DeepSeek V4 Flash (high) vs GPT-6 Luna (medium)
Results updated 30 October 2026 · In Claude Code and Codex, so part of any gap can come from the harness · How we benchmark
Verdict
GPT-6 Luna (medium) edges DeepSeek V4 Flash (high) by 1 point, and is cheaper and faster.
- Cost per task
- $0.080 vs $0.0026
- Time per task
- 5.4 min vs 0.4 min
- Tokens per attempt
- 634k vs 86k
- Image input
- No vs Yes
DeepSeek V4 Flash (high), Claude Code scores higher on GS-QA2; GPT-6 Luna (medium), Codex scores higher on GeoBenchX; they are level on Earth-Bench V2. GPT-6 Luna (medium), Codex is cheaper and faster: $0.0026 against $0.080 per task, and 0.4 min against 5.4 min.
Where they differ most
By kind of work, share of tasks passed, 0 to 100 · Higher is better
| Kind of work | DeepSeek V4 Flash (high) | GPT-6 Luna (medium) |
|---|---|---|
| Spatial statistics | 38 | 23 |
| Recognising infeasible requests | 35 | 43 |
| Making maps | 68 | 75 |
| Climate and time series | 53 | 47 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: DeepSeek V4 Flash (high), Claude Code · GPT-6 Luna (medium), Codex.
Side by side
| Metric | DeepSeek V4 Flash (high), Claude Code | GPT-6 Luna (medium), Codex |
|---|---|---|
| Geospatial Agent Index: Text only | 56 | 57 |
| Geospatial Agent Index: Text + vision | No image input | 59 |
| Earth-Bench V2 score | 50 | 50 |
| GeoBenchX score | 50 | 57 |
| GS-QA2 score | 69 | 65 |
| Vision agents score | No image input | 64 |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 84 | 94 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 89 | 89 |
| OGD4All score (standalone; in no index) | 70 | No data |
| GeoNatureAgent score (standalone; in no index) | 71 | No data |
| MultiGlobeQA score (standalone; in no index) | 64 | No data |
| GeoAgentEval score (standalone; in no index) | 80 | No data |
| Remote sensing and raster analysis score | 67 | 63 |
| GIS analysis score | 62 | 65 |
| Climate and time series score | 53 | 47 |
| Spatial statistics score | 38 | 23 |
| Making maps score | 68 | 75 |
| Recognising infeasible requests score | 35 | 43 |
| Cost per task | $0.080 | $0.0026 |
| Time per task (at 100 output tokens/s) | 5.4 min | 0.4 min |
| Tokens per attempt | 634k | 86k |
| Turns per attempt | 16.1 | 6.4 |
| Input price per 1M tokens | $0.44 | $0.10 |
| Output price per 1M tokens | $1.32 | $0.50 |
| Context window | 1,048,576 | 1,050,000 |
| Image input | No | Yes |
| Developer | DeepSeek | OpenAI |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, DeepSeek V4 Flash (high), Claude Code or GPT-6 Luna (medium), Codex?
GPT-6 Luna (medium), Codex, 57 to 56.
Which is cheaper per task, DeepSeek V4 Flash (high), Claude Code or GPT-6 Luna (medium), Codex?
GPT-6 Luna (medium), Codex, $0.0026 to $0.080.
Which writes fewer output tokens per task, DeepSeek V4 Flash (high), Claude Code or GPT-6 Luna (medium), Codex?
GPT-6 Luna (medium), Codex, 1k to 19k.
Which has the larger context window, DeepSeek V4 Flash (high), Claude Code or GPT-6 Luna (medium), Codex?
GPT-6 Luna (medium), Codex, 1,050,000 tokens to 1,048,576 tokens.