 (max), Terminus-2	Z.ai	72
GLM-5.2 (high), Terminus-2	Z.ai	71
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	70
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	70
DeepSeek V4 Flash (max), Terminus-2	DeepSeek	69
GLM-5.3 (low), Terminus-2	Z.ai	68
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	68
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	67
Muse Glimmer 30B (high), Terminus-2	Meta	64
GLM-5.2 (Non-reasoning), Terminus-2	Z.ai	63
GLM-5.3-Flash (low), Terminus-2	Z.ai	63
GLM-5.3-Flash (max), Terminus-2	Z.ai	63
Mistral Large 4 (high), Terminus-2	Mistral AI	63
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	56
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	56
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	53
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	52
Step 5 Preview (Reasoning), Terminus-2	StepFun	52
Gemini 3.1 Flash-Lite (low), Terminus-2	Google	48
gpt-oss-120b (high), Terminus-2	OpenAI	46
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	43
gpt-oss-20b (high), Terminus-2	OpenAI	33
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	33
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Google	32
gpt-oss-120b (low), Terminus-2	OpenAI	31
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	21
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	13
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	8

Provisional, so not on the index chart: Gemini 3.1 Pro (preview) (high), Gemini 3.8 Flash (high), Claude Sonnet 4.6 (high), Gemini 3.1 Pro (preview) (low). Their finished evaluations are in the charts below.

Same model elsewhere

The models we benchmark in Terminus-2 and in another harness, at the same reasoning level: an agent is a harness plus a model, so the gap between the bars is the harness. Terminus-2 is highlighted; gaps under about 3 points are within noise.

gpt-oss-20b (high)

higher is better

Data table: gpt-oss-20b (high)

gpt-oss-20b (high)
Model	Developer	gpt-oss-20b (high)
gpt-oss-20b (high), Terminus-2	OpenAI	33
gpt-oss-20b (high), Codex	OpenAI	9

DeepSeek V4 Flash (high)

higher is better

Data table: DeepSeek V4 Flash (high)

DeepSeek V4 Flash (high)
Model	Developer	DeepSeek V4 Flash (high)
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	72
DeepSeek V4 Flash (high), Codex	DeepSeek	66
DeepSeek V4 Flash (high), Claude Code	DeepSeek	59

gpt-oss-120b (high)

higher is better

Data table: gpt-oss-120b (high)

gpt-oss-120b (high)
Model	Developer	gpt-oss-120b (high)
gpt-oss-120b (high), Terminus-2	OpenAI	46
gpt-oss-120b (high), Codex	OpenAI	43

GLM-4.7-Flash (Reasoning)

higher is better

Data table: GLM-4.7-Flash (Reasoning)

GLM-4.7-Flash (Reasoning)
Model	Developer	GLM-4.7-Flash (Reasoning)
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	43
GLM-4.7-Flash (Reasoning), Claude Code	Z.ai	41

Kimi K2.6 (Reasoning)

higher is better

Data table: Kimi K2.6 (Reasoning)

Kimi K2.6 (Reasoning)
Model	Developer	Kimi K2.6 (Reasoning)
Kimi K2.6 (Reasoning), Kimi Code	Moonshot AI	70
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	70

Kimi K2.7 Code (Reasoning)

higher is better

Data table: Kimi K2.7 Code (Reasoning)

Kimi K2.7 Code (Reasoning)
Model	Developer	Kimi K2.7 Code (Reasoning)
Kimi K2.7 Code (Reasoning), Kimi Code	Moonshot AI	72
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	72

GLM-5.2 (max)

higher is better

Data table: GLM-5.2 (max)

GLM-5.2 (max)
Model	Developer	GLM-5.2 (max)
GLM-5.2 (max), Claude Code	Z.ai	74
GLM-5.2 (max), Terminus-2	Z.ai	72

Kimi K3 (Reasoning)

higher is better

Data table: Kimi K3 (Reasoning)

Kimi K3 (Reasoning)
Model	Developer	Kimi K3 (Reasoning)
Kimi K3 (Reasoning), Kimi Code	Moonshot AI	75
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	73

GLM-5.3 (max)

higher is better

Data table: GLM-5.3 (max)

GLM-5.3 (max)
Model	Developer	GLM-5.3 (max)
GLM-5.3 (max), Claude Code	Z.ai	74
GLM-5.3 (max), Terminus-2	Z.ai	72

DeepSeek V4 Pro (high)

higher is better

Data table: DeepSeek V4 Pro (high)

DeepSeek V4 Pro (high)
Model	Developer	DeepSeek V4 Pro (high)
DeepSeek V4 Pro (high), Codex	DeepSeek	73
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	70

GLM-5.3-Flash (max)

higher is better

Data table: GLM-5.3-Flash (max)

GLM-5.3-Flash (max)
Model	Developer	GLM-5.3-Flash (max)
GLM-5.3-Flash (max), Claude Code	Z.ai	72
GLM-5.3-Flash (max), Terminus-2	Z.ai	63

By evaluation

Earth-Bench V2

higher is better

Striped: provisional, not yet ranked

Data table: Earth-Bench V2

Earth-Bench V2
Model	Developer	Earth-Bench V2
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	75
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	73
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	70
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	69
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	69
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	68
GLM-5.2 (high), Terminus-2	Z.ai	68
GLM-5.2 (max), Terminus-2	Z.ai	68
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	68
DeepSeek V4 Flash (max), Terminus-2	DeepSeek	67
GLM-5.3 (max), Terminus-2	Z.ai	67
GLM-5.3 (low), Terminus-2	Z.ai	62
Muse Glimmer 30B (high), Terminus-2	Meta	60
Step 5 Preview (Reasoning), Terminus-2	StepFun	60
Claude Sonnet 4.6 (high), Terminus-2	Anthropic	59
Gemini 3.1 Pro (preview) (high), Terminus-2	Google	59
GLM-5.2 (Non-reasoning), Terminus-2	Z.ai	58
GLM-5.3-Flash (low), Terminus-2	Z.ai	56
Gemini 3.8 Flash (low), Terminus-2	Google	54
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	54
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	53
GLM-5.3-Flash (max), Terminus-2	Z.ai	51
gpt-oss-120b (high), Terminus-2	OpenAI	50
Mistral Large 4 (high), Terminus-2	Mistral AI	50
Gemini 3.8 Flash (high), Terminus-2	Google	48
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	46
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	46
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	44
Gemini 3.1 Flash-Lite (low), Terminus-2	Google	43
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	41
Gemini 3.1 Pro (preview) (low), Terminus-2	Google	41
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Google	34
gpt-oss-20b (high), Terminus-2	OpenAI	32
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	31
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	31
gpt-oss-120b (low), Terminus-2	OpenAI	28
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	16
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	9

GeoBenchX

higher is better

Data table: GeoBenchX

GeoBenchX
Model	Developer	GeoBenchX
Gemini 3.8 Flash (low), Terminus-2	Google	76
Gemini 3.1 Pro (preview) (low), Terminus-2	Google	74
Gemini 3.1 Pro (preview) (high), Terminus-2	Google	73
Gemini 3.8 Flash (high), Terminus-2	Google	68
GLM-5.3 (max), Terminus-2	Z.ai	68
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	67
GLM-5.2 (max), Terminus-2	Z.ai	67
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	67
GLM-5.2 (high), Terminus-2	Z.ai	67
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	65
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	65
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	65
GLM-5.3 (low), Terminus-2	Z.ai	64
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	64
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	64
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	63
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	63
DeepSeek V4 Flash (max), Terminus-2	DeepSeek	62
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	61
Muse Glimmer 30B (high), Terminus-2	Meta	61
GLM-5.2 (Non-reasoning), Terminus-2	Z.ai	60
Mistral Large 4 (high), Terminus-2	Mistral AI	59
GLM-5.3-Flash (low), Terminus-2	Z.ai	58
GLM-5.3-Flash (max), Terminus-2	Z.ai	57
Claude Sonnet 4.6 (high), Terminus-2	Anthropic	57
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	53
Gemini 3.1 Flash-Lite (low), Terminus-2	Google	51
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	46
gpt-oss-120b (high), Terminus-2	OpenAI	45
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Google	44
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	42
gpt-oss-20b (high), Terminus-2	OpenAI	38
gpt-oss-120b (low), Terminus-2	OpenAI	37
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	29
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	23
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	20
Step 5 Preview (Reasoning), Terminus-2	StepFun	18
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	6

GS-QA2

higher is better

Data table: GS-QA2

GS-QA2
Model	Developer	GS-QA2
Gemini 3.8 Flash (low), Terminus-2	Google	86
GLM-5.3 (max), Terminus-2	Z.ai	83
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	82
GLM-5.2 (max), Terminus-2	Z.ai	80
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	80
GLM-5.3 (low), Terminus-2	Z.ai	80
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	80
GLM-5.3-Flash (max), Terminus-2	Z.ai	80
DeepSeek V4 Flash (max), Terminus-2	DeepSeek	80
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	79
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	79
Step 5 Preview (Reasoning), Terminus-2	StepFun	79
Mistral Large 4 (high), Terminus-2	Mistral AI	79
GLM-5.2 (high), Terminus-2	Z.ai	78
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	77
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	77
GLM-5.3-Flash (low), Terminus-2	Z.ai	76
GLM-5.2 (Non-reasoning), Terminus-2	Z.ai	72
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	71
Muse Glimmer 30B (high), Terminus-2	Meta	70
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	70
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	59
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	58
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	55
Gemini 3.1 Flash-Lite (low), Terminus-2	Google	49
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	43
gpt-oss-120b (high), Terminus-2	OpenAI	42
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	39
gpt-oss-20b (high), Terminus-2	OpenAI	30
gpt-oss-120b (low), Terminus-2	OpenAI	29
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Google	18
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	10
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	9
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	4

GeoAgentBench

higher is better

Data table: GeoAgentBench

GeoAgentBench
Model	Developer	GeoAgentBench
Gemini 3.8 Flash (low), Terminus-2	Google	98
Claude Sonnet 4.6 (high), Terminus-2	Anthropic	96
DeepSeek V4 Flash (max), Terminus-2	DeepSeek	96
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	96
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	96
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	96
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	94
GLM-5.2 (high), Terminus-2	Z.ai	94
GLM-5.2 (max), Terminus-2	Z.ai	94
GLM-5.3-Flash (low), Terminus-2	Z.ai	94
GLM-5.3 (max), Terminus-2	Z.ai	94
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	94
GLM-5.2 (Non-reasoning), Terminus-2	Z.ai	92
GLM-5.3 (low), Terminus-2	Z.ai	92
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	92
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	90
Gemini 3.8 Flash (high), Terminus-2	Google	88
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	88
Gemini 3.1 Pro (preview) (high), Terminus-2	Google	86
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	86
Gemini 3.1 Pro (preview) (low), Terminus-2	Google	84
GLM-5.3-Flash (max), Terminus-2	Z.ai	84
Step 5 Preview (Reasoning), Terminus-2	StepFun	84
Mistral Large 4 (high), Terminus-2	Mistral AI	80
Muse Glimmer 30B (high), Terminus-2	Meta	79
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	70
gpt-oss-120b (high), Terminus-2	OpenAI	70
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	68
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	62
gpt-oss-120b (low), Terminus-2	OpenAI	60
gpt-oss-20b (high), Terminus-2	OpenAI	53
Gemini 3.1 Flash-Lite (low), Terminus-2	Google	52
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	46
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Google	42
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	42
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	28
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	10
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	8

GeoAnalystBench

higher is better

Striped: provisional, not yet ranked

Data table: GeoAnalystBench

GeoAnalystBench
Model	Developer	GeoAnalystBench
Claude Sonnet 4.6 (high), Terminus-2	Anthropic	100
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	100
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	100
GLM-5.2 (high), Terminus-2	Z.ai	100
DeepSeek V4 Flash (max), Terminus-2	DeepSeek	95
Gemini 3.1 Pro (preview) (high), Terminus-2	Google	95
Gemini 3.8 Flash (low), Terminus-2	Google	95
GLM-5.2 (max), Terminus-2	Z.ai	95
GLM-5.3-Flash (low), Terminus-2	Z.ai	95
GLM-5.3 (low), Terminus-2	Z.ai	95
Mistral Large 4 (high), Terminus-2	Mistral AI	94
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	89
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	89
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	89
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	89
Step 5 Preview (Reasoning), Terminus-2	StepFun	89
GLM-5.3-Flash (max), Terminus-2	Z.ai	84
GLM-5.3 (max), Terminus-2	Z.ai	84
Gemini 3.1 Pro (preview) (low), Terminus-2	Google	79
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	79
GLM-5.2 (Non-reasoning), Terminus-2	Z.ai	74
Gemini 3.1 Flash-Lite (low), Terminus-2	Google	63
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	63
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	63
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	58
Gemini 3.8 Flash (high), Terminus-2	Google	58
Muse Glimmer 30B (high), Terminus-2	Meta	58
gpt-oss-120b (high), Terminus-2	OpenAI	53
gpt-oss-120b (low), Terminus-2	OpenAI	47
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	47
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	47
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Google	42
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	37
gpt-oss-20b (high), Terminus-2	OpenAI	37
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	37
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	32
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	16
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	0

OGD4All

higher is better

Data table: OGD4All

OGD4All
Model	Developer	OGD4All
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	96
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	96
Muse Glimmer 30B (high), Terminus-2	Meta	95
Mistral Large 4 (high), Terminus-2	Mistral AI	95
Step 5 Preview (Reasoning), Terminus-2	StepFun	94
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	92
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	91
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	90
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	47

GeoNatureAgent

higher is better

Data table: GeoNatureAgent

GeoNatureAgent
Model	Developer	GeoNatureAgent
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	100
Mistral Large 4 (high), Terminus-2	Mistral AI	100
Muse Glimmer 30B (high), Terminus-2	Meta	100
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	100
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	92
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	92
Step 5 Preview (Reasoning), Terminus-2	StepFun	88
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	64
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	0

MultiGlobeQA

higher is better

Data table: MultiGlobeQA

MultiGlobeQA
Model	Developer	MultiGlobeQA
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	74
Step 5 Preview (Reasoning), Terminus-2	StepFun	73
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	73
Muse Glimmer 30B (high), Terminus-2	Meta	72
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	72
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	72
Mistral Large 4 (high), Terminus-2	Mistral AI	71
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	68
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	48

GeoAgentEval

higher is better

Data table: GeoAgentEval

GeoAgentEval
Model	Developer	GeoAgentEval
Mistral Large 4 (high), Terminus-2	Mistral AI	92
Muse Glimmer 30B (high), Terminus-2	Meta	92
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	92
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	92
Step 5 Preview (Reasoning), Terminus-2	StepFun	91
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	83
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	80
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	48
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	36

Cost

Cost per task

Average cost per agent task (USD) · Lower is better

Striped: provisional, not yet ranked

How to read this chart

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. Lower is better. Costs are pay-per-token list prices, not any discount or credit a customer might have.

Bars. Every entry in Terminus-2 with a result, highlighted. Provisional: not yet ranked.

Data table: Cost per task

Cost per task
Model	Developer	Cost per task
gpt-oss-120b (low), Terminus-2	OpenAI	$0.0046
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Google	$0.0066
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	$0.0096
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	$0.013
gpt-oss-20b (high), Terminus-2	OpenAI	$0.017
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	$0.018
Gemini 3.1 Flash-Lite (low), Terminus-2	Google	$0.018
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	$0.021
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	$0.023
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	$0.026
gpt-oss-120b (high), Terminus-2	OpenAI	$0.026
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	$0.031
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	$0.037
GLM-5.3-Flash (low), Terminus-2	Z.ai	$0.048
Gemini 3.1 Pro (preview) (low), Terminus-2	Google	$0.057
DeepSeek V4 Flash (max), Terminus-2	DeepSeek	$0.062
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	$0.065
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	$0.067
Muse Glimmer 30B (high), Terminus-2	Meta	$0.078
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	$0.089
GLM-5.3-Flash (max), Terminus-2	Z.ai	$0.094
GLM-5.2 (Non-reasoning), Terminus-2	Z.ai	$0.099
Mistral Large 4 (high), Terminus-2	Mistral AI	$0.10
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	$0.11
Gemini 3.8 Flash (low), Terminus-2	Google	$0.11
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	$0.12
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	$0.13
Step 5 Preview (Reasoning), Terminus-2	StepFun	$0.13
GLM-5.2 (high), Terminus-2	Z.ai	$0.22
Gemini 3.1 Pro (preview) (high), Terminus-2	Google	$0.22
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	$0.23
Claude Sonnet 4.6 (high), Terminus-2	Anthropic	$0.25
GLM-5.3 (low), Terminus-2	Z.ai	$0.27
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	$0.30
Gemini 3.8 Flash (high), Terminus-2	Google	$0.31
GLM-5.2 (max), Terminus-2	Z.ai	$0.32
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	$0.35
GLM-5.3 (max), Terminus-2	Z.ai	$0.38

Geospatial Agent Index vs. cost per task

Geospatial Agent Index, Text only; cost per task (USD) · Higher and further left is better

Most attractive quadrant
Pareto line

How to read this chart

Why a log scale. Costs and times differ by a factor of ten or more between models, so equal steps on the axis are equal ratios.

How to read this chart. Each point is one entry. 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
Model	Geospatial Agent Index	Cost per task (USD)	On the Pareto line
DeepSeek V4 Flash (max), Terminus-2	69	$0.062	Yes
DeepSeek V4 Flash (high), Terminus-2	72	$0.065	Yes
DeepSeek V4 Pro (max), Terminus-2	72	$0.35	No
DeepSeek V4 Pro (high), Terminus-2	70	$0.30	No
Gemini 3.1 Flash-Lite (low), Terminus-2	48	$0.018	No
Gemini 3.1 Flash-Lite (minimal), Terminus-2	32	$0.0066	Yes
Gemini 3.1 Flash-Lite (high), Terminus-2	56	$0.031	Yes
Gemini 3.8 Flash (low), Terminus-2	72	$0.11	No
Gemma 4 26B A4B (Reasoning), Terminus-2	56	$0.021	Yes
GLM-4.7-Flash (Reasoning), Terminus-2	43	$0.037	No
GLM-5.2 (high), Terminus-2	71	$0.22	No
GLM-5.2 (Non-reasoning), Terminus-2	63	$0.099	No
GLM-5.2 (max), Terminus-2	72	$0.32	No
GLM-5.3-Flash (low), Terminus-2	63	$0.048	Yes
GLM-5.3-Flash (max), Terminus-2	63	$0.094	No
GLM-5.3 (low), Terminus-2	68	$0.27	No
GLM-5.3 (max), Terminus-2	72	$0.38	No
gpt-oss-120b (low), Terminus-2	31	$0.0046	Yes
gpt-oss-120b (high), Terminus-2	46	$0.026	No
gpt-oss-20b (high), Terminus-2	33	$0.017	No
Kimi K2.6 (Reasoning), Terminus-2	70	$0.089	No
Kimi K2.7 Code (Reasoning), Terminus-2	72	$0.11	No
Kimi K3 (Reasoning), Terminus-2	73	$0.23	Yes
Llama 3.3 70B (Non-reasoning), Terminus-2	8	$0.023	No
MiniMax M3.1 Flash (max), Terminus-2	53	$0.018	Yes
Mistral Large 3 (Non-reasoning), Terminus-2	33	$0.026	No
Mistral Large 4 (high), Terminus-2	63	$0.10	No
Muse Glimmer 30B (high), Terminus-2	64	$0.078	No
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	21	$0.13	No
Nemotron 3 Ultra (high), Terminus-2	68	$0.067	No
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	13	$0.013	No
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	52	$0.0096	Yes
Qwen 3.8 27B (xhigh), Terminus-2	67	$0.12	No
Step 5 Preview (Reasoning), Terminus-2	52	$0.13	No

Time

Time per task

Average minutes per agent task, at 100 output tokens/s · Lower is better

Striped: provisional, not yet ranked

How to read this chart

What this metric means. Time as if every model wrote 100 output tokens per second: output tokens divided by 100, plus the time the agent's commands took. Serving speed differs between providers and is not ranked.

Bars. Every entry in Terminus-2 with a result, highlighted. Provisional: not yet ranked.

Data table: Time per task

Time per task
Model	Developer	Time per task (at 100 output tokens/s)
gpt-oss-120b (low), Terminus-2	OpenAI	0.5 min
Gemini 3.1 Pro (preview) (low), Terminus-2	Google	0.5 min
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Google	0.6 min
Mistral Large 3 (Non-reasoning), Terminus-2	Mistral AI	1.1 min
GLM-5.2 (Non-reasoning), Terminus-2	Z.ai	1.5 min
Gemini 3.1 Flash-Lite (low), Terminus-2	Google	1.7 min
Gemini 3.8 Flash (low), Terminus-2	Google	1.9 min
Llama 3.3 70B (Non-reasoning), Terminus-2	Meta	2.1 min
Nemotron 3 Ultra (high), Terminus-2	NVIDIA	2.4 min
Gemini 3.1 Pro (preview) (high), Terminus-2	Google	2.5 min
GLM-5.3 (low), Terminus-2	Z.ai	2.7 min
Kimi K3 (Reasoning), Terminus-2	Moonshot AI	2.7 min
gpt-oss-120b (high), Terminus-2	OpenAI	3.0 min
Muse Glimmer 30B (high), Terminus-2	Meta	3.0 min
Gemini 3.1 Flash-Lite (high), Terminus-2	Google	3.1 min
Claude Sonnet 4.6 (high), Terminus-2	Anthropic	3.2 min
Kimi K2.6 (Reasoning), Terminus-2	Moonshot AI	3.5 min
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	NVIDIA	3.8 min
Kimi K2.7 Code (Reasoning), Terminus-2	Moonshot AI	4.0 min
DeepSeek V4 Pro (high), Terminus-2	DeepSeek	4.0 min
GLM-5.3 (max), Terminus-2	Z.ai	4.2 min
GLM-5.3-Flash (low), Terminus-2	Z.ai	4.2 min
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	NVIDIA	4.6 min
GLM-5.2 (high), Terminus-2	Z.ai	4.7 min
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	NVIDIA	4.8 min
DeepSeek V4 Pro (max), Terminus-2	DeepSeek	4.9 min
GLM-4.7-Flash (Reasoning), Terminus-2	Z.ai	5.4 min
Gemma 4 26B A4B (Reasoning), Terminus-2	Google	5.6 min
DeepSeek V4 Flash (max), Terminus-2	DeepSeek	5.7 min
DeepSeek V4 Flash (high), Terminus-2	DeepSeek	5.9 min
GLM-5.2 (max), Terminus-2	Z.ai	6.1 min
Gemini 3.8 Flash (high), Terminus-2	Google	6.7 min
gpt-oss-20b (high), Terminus-2	OpenAI	7.2 min
Qwen 3.8 27B (xhigh), Terminus-2	Alibaba (Qwen)	7.4 min
GLM-5.3-Flash (max), Terminus-2	Z.ai	8.5 min
MiniMax M3.1 Flash (max), Terminus-2	MiniMax	8.7 min
Mistral Large 4 (high), Terminus-2	Mistral AI	11.4 min
Step 5 Preview (Reasoning), Terminus-2	StepFun	13.1 min

Geospatial Agent Index vs. time per task

Geospatial Agent Index, Text only; time per task (minutes) · Higher and further left is better

Most attractive quadrant
Pareto line

How to read this chart

Why a log scale. Costs and times differ by a factor of ten or more between models, so equal steps on the axis are equal ratios.

How to read this chart. Each point is one entry. Further left means a lower average time 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 faster shown model beats on score.

Data table: Geospatial Agent Index vs. time per task

Geospatial Agent Index vs. time per task
Model	Geospatial Agent Index	Time per task (minutes)	On the Pareto line
DeepSeek V4 Flash (max), Terminus-2	69	5.7 min	No
DeepSeek V4 Flash (high), Terminus-2	72	5.9 min	No
DeepSeek V4 Pro (max), Terminus-2	72	4.9 min	No
DeepSeek V4 Pro (high), Terminus-2	70	4.0 min	No
Gemini 3.1 Flash-Lite (low), Terminus-2	48	1.7 min	No
Gemini 3.1 Flash-Lite (minimal), Terminus-2	32	0.6 min	Yes
Gemini 3.1 Flash-Lite (high), Terminus-2	56	3.1 min	No
Gemini 3.8 Flash (low), Terminus-2	72	1.9 min	Yes
Gemma 4 26B A4B (Reasoning), Terminus-2	56	5.6 min	No
GLM-4.7-Flash (Reasoning), Terminus-2	43	5.4 min	No
GLM-5.2 (high), Terminus-2	71	4.7 min	No
GLM-5.2 (Non-reasoning), Terminus-2	63	1.5 min	Yes
GLM-5.2 (max), Terminus-2	72	6.1 min	No
GLM-5.3-Flash (low), Terminus-2	63	4.2 min	No
GLM-5.3-Flash (max), Terminus-2	63	8.5 min	No
GLM-5.3 (low), Terminus-2	68	2.7 min	No
GLM-5.3 (max), Terminus-2	72	4.2 min	No
gpt-oss-120b (low), Terminus-2	31	0.5 min	Yes
gpt-oss-120b (high), Terminus-2	46	3.0 min	No
gpt-oss-20b (high), Terminus-2	33	7.2 min	No
Kimi K2.6 (Reasoning), Terminus-2	70	3.5 min	No
Kimi K2.7 Code (Reasoning), Terminus-2	72	4.0 min	No
Kimi K3 (Reasoning), Terminus-2	73	2.7 min	Yes
Llama 3.3 70B (Non-reasoning), Terminus-2	8	2.1 min	No
MiniMax M3.1 Flash (max), Terminus-2	53	8.7 min	No
Mistral Large 3 (Non-reasoning), Terminus-2	33	1.1 min	Yes
Mistral Large 4 (high), Terminus-2	63	11.4 min	No
Muse Glimmer 30B (high), Terminus-2	64	3.0 min	No
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	21	4.6 min	No
Nemotron 3 Ultra (high), Terminus-2	68	2.4 min	No
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	13	4.8 min	No
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	52	3.8 min	No
Qwen 3.8 27B (xhigh), Terminus-2	67	7.4 min	No
Step 5 Preview (Reasoning), Terminus-2	52	13.1 min	No

Every benchmark

Score by benchmark

Share answered correctly, 0 to 100 · Darker is better; hatched: provisional, evaluation not finished

Earth-Bench V2 images

ThinkGeo images

Geospatial knowledge

Kimi K3 (Reasoning), Terminus-2

–

GLM-5.3 (max), Terminus-2

–

DeepSeek V4 Flash (high), Terminus-2

–

DeepSeek V4 Pro (max), Terminus-2

100

–

Gemini 3.8 Flash (low), Terminus-2

–

Kimi K2.7 Code (Reasoning), Terminus-2

–

GLM-5.2 (max), Terminus-2

–

GLM-5.2 (high), Terminus-2

100

–

DeepSeek V4 Pro (high), Terminus-2

100

–

Kimi K2.6 (Reasoning), Terminus-2

–

DeepSeek V4 Flash (max), Terminus-2

–

GLM-5.3 (low), Terminus-2

–

Nemotron 3 Ultra (high), Terminus-2

–

Qwen 3.8 27B (xhigh), Terminus-2

100

Muse Glimmer 30B (high), Terminus-2

100

–

GLM-5.2 (Non-reasoning), Terminus-2

–

GLM-5.3-Flash (low), Terminus-2

–

GLM-5.3-Flash (max), Terminus-2

–

Mistral Large 4 (high), Terminus-2

100

–

Gemini 3.1 Flash-Lite (high), Terminus-2

–

Gemma 4 26B A4B (Reasoning), Terminus-2

–

MiniMax M3.1 Flash (max), Terminus-2

100

–

Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2

–

Step 5 Preview (Reasoning), Terminus-2

–

Gemini 3.1 Flash-Lite (low), Terminus-2

–

gpt-oss-120b (high), Terminus-2

–

GLM-4.7-Flash (Reasoning), Terminus-2

–

gpt-oss-20b (high), Terminus-2

–

Mistral Large 3 (Non-reasoning), Terminus-2

–

Gemini 3.1 Flash-Lite (minimal), Terminus-2

–

gpt-oss-120b (low), Terminus-2

–

Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2

–

Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2

–

Llama 3.3 70B (Non-reasoning), Terminus-2

–

Gemini 3.1 Pro (preview) (high), Terminus-2 provisional

–

Gemini 3.8 Flash (high), Terminus-2 provisional

–

Claude Sonnet 4.6 (high), Terminus-2 provisional

–

100

–

Gemini 3.1 Pro (preview) (low), Terminus-2 provisional

–

Data table: Score by benchmark

Every benchmark
Model	Status	Earth-Bench V2	GeoBenchX	GS-QA2	GeoAgentBench	GeoAnalystBench	OGD4All	GeoNatureAgent	MultiGlobeQA	GeoAgentEval	Earth-Bench V2 images	ThinkGeo images	Geospatial knowledge
Kimi K3 (Reasoning), Terminus-2	Ranked	75	64	80	94	89	No data	No data	No data	No data	66	30	92
GLM-5.3 (max), Terminus-2	Ranked	67	68	83	94	84	No data	No data	No data	No data	No data	No data	92
DeepSeek V4 Flash (high), Terminus-2	Ranked	73	65	79	94	89	No data	No data	No data	No data	No data	No data	91
DeepSeek V4 Pro (max), Terminus-2	Ranked	69	67	80	96	100	No data	No data	No data	No data	No data	No data	92
Gemini 3.8 Flash (low), Terminus-2	Ranked	54	76	86	98	95	No data	No data	No data	No data	66	34	94
Kimi K2.7 Code (Reasoning), Terminus-2	Ranked	70	63	82	92	89	No data	No data	No data	No data	56	31	92
GLM-5.2 (max), Terminus-2	Ranked	68	67	80	94	95	No data	No data	No data	No data	No data	No data	89
GLM-5.2 (high), Terminus-2	Ranked	68	67	78	94	100	No data	No data	No data	No data	No data	No data	91
DeepSeek V4 Pro (high), Terminus-2	Ranked	68	65	77	96	100	No data	No data	No data	No data	No data	No data	92
Kimi K2.6 (Reasoning), Terminus-2	Ranked	68	64	77	90	79	No data	No data	No data	No data	60	28	92
DeepSeek V4 Flash (max), Terminus-2	Ranked	67	62	80	96	95	No data	No data	No data	No data	No data	No data	91
GLM-5.3 (low), Terminus-2	Ranked	62	64	80	92	95	No data	No data	No data	No data	No data	No data	91
Nemotron 3 Ultra (high), Terminus-2	Ranked	69	65	70	96	63	92	92	72	92	No data	No data	91
Qwen 3.8 27B (xhigh), Terminus-2	Ranked	54	67	79	62	89	96	100	74	92	69	37	91
Muse Glimmer 30B (high), Terminus-2	Ranked	60	61	70	79	58	95	100	72	92	No data	No data	No data
GLM-5.2 (Non-reasoning), Terminus-2	Ranked	58	60	72	92	74	No data	No data	No data	No data	No data	No data	89
GLM-5.3-Flash (low), Terminus-2	Ranked	56	58	76	94	95	No data	No data	No data	No data	56	36	85
GLM-5.3-Flash (max), Terminus-2	Ranked	51	57	80	84	84	No data	No data	No data	No data	53	35	92
Mistral Large 4 (high), Terminus-2	Ranked	50	59	79	80	94	95	100	71	92	No data	No data	No data
Gemini 3.1 Flash-Lite (high), Terminus-2	Ranked	53	61	55	70	58	No data	No data	No data	No data	52	24	90
Gemma 4 26B A4B (Reasoning), Terminus-2	Ranked	46	63	58	86	63	No data	No data	No data	No data	64	22	89
MiniMax M3.1 Flash (max), Terminus-2	Ranked	41	46	71	88	47	96	100	72	83	No data	No data	89
Nemotron 3.5 Super VL 120B (Reasoning), Terminus-2	Ranked	46	53	59	68	37	91	92	73	80	No data	No data	92
Step 5 Preview (Reasoning), Terminus-2	Ranked	60	18	79	84	89	94	88	73	91	No data	No data	92
Gemini 3.1 Flash-Lite (low), Terminus-2	Ranked	43	51	49	52	63	No data	No data	No data	No data	54	15	90
gpt-oss-120b (high), Terminus-2	Ranked	50	45	42	70	53	No data	No data	No data	No data	No data	No data	90
GLM-4.7-Flash (Reasoning), Terminus-2	Ranked	44	42	43	42	37	No data	No data	No data	No data	No data	No data	86
gpt-oss-20b (high), Terminus-2	Ranked	32	38	30	53	37	No data	No data	No data	No data	No data	No data	85
Mistral Large 3 (Non-reasoning), Terminus-2	Ranked	31	29	39	46	47	No data	No data	No data	No data	52	15	89
Gemini 3.1 Flash-Lite (minimal), Terminus-2	Ranked	34	44	18	42	42	No data	No data	No data	No data	56	26	89
gpt-oss-120b (low), Terminus-2	Ranked	28	37	29	60	47	No data	No data	No data	No data	No data	No data	88
Nemotron 3 Nano Omni 30B (Reasoning), Terminus-2	Ranked	31	23	9	28	16	47	0	48	48	No data	No data	85
Nemotron 3.5 Lightning 30B (Reasoning), Terminus-2	Ranked	16	20	4	10	32	90	64	68	36	No data	No data	90
Llama 3.3 70B (Non-reasoning), Terminus-2	Ranked	9	6	10	8	0	No data	No data	No data	No data	No data	No data	82
Gemini 3.1 Pro (preview) (high), Terminus-2	Provisional	59 (provisional)	73	No data	86	95	No data	No data	No data	No data	No data	No data	No data
Gemini 3.8 Flash (high), Terminus-2	Provisional	48 (provisional)	68	No data	88	58	No data	No data	No data	No data	No data	No data	No data
Claude Sonnet 4.6 (high), Terminus-2	Provisional	59 (provisional)	57	No data	96	100	No data	No data	No data	No data	No data	No data	98
Gemini 3.1 Pro (preview) (low), Terminus-2	Provisional	41 (provisional)	74	No data	84	79	No data	No data	No data	No data	No data	No data	No data

About Terminus-2

Made by: Harbor, open source
What it is: Harbor's stock terminal agent, open source. The model works in a terminal session: it reads the screen and types commands. It is the harness for models whose developer has no agent of its own.
Version: 2.0.0, the same pinned build for every run
Models we run in it: 27 from Moonshot AI, Z.ai, DeepSeek, Google, NVIDIA, Alibaba (Qwen), Meta, Mistral AI, MiniMax, StepFun, OpenAI, Anthropic
How we run it: The same tasks, limits and checks as every other harness: only the agent program changes. Methodology.

Frequently asked questions

Which models do you benchmark in Terminus-2?

27 models: Kimi K3 (Reasoning), GLM-5.3 (max), DeepSeek V4 Flash (high), DeepSeek V4 Pro (max), Gemini 3.8 Flash (low), Kimi K2.7 Code (Reasoning), GLM-5.2 (max), Kimi K2.6 (Reasoning), Nemotron 3 Ultra (high), Qwen 3.8 27B (xhigh), Muse Glimmer 30B (high), GLM-5.3-Flash (low), Mistral Large 4 (high), Gemini 3.1 Flash-Lite (high), Gemma 4 26B A4B (Reasoning), MiniMax M3.1 Flash (max), Nemotron 3.5 Super VL 120B (Reasoning), Step 5 Preview (Reasoning), gpt-oss-120b (high), GLM-4.7-Flash (Reasoning), gpt-oss-20b (high), Mistral Large 3 (Non-reasoning), Nemotron 3 Nano Omni 30B (Reasoning), Nemotron 3.5 Lightning 30B (Reasoning), Llama 3.3 70B (Non-reasoning), Gemini 3.1 Pro (preview) (high) and Claude Sonnet 4.6 (high).

Which model is best in Terminus-2?

Kimi K3 (Reasoning), with 73.1 on the Geospatial Agent Index, 25th of 83 entries.

What is the cheapest model in Terminus-2?

gpt-oss-120b (low), at $0.0046 per agent task at list prices.

What is the fastest model in Terminus-2?

gpt-oss-120b (low), at 0.5 min per agent task.

Is Terminus-2 better than other harnesses?

It depends on the model. gpt-oss-20b (high) scores 33.5 in Terminus-2 and 9.3 in Codex; GLM-5.3-Flash (max) scores 72.1 in Claude Code and 62.7 in Terminus-2. Gaps under about 3 points are within noise.

Who makes Terminus-2?

Harbor, open source. Harbor's stock terminal agent, open source. The model works in a terminal session: it reads the screen and types commands. It is the harness for models whose developer has no agent of its own.

Which version of Terminus-2 do you run?

2.0.0, pinned: every run uses the same build.

Are any models in Terminus-2 still being benchmarked?

Yes. Gemini 3.1 Pro (preview) (high); Gemini 3.8 Flash (high); Claude Sonnet 4.6 (high); Gemini 3.1 Pro (preview) (low) are still being benchmarked, so the results are provisional: not yet ranked.