Mistral Large 3 (Non-reasoning), Terminus-2 vs. DeepSeek V4 Pro (high), Codex
Mistral Large 3 (Non-reasoning), Terminus-2 against DeepSeek V4 Pro (high), Codex: Geospatial Agent Index, each evaluation, cost, time and tokens, on the same tasks and limits. The harnesses differ, so part of any gap can come from the harness. Methodology.
Side by side
| Metric | Mistral Large 3 (Non-reasoning), Terminus-2 | DeepSeek V4 Pro (high), Codex |
|---|---|---|
| Geospatial Agent Index: Text only | 33, not every task attempted yet | 62, not every task attempted yet (116 of 488 tasks attempted) |
| Geospatial Agent Index: Text + vision | Still to run: GS-QA2, Vision agents | No image input |
| Earth-Bench V2 score | 31 | 57 |
| GeoBenchX score | 29 | 66 |
| GS-QA2 score | 39 | No data |
| Vision agents score | No data | No image input |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 46 | No data |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 47 | No data |
| Raster and remote sensing score | 29 | 62 |
| Vector and overlay score | 41 | 0 |
| Climate and time series score | 34 | 42 |
| Spatial statistics and interpolation score | 8 | 17 |
| Mapping and cartography score | 33 | 85 |
| Recognising infeasible requests score | 33 | 65 |
| Cost per task | $0.026 | $0.64 |
| Time per task (at 100 output tokens/s) | 1.1 min | 4.9 min |
| Tokens per attempt | 43k | 828k |
| Turns per attempt | 5.2 | 20.8 |
| Attempts scored | 487 | 116 |
| Input price per 1M tokens | $0.50 | $1.32 |
| Output price per 1M tokens | $1.50 | $3.96 |
| Context window | 256,000 | 1,048,576 |
| Image input | Yes | No |
| Developer | Mistral AI | DeepSeek |
* Not every task attempted yet: the score may still change. What the asterisk means.
Charts
Not yet ranked. Mistral Large 3 (Non-reasoning), Terminus-2: 220 of 221 GS-QA2 tasks attempted; DeepSeek V4 Pro (high), Codex: 28 of 117 Earth-Bench V2 tasks; 88 of 150 GeoBenchX tasks; GS-QA2 not started attempted. Grey placeholders mark these entries in the index, cost and output token charts; their height is not a value. Bars appear once every task has an attempt. Results so far are in the side-by-side table.
Geospatial Agent Index
Equal average of Earth-Bench V2, GeoBenchX and GS-QA2, 0 to 100 · Higher is better
Top eight. Every model is in the table.
Data table: Geospatial Agent Index
| Model | Developer | Geospatial Agent Index | Status | Evaluations scored | Tasks attempted |
|---|
Cost per task
Average cost per task (USD) · Lower is better
Top eight. Every model is in the table.
How to read this chart
Not shown: not yet ranked. 2 entries without an attempt on every index task. What is still to run is on the models page.
Data table: Cost per task
| Model | Developer | Cost per task | Tasks attempted |
|---|
Output tokens per task
Average output tokens per attempt, reasoning included · Lower is better
Top eight. Every model is in the table.
How to read this chart
Not shown: not yet ranked. 2 entries without an attempt on every index task. What is still to run is on the models page.
Data table: Output tokens per task
| Model | Developer | Output tokens per task | Tasks attempted |
|---|
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Mistral Large 3 (Non-reasoning), Terminus-2 or DeepSeek V4 Pro (high), Codex?
Which is cheaper per task, Mistral Large 3 (Non-reasoning), Terminus-2 or DeepSeek V4 Pro (high), Codex?
Which writes fewer output tokens per task, Mistral Large 3 (Non-reasoning), Terminus-2 or DeepSeek V4 Pro (high), Codex?
Which has the larger context window, Mistral Large 3 (Non-reasoning), Terminus-2 or DeepSeek V4 Pro (high), Codex?
More: Mistral Large 3 (Non-reasoning), Terminus-2 · DeepSeek V4 Pro (high), Codex · All comparisons