Gemma 4 31B (Reasoning) vs Claude Haiku 4.5 (Reasoning)
Results updated 11 November 2026 · In Terminus-2 and Claude Code, so part of any gap can come from the harness · How we benchmark
Verdict
Claude Haiku 4.5 (Reasoning) beats Gemma 4 31B (Reasoning) by 3 points, and is faster; Gemma 4 31B (Reasoning) is cheaper.
- Cost per task
- $0.015 vs $0.092
- Time per task
- 2.1 min vs 1.6 min
- Tokens per attempt
- 93k vs 313k
- Image input
- Yes vs Yes
Gemma 4 31B (Reasoning), Terminus-2 scores higher on GeoBenchX; Claude Haiku 4.5 (Reasoning), Claude Code scores higher on Earth-Bench V2 and GS-QA2. Gemma 4 31B (Reasoning), Terminus-2 is cheaper ($0.015 against $0.092 per task), but Claude Haiku 4.5 (Reasoning), Claude Code is faster (1.6 min against 2.1 min).
Where they differ most
By kind of work, share of tasks passed, 0 to 100 · Higher is better
| Kind of work | Gemma 4 31B (Reasoning) | Claude Haiku 4.5 (Reasoning) |
|---|---|---|
| Change detection | 52 | 68 |
| Making maps | 70 | 56 |
| GIS analysis | 56 | 67 |
| Remote sensing and raster analysis | 59 | 64 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: Gemma 4 31B (Reasoning), Terminus-2 · Claude Haiku 4.5 (Reasoning), Claude Code.
Side by side
| Metric | Gemma 4 31B (Reasoning), Terminus-2 | Claude Haiku 4.5 (Reasoning), Claude Code |
|---|---|---|
| Geospatial Agent Index: Text only | 55 | 58 |
| Geospatial Agent Index: Text + vision | No data | 48 |
| Earth-Bench V2 score | 44 | 51 |
| GeoBenchX score | 61 | 53 |
| GS-QA2 score | 61 | 69 |
| Vision agents score | No data | 19 |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 83 | 78 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 63 | 74 |
| OGD4All score (standalone; in no index) | 85 | 95 |
| GeoNatureAgent score (standalone; in no index) | 100 | 93 |
| MultiGlobeQA score (standalone; in no index) | 30 | 70 |
| GeoAgentEval score (standalone; in no index) | 88 | 88 |
| ImpactMesh score (standalone; in no index) | No data | No data |
| Remote sensing and raster analysis score | 59 | 64 |
| Change detection score | 52 | 68 |
| GIS analysis score | 56 | 67 |
| Climate and time series score | 47 | 46 |
| Making maps score | 70 | 56 |
| Cost per task | $0.015 | $0.092 |
| Time per task (at 100 output tokens/s) | 2.1 min | 1.6 min |
| Tokens per attempt | 93k | 313k |
| Turns per attempt | 9.0 | 10.6 |
| Input price per 1M tokens | $0.17 | $1.00 |
| Output price per 1M tokens | $0.40 | $5.00 |
| Context window | 262,144 | 200,000 |
| Image input | Yes | Yes |
| Developer | Anthropic |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Gemma 4 31B (Reasoning), Terminus-2 or Claude Haiku 4.5 (Reasoning), Claude Code?
Claude Haiku 4.5 (Reasoning), Claude Code, 58 to 55.
Which is cheaper per task, Gemma 4 31B (Reasoning), Terminus-2 or Claude Haiku 4.5 (Reasoning), Claude Code?
Gemma 4 31B (Reasoning), Terminus-2, $0.015 to $0.092.
Which writes fewer output tokens per task, Gemma 4 31B (Reasoning), Terminus-2 or Claude Haiku 4.5 (Reasoning), Claude Code?
Claude Haiku 4.5 (Reasoning), Claude Code, 7k to 8k.
Which has the larger context window, Gemma 4 31B (Reasoning), Terminus-2 or Claude Haiku 4.5 (Reasoning), Claude Code?
Gemma 4 31B (Reasoning), Terminus-2, 262,144 tokens to 200,000 tokens.