Qwen 3.8 27B (Reasoning) vs GPT-6 Luna (high)
Results updated 24 October 2026 · In Qwen Code and Codex, so part of any gap can come from the harness · How we benchmark
Verdict
GPT-6 Luna (high) edges Qwen 3.8 27B (Reasoning) by 2 points, and is cheaper and faster.
- Cost per task
- $0.22 vs $0.0040
- Time per task
- 4.7 min (estimated) vs 0.7 min
- Tokens per attempt
- 840k vs 120k
- Image input
- Yes vs Yes
Qwen 3.8 27B (Reasoning), Qwen Code scores higher on Earth-Bench V2; GPT-6 Luna (high), Codex scores higher on GeoBenchX and GS-QA2. GPT-6 Luna (high), Codex is cheaper and faster: $0.0040 against $0.22 per task, and 0.7 min against 4.7 min.
Where they differ most
By kind of work, share of tasks passed, 0 to 100 · Higher is better
| Kind of work | Qwen 3.8 27B (Reasoning) | GPT-6 Luna (high) |
|---|---|---|
| Making maps | 65 | 79 |
| Recognising infeasible requests | 62 | 75 |
| Remote sensing and raster analysis | 77 | 73 |
| Climate and time series | 63 | 66 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: Qwen 3.8 27B (Reasoning), Qwen Code · GPT-6 Luna (high), Codex.
Side by side
| Metric | Qwen 3.8 27B (Reasoning), Qwen Code | GPT-6 Luna (high), Codex |
|---|---|---|
| Geospatial Agent Index: Text only | 69 | 71 |
| Geospatial Agent Index: Text + vision | No data | 70 |
| Earth-Bench V2 score | 69 | 63 |
| GeoBenchX score | 63 | 74 |
| GS-QA2 score | 75 | 77 |
| Vision agents score | No data | 67 |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 94 | 96 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 79 | 100 |
| OGD4All score (standalone; in no index) | 97 | No data |
| GeoNatureAgent score (standalone; in no index) | 100 | No data |
| MultiGlobeQA score (standalone; in no index) | 74 | No data |
| GeoAgentEval score (standalone; in no index) | 88 | No data |
| ImpactMesh score (standalone; in no index) | No data | No data |
| Remote sensing and raster analysis score | 77 | 73 |
| GIS analysis score | 75 | 77 |
| Climate and time series score | 63 | 66 |
| Spatial statistics score | 54 | 54 |
| Making maps score | 65 | 79 |
| Recognising infeasible requests score | 62 | 75 |
| Cost per task | $0.22 | $0.0040 |
| Time per task (at 100 output tokens/s) | 4.7 min (estimated) | 0.7 min |
| Tokens per attempt | 840k | 120k |
| Turns per attempt | 17.6 | 7.4 |
| Input price per 1M tokens | $0.45 | $0.10 |
| Output price per 1M tokens | $3.20 | $0.50 |
| Context window | 262,144 | 1,050,000 |
| Image input | Yes | Yes |
| Developer | Alibaba (Qwen) | OpenAI |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Qwen 3.8 27B (Reasoning), Qwen Code or GPT-6 Luna (high), Codex?
GPT-6 Luna (high), Codex, 71 to 69.
Which is cheaper per task, Qwen 3.8 27B (Reasoning), Qwen Code or GPT-6 Luna (high), Codex?
GPT-6 Luna (high), Codex, $0.0040 to $0.22.
Which writes fewer output tokens per task, Qwen 3.8 27B (Reasoning), Qwen Code or GPT-6 Luna (high), Codex?
GPT-6 Luna (high), Codex, 3k to 28k.
Which has the larger context window, Qwen 3.8 27B (Reasoning), Qwen Code or GPT-6 Luna (high), Codex?
GPT-6 Luna (high), Codex, 1,050,000 tokens to 262,144 tokens.