Mistral Large 4 (high) vs Qwen 3.8 27B (xhigh)
Results updated 3 November 2026 · In Terminus-2 and Qwen Code, so part of any gap can come from the harness · How we benchmark
Verdict
Qwen 3.8 27B (xhigh) beats Mistral Large 4 (high) by 9 points, and is faster; Mistral Large 4 (high) is cheaper.
- Cost per task
- $0.10 vs $0.22
- Time per task
- 12.0 min vs 4.7 min (estimated)
- Tokens per attempt
- 226k vs 840k
- Image input
- Yes vs Yes
Mistral Large 4 (high), Terminus-2 scores higher on GS-QA2; Qwen 3.8 27B (xhigh), Qwen Code scores higher on Earth-Bench V2 and GeoBenchX. Mistral Large 4 (high), Terminus-2 is cheaper ($0.10 against $0.22 per task), but Qwen 3.8 27B (xhigh), Qwen Code is faster (4.7 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) | Qwen 3.8 27B (xhigh) |
|---|---|---|
| Recognising infeasible requests | 23 | 60 |
| Climate and time series | 40 | 61 |
| Making maps | 82 | 63 |
| Remote sensing and raster analysis | 67 | 76 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: Mistral Large 4 (high), Terminus-2 · Qwen 3.8 27B (xhigh), Qwen Code.
Side by side
| Metric | Mistral Large 4 (high), Terminus-2 | Qwen 3.8 27B (xhigh), Qwen Code |
|---|---|---|
| Geospatial Agent Index: Text only | 59 | 68 |
| Geospatial Agent Index: Text + vision | No data | No data |
| Earth-Bench V2 score | 46 | 68 |
| GeoBenchX score | 55 | 62 |
| GS-QA2 score | 77 | 75 |
| Vision agents score | No data | No data |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 80 | 94 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 94 | 79 |
| OGD4All score (standalone; in no index) | 96 | 95 |
| GeoNatureAgent score (standalone; in no index) | 100 | 100 |
| MultiGlobeQA score (standalone; in no index) | 62 | 74 |
| GeoAgentEval score (standalone; in no index) | 60 | 88 |
| Remote sensing and raster analysis score | 67 | 76 |
| GIS analysis score | 82 | 75 |
| Climate and time series score | 40 | 61 |
| Spatial statistics score | 46 | 54 |
| Making maps score | 82 | 63 |
| Recognising infeasible requests score | 23 | 60 |
| Cost per task | $0.10 | $0.22 |
| Time per task (at 100 output tokens/s) | 12.0 min | 4.7 min (estimated) |
| Tokens per attempt | 226k | 840k |
| Turns per attempt | 13.8 | 17.6 |
| Input price per 1M tokens | $0.70 | $0.45 |
| Output price per 1M tokens | $2.10 | $3.20 |
| Context window | 262,144 | 262,144 |
| Image input | Yes | Yes |
| Developer | Mistral AI | Alibaba (Qwen) |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Mistral Large 4 (high), Terminus-2 or Qwen 3.8 27B (xhigh), Qwen Code?
Qwen 3.8 27B (xhigh), Qwen Code, 68 to 59.
Which is cheaper per task, Mistral Large 4 (high), Terminus-2 or Qwen 3.8 27B (xhigh), Qwen Code?
Mistral Large 4 (high), Terminus-2, $0.10 to $0.22.
Which writes fewer output tokens per task, Mistral Large 4 (high), Terminus-2 or Qwen 3.8 27B (xhigh), Qwen Code?
Mistral Large 4 (high), Terminus-2, 14k to 28k.
Which has the larger context window, Mistral Large 4 (high), Terminus-2 or Qwen 3.8 27B (xhigh), Qwen Code?
They are level, at 262,144 tokens.