Step 5 Preview (Reasoning) vs DeepSeek V4 Flash (high)
Results updated 3 November 2026 · In Terminus-2 and Claude Code, so part of any gap can come from the harness · How we benchmark
Verdict
DeepSeek V4 Flash (high) beats Step 5 Preview (Reasoning) by 5 points, and is cheaper and faster.
- Cost per task
- $0.13 vs $0.080
- Time per task
- 13.1 min (estimated) vs 5.4 min
- Tokens per attempt
- 311k vs 634k
- Image input
- Yes vs No
Step 5 Preview (Reasoning), Terminus-2 scores higher on Earth-Bench V2 and GS-QA2; DeepSeek V4 Flash (high), Claude Code scores higher on GeoBenchX. DeepSeek V4 Flash (high), Claude Code is cheaper and faster: $0.080 against $0.13 per task, and 5.4 min against 13.1 min.
Where they differ most
By kind of work, share of tasks passed, 0 to 100 · Higher is better
| Kind of work | Step 5 Preview (Reasoning) | DeepSeek V4 Flash (high) |
|---|---|---|
| Making maps | 21 | 68 |
| Spatial statistics | 0 | 38 |
| GIS analysis | 82 | 62 |
| Recognising infeasible requests | 23 | 35 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: Step 5 Preview (Reasoning), Terminus-2 · DeepSeek V4 Flash (high), Claude Code.
Side by side
| Metric | Step 5 Preview (Reasoning), Terminus-2 | DeepSeek V4 Flash (high), Claude Code |
|---|---|---|
| Geospatial Agent Index: Text only | 51 | 56 |
| Geospatial Agent Index: Text + vision | No data | No image input |
| Earth-Bench V2 score | 56 | 50 |
| GeoBenchX score | 18 | 50 |
| GS-QA2 score | 79 | 69 |
| Vision agents score | No data | No image input |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 84 | 84 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 89 | 89 |
| OGD4All score (standalone; in no index) | 93 | 70 |
| GeoNatureAgent score (standalone; in no index) | 88 | 71 |
| MultiGlobeQA score (standalone; in no index) | 70 | 64 |
| GeoAgentEval score (standalone; in no index) | 91 | 80 |
| Remote sensing and raster analysis score | 62 | 67 |
| GIS analysis score | 82 | 62 |
| Climate and time series score | 54 | 53 |
| Spatial statistics score | 0 | 38 |
| Making maps score | 21 | 68 |
| Recognising infeasible requests score | 23 | 35 |
| Cost per task | $0.13 | $0.080 |
| Time per task (at 100 output tokens/s) | 13.1 min (estimated) | 5.4 min |
| Tokens per attempt | 311k | 634k |
| Turns per attempt | 15.1 | 16.1 |
| Input price per 1M tokens | $1.00 | $0.44 |
| Output price per 1M tokens | $2.70 | $1.32 |
| Context window | 1,024,000 | 1,048,576 |
| Image input | Yes | No |
| Developer | StepFun | DeepSeek |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Step 5 Preview (Reasoning), Terminus-2 or DeepSeek V4 Flash (high), Claude Code?
DeepSeek V4 Flash (high), Claude Code, 56 to 51.
Which is cheaper per task, Step 5 Preview (Reasoning), Terminus-2 or DeepSeek V4 Flash (high), Claude Code?
DeepSeek V4 Flash (high), Claude Code, $0.080 to $0.13.
Which writes fewer output tokens per task, Step 5 Preview (Reasoning), Terminus-2 or DeepSeek V4 Flash (high), Claude Code?
DeepSeek V4 Flash (high), Claude Code, 19k to 24k.
Which has the larger context window, Step 5 Preview (Reasoning), Terminus-2 or DeepSeek V4 Flash (high), Claude Code?
DeepSeek V4 Flash (high), Claude Code, 1,048,576 tokens to 1,024,000 tokens.