Kimi K3 (Reasoning) vs Claude Haiku 5.5 (xhigh)
Results updated 31 October 2026 · In Kimi Code and Claude Code, so part of any gap can come from the harness · How we benchmark
Verdict
Claude Haiku 5.5 (xhigh) edges Kimi K3 (Reasoning) by 1 point; Kimi K3 (Reasoning) is cheaper and faster.
- Cost per task
- $0.00 vs $0.017
- Time per task
- 1.5 min (estimated) vs 2.5 min
- Tokens per attempt
- 345k vs 458k
- Image input
- Yes vs Yes
Kimi K3 (Reasoning), Kimi Code scores higher on Earth-Bench V2 and GS-QA2; Claude Haiku 5.5 (xhigh), Claude Code scores higher on GeoBenchX. Kimi K3 (Reasoning), Kimi Code is cheaper and faster: $0.00 against $0.017 per task, and 1.5 min against 2.5 min.
Where they differ most
By kind of work, share of tasks passed, 0 to 100 · Higher is better
| Kind of work | Kimi K3 (Reasoning) | Claude Haiku 5.5 (xhigh) |
|---|---|---|
| Recognising infeasible requests | 52 | 83 |
| Climate and time series | 78 | 69 |
| Spatial statistics | 38 | 46 |
| Remote sensing and raster analysis | 81 | 76 |
Where the tasks are set: each kind of work links to its tasks and scoresCharts for each: Kimi K3 (Reasoning), Kimi Code · Claude Haiku 5.5 (xhigh), Claude Code.
Side by side
| Metric | Kimi K3 (Reasoning), Kimi Code | Claude Haiku 5.5 (xhigh), Claude Code |
|---|---|---|
| Geospatial Agent Index: Text only | 75 | 76 |
| Geospatial Agent Index: Text + vision | No data | 77 |
| Earth-Bench V2 score | 74 | 68 |
| GeoBenchX score | 65 | 77 |
| GS-QA2 score | 85 | 81 |
| Vision agents score | No data | 80 |
| GeoAgentBench score (Guided workflows; not in the Agent Index) | 96 | 100 |
| GeoAnalystBench score (Guided workflows; not in the Agent Index) | 100 | 95 |
| OGD4All score (standalone; in no index) | 97 | 97 |
| GeoNatureAgent score (standalone; in no index) | 100 | 100 |
| MultiGlobeQA score (standalone; in no index) | 77 | 75 |
| GeoAgentEval score (standalone; in no index) | 96 | 92 |
| Remote sensing and raster analysis score | 81 | 76 |
| GIS analysis score | 84 | 82 |
| Climate and time series score | 78 | 69 |
| Spatial statistics score | 38 | 46 |
| Making maps score | 82 | 86 |
| Recognising infeasible requests score | 52 | 83 |
| Cost per task | $0.00 | $0.017 |
| Time per task (at 100 output tokens/s) | 1.5 min (estimated) | 2.5 min |
| Tokens per attempt | 345k | 458k |
| Turns per attempt | No data | 9.6 |
| Input price per 1M tokens | $3.00 | $0.10 |
| Output price per 1M tokens | $15.00 | $0.50 |
| Context window | 1,000,000 | 1,000,000 |
| Image input | Yes | Yes |
| Developer | Moonshot AI | Anthropic |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Kimi K3 (Reasoning), Kimi Code or Claude Haiku 5.5 (xhigh), Claude Code?
Claude Haiku 5.5 (xhigh), Claude Code, 76 to 75.
Which is cheaper per task, Kimi K3 (Reasoning), Kimi Code or Claude Haiku 5.5 (xhigh), Claude Code?
Kimi K3 (Reasoning), Kimi Code, $0.00 to $0.017.
Which writes fewer output tokens per task, Kimi K3 (Reasoning), Kimi Code or Claude Haiku 5.5 (xhigh), Claude Code?
Kimi K3 (Reasoning), Kimi Code, 9k to 12k.
Which has the larger context window, Kimi K3 (Reasoning), Kimi Code or Claude Haiku 5.5 (xhigh), Claude Code?
They are level, at 1,000,000 tokens.