Nemotron 3 Ultra (high) vs DeepSeek V4.1 Flash (max)
Results updated 31 October 2026 · In Terminus-2 and Claude Code, so part of any gap can come from the harness · How we benchmark
Verdict
Nemotron 3 Ultra (high) and DeepSeek V4.1 Flash (max) are level on the Geospatial Agent Index, at 68.
- Cost per task
- $0.067 vs $0.20
- Time per task
- 2.4 min vs 10.9 min
- Tokens per attempt
- 144k vs 1.58M
- Image input
- No vs Yes
Nemotron 3 Ultra (high), Terminus-2 scores higher on Earth-Bench V2 and GeoBenchX; DeepSeek V4.1 Flash (max), Claude Code scores higher on GS-QA2. Nemotron 3 Ultra (high), Terminus-2 is cheaper and faster: $0.067 against $0.20 per task, and 2.4 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 | Nemotron 3 Ultra (high) | DeepSeek V4.1 Flash (max) |
|---|---|---|
| Recognising infeasible requests | 68 | 40 |
| GIS analysis | 70 | 92 |
| Remote sensing and raster analysis | 70 | 75 |
| Climate and time series | 68 | 64 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: Nemotron 3 Ultra (high), Terminus-2 · DeepSeek V4.1 Flash (max), Claude Code.
Side by side
| Metric | Nemotron 3 Ultra (high), Terminus-2 | DeepSeek V4.1 Flash (max), Claude Code |
|---|---|---|
| Geospatial Agent Index: Text only | 68 | 68 |
| Geospatial Agent Index: Text + vision | No data | No data |
| Earth-Bench V2 score | 69 | 58 |
| GeoBenchX score | 65 | 55 |
| GS-QA2 score | 70 | 89 |
| Vision agents score | No data | No data |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 96 | 96 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 63 | 100 |
| OGD4All score (standalone; in no index) | 92 | 79 |
| GeoNatureAgent score (standalone; in no index) | 92 | 100 |
| MultiGlobeQA score (standalone; in no index) | 72 | 64 |
| GeoAgentEval score (standalone; in no index) | 92 | 84 |
| Remote sensing and raster analysis score | 70 | 75 |
| GIS analysis score | 70 | 92 |
| Climate and time series score | 68 | 64 |
| Spatial statistics score | 31 | 31 |
| Making maps score | 70 | 72 |
| Recognising infeasible requests score | 68 | 40 |
| Cost per task | $0.067 | $0.20 |
| Time per task (at 100 output tokens/s) | 2.4 min | 10.9 min |
| Tokens per attempt | 144k | 1.58M |
| Turns per attempt | 12.0 | 26.0 |
| Input price per 1M tokens | $0.50 | $0.30 |
| Output price per 1M tokens | $2.20 | $1.20 |
| Context window | 262,144 | 1,048,576 |
| Image input | No | Yes |
| Developer | NVIDIA | DeepSeek |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Nemotron 3 Ultra (high), Terminus-2 or DeepSeek V4.1 Flash (max), Claude Code?
Nemotron 3 Ultra (high), Terminus-2, 68 to 68.
Which is cheaper per task, Nemotron 3 Ultra (high), Terminus-2 or DeepSeek V4.1 Flash (max), Claude Code?
Nemotron 3 Ultra (high), Terminus-2, $0.067 to $0.20.
Which writes fewer output tokens per task, Nemotron 3 Ultra (high), Terminus-2 or DeepSeek V4.1 Flash (max), Claude Code?
Nemotron 3 Ultra (high), Terminus-2, 9k to 48k.
Which has the larger context window, Nemotron 3 Ultra (high), Terminus-2 or DeepSeek V4.1 Flash (max), Claude Code?
DeepSeek V4.1 Flash (max), Claude Code, 1,048,576 tokens to 262,144 tokens.