DeepSeek V4.1 Flash (max) vs GPT-5.6 Luna (medium)
Results updated 31 October 2026 · In Claude Code and Codex, so part of any gap can come from the harness · How we benchmark
Verdict
DeepSeek V4.1 Flash (max) edges GPT-5.6 Luna (medium) by 2 points; GPT-5.6 Luna (medium) is cheaper and faster.
- Cost per task
- $0.20 vs $0.0094
- Time per task
- 10.9 min vs 0.8 min
- Tokens per attempt
- 1.58M vs 140k
- Image input
- Yes vs Yes
DeepSeek V4.1 Flash (max), Claude Code scores higher on Earth-Bench V2 and GS-QA2; GPT-5.6 Luna (medium), Codex scores higher on GeoBenchX. GPT-5.6 Luna (medium), Codex is cheaper and faster: $0.0094 against $0.20 per task, and 0.8 min against 10.9 min.
Where they differ most
By kind of work, share of tasks passed, 0 to 100 · Higher is better
| Kind of work | DeepSeek V4.1 Flash (max) | GPT-5.6 Luna (medium) |
|---|---|---|
| GIS analysis | 92 | 76 |
| Climate and time series | 64 | 54 |
| Spatial statistics | 31 | 38 |
| Making maps | 72 | 79 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: DeepSeek V4.1 Flash (max), Claude Code · GPT-5.6 Luna (medium), Codex.
Side by side
| Metric | DeepSeek V4.1 Flash (max), Claude Code | GPT-5.6 Luna (medium), Codex |
|---|---|---|
| Geospatial Agent Index: Text only | 68 | 66 |
| Geospatial Agent Index: Text + vision | No data | 67 |
| Earth-Bench V2 score | 58 | 56 |
| GeoBenchX score | 55 | 63 |
| GS-QA2 score | 89 | 80 |
| Vision agents score | No data | 69 |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 96 | 98 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 100 | 100 |
| OGD4All score (standalone; in no index) | 79 | No data |
| GeoNatureAgent score (standalone; in no index) | 100 | No data |
| MultiGlobeQA score (standalone; in no index) | 64 | No data |
| GeoAgentEval score (standalone; in no index) | 84 | No data |
| Remote sensing and raster analysis score | 75 | 77 |
| GIS analysis score | 92 | 76 |
| Climate and time series score | 64 | 54 |
| Spatial statistics score | 31 | 38 |
| Making maps score | 72 | 79 |
| Recognising infeasible requests score | 40 | 47 |
| Cost per task | $0.20 | $0.0094 |
| Time per task (at 100 output tokens/s) | 10.9 min | 0.8 min |
| Tokens per attempt | 1.58M | 140k |
| Turns per attempt | 26.0 | 9.4 |
| Input price per 1M tokens | $0.30 | $0.20 |
| Output price per 1M tokens | $1.20 | $1.20 |
| Context window | 1,048,576 | 1,050,000 |
| Image input | Yes | Yes |
| Developer | DeepSeek | OpenAI |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, DeepSeek V4.1 Flash (max), Claude Code or GPT-5.6 Luna (medium), Codex?
DeepSeek V4.1 Flash (max), Claude Code, 68 to 66.
Which is cheaper per task, DeepSeek V4.1 Flash (max), Claude Code or GPT-5.6 Luna (medium), Codex?
GPT-5.6 Luna (medium), Codex, $0.0094 to $0.20.
Which writes fewer output tokens per task, DeepSeek V4.1 Flash (max), Claude Code or GPT-5.6 Luna (medium), Codex?
GPT-5.6 Luna (medium), Codex, 3k to 48k.
Which has the larger context window, DeepSeek V4.1 Flash (max), Claude Code or GPT-5.6 Luna (medium), Codex?
GPT-5.6 Luna (medium), Codex, 1,050,000 tokens to 1,048,576 tokens.