Mistral Large 4 (high) vs GPT-5.6 Luna (low)
Results updated 3 November 2026 · In Terminus-2 and Codex, so part of any gap can come from the harness · How we benchmark
Verdict
GPT-5.6 Luna (low) edges Mistral Large 4 (high) by 2 points, and is cheaper and faster.
- Cost per task
- $0.10 vs $0.0058
- Time per task
- 12.0 min vs 0.5 min
- Tokens per attempt
- 226k vs 84k
- Image input
- Yes vs Yes
Mistral Large 4 (high), Terminus-2 scores higher on Earth-Bench V2 and GS-QA2; GPT-5.6 Luna (low), Codex scores higher on GeoBenchX. GPT-5.6 Luna (low), Codex is cheaper and faster: $0.0058 against $0.10 per task, and 0.5 min against 12.0 min.
Where they differ most
By kind of work, share of tasks passed, 0 to 100 · Higher is better
| Kind of work | Mistral Large 4 (high) | GPT-5.6 Luna (low) |
|---|---|---|
| Recognising infeasible requests | 23 | 50 |
| Spatial statistics | 46 | 23 |
| GIS analysis | 82 | 72 |
| Remote sensing and raster analysis | 67 | 70 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: Mistral Large 4 (high), Terminus-2 · GPT-5.6 Luna (low), Codex.
Side by side
| Metric | Mistral Large 4 (high), Terminus-2 | GPT-5.6 Luna (low), Codex |
|---|---|---|
| Geospatial Agent Index: Text only | 59 | 61 |
| Geospatial Agent Index: Text + vision | No data | 62 |
| Earth-Bench V2 score | 46 | 44 |
| GeoBenchX score | 55 | 62 |
| GS-QA2 score | 77 | 76 |
| Vision agents score | No data | 64 |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 80 | 86 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 94 | 95 |
| OGD4All score (standalone; in no index) | 96 | No data |
| GeoNatureAgent score (standalone; in no index) | 100 | No data |
| MultiGlobeQA score (standalone; in no index) | 62 | No data |
| GeoAgentEval score (standalone; in no index) | 60 | No data |
| Remote sensing and raster analysis score | 67 | 70 |
| GIS analysis score | 82 | 72 |
| Climate and time series score | 40 | 42 |
| Spatial statistics score | 46 | 23 |
| Making maps score | 82 | 81 |
| Recognising infeasible requests score | 23 | 50 |
| Cost per task | $0.10 | $0.0058 |
| Time per task (at 100 output tokens/s) | 12.0 min | 0.5 min |
| Tokens per attempt | 226k | 84k |
| Turns per attempt | 13.8 | 7.0 |
| Input price per 1M tokens | $0.70 | $0.20 |
| Output price per 1M tokens | $2.10 | $1.20 |
| Context window | 262,144 | 1,050,000 |
| Image input | Yes | Yes |
| Developer | Mistral AI | OpenAI |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Mistral Large 4 (high), Terminus-2 or GPT-5.6 Luna (low), Codex?
GPT-5.6 Luna (low), Codex, 61 to 59.
Which is cheaper per task, Mistral Large 4 (high), Terminus-2 or GPT-5.6 Luna (low), Codex?
GPT-5.6 Luna (low), Codex, $0.0058 to $0.10.
Which writes fewer output tokens per task, Mistral Large 4 (high), Terminus-2 or GPT-5.6 Luna (low), Codex?
GPT-5.6 Luna (low), Codex, 2k to 14k.
Which has the larger context window, Mistral Large 4 (high), Terminus-2 or GPT-5.6 Luna (low), Codex?
GPT-5.6 Luna (low), Codex, 1,050,000 tokens to 262,144 tokens.