The short version:Kimi K3.1 Pro is the flagship MoE model Moonshot (Kimi) released on October 23, 2026. Official pricing is $2 per million input tokens and $10 per million output tokens ($4/$20 for the high-speed tier), with a context window of 1,048,576 tokens, about 2.8T total parameters, and about 104B active. Its Artificial Analysis Intelligence Index score is around 55, behind Opus 5.5, Sonnet 5.5, and Muse Wonder Full. Third parties measured about 67% on Terminal-Bench 4.0, while the official self-reported score (Claude Code harness) is 65.7%.
Numbers in this post are current as of 2026-10-24 and based on Kimi's official launch blog post,platform pricing page,quickstart, and the Artificial Analysis post,OpenRouter model page. Third-party leaderboards change, so defer to the original pages. I'll update this post as more third-party evaluations come out.
What Is K3.1 Pro
Key points from the official announcement:
- Officially released on October 23, 2026, positioned as the next-generation Kimi flagship
- A MoE with about 2.8T total parameters and about 104B active
- 1M context; accepts text, image, and video input and outputs text
- Thinking is always on, and can be set via
reasoning_efforttolow/high/max(defaultmax) - Available in a standard version,
kimi-k3.1-pro, and a high-speed version,kimi-k3.1-pro-highspeed - Weights are promised to be open by November 4, 2026(not yet released as of 2026-10-24)
- The API is compatible with OpenAI and Anthropic styles, with the endpoint
https://api.moonshot.ai/v1
Compared with its predecessor kimi-k3, the official price is lower (K3 is $3/$15), and it reports higher self-reported scores on long-horizon coding, office Q&A, and similar tests.
Spec Cheat Sheet
| Item | Official details |
|---|---|
| Model name (Moonshot API) | kimi-k3.1-pro / kimi-k3.1-pro-highspeed |
| Context | 1,048,576 tokens |
| Parameters | ~2.8T total / 104B active (MoE) |
| Input modalities | Text, image, video |
| Output | Text |
| Thinking | Always on;reasoning_effort: low / high / max(default max) |
| Protocols | OpenAI-compatible, Anthropic-compatible |
Official pricing
Unit: per million tokens (MTok). Checked against the platform pricing page on 2026-10-24,with no tiers by context length.
| Model | Input | Output | Cache Write | Cache Hit |
|---|---|---|---|---|
| kimi-k3.1-pro | $2.00 | $10.00 | $2.00 | $0.20 |
| kimi-k3.1-pro-highspeed | $4.00 | $20.00 | $4.00 | $0.20 |
| kimi-k3(for comparison) | $3.00 | $15.00 | $3.00 | $0.30 |
You can also call it through OpenRouter at moonshotai/kimi-k3.1-pro. Check the OpenRouter page at the time for exact pricing.
Benchmarks
Self-reported by the vendor (marked as official)
These come from the Kimi launch blog, so read them as official claims:
| Benchmark | K3.1 Pro (official) | Notes |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Claude Code harness |
| Kimi Code Bench 3.0 Long-horizon | 61.2% | vs. K3 at 55.95% |
| BenchCAD (with tools) | 92.7% | Official self-test |
| OfficeQA Pro v2 | 64.4 | Official self-test |
| Hallu Bench Hard (claim-level) | 8.3% | vs. K3 at 23.4%; internal stress test |
Third party: Artificial Analysis / OpenRouter
| Metric | Value (verified 2026-10-24) | Source |
|---|---|---|
| Intelligence Index | 55 (AA post) / ~55.1 (OpenRouter embed) | AA post; OpenRouter |
| Approximate ranking | AA post says about 4th; OpenRouter embed showed about 3rd of 352 at the time | Leaderboards move |
| Comparison | Below Opus 5.5 (58), Sonnet 5.5 (56), and Muse Wonder Full (56) | AA post |
| Terminal-Bench 4.0 | ~67% | AA; higher than the official self-test of 65.7% |
| GDPval-AA | ~1,846 Elo; OpenRouter shows ~67.3% | AA / OpenRouter |
| AutomationBench-AA | ~73% (K3 at ~58%) | AA post |
| HLE | ~52% (OpenRouter ~52.5%) | AA / OpenRouter |
| AA-Omniscience Index | ~30 (accuracy ~51%, hallucination ~44%) | AA |
| Cost per task (order of magnitude) | ~$1.98–$2.11 | AA post |
A note on the numbers: The official Terminal-Bench score and AA's differ by about 1.3 percentage points. That's common with harness and sampling differences, so just label both clearly. Knowledge-related hallucination is still not low on AA-Omniscience, so add verification for factual Q&A.
Strengths, weaknesses, and where it fits
Good for:
- Long-context agents and multi-file coding, plus terminal-oriented tasks (Terminal-Bench is a standout)
- A model that gets real work done at a lower price than flagship models (cheaper than Opus at a comparable tier)
- Projects already on Moonshot / OpenRouter that need an OpenAI- or Anthropic-compatible protocol
Less suitable, or use with care:
- Factual Q&A without verification (AA's hallucination rate is still on the high side)
- The default
reasoning_effort=maxis slower and uses more tokens, so step down tolow/high - If you need local or open-weight deployment: the weights are promised before November 4, and as of this writing they have not been released
How to connect
Moonshot official API
curl https://api.moonshot.ai/v1/chat/completions \
-H "Authorization: Bearer $MOONSHOT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k3.1-pro",
"messages": [{"role": "user", "content": "你好"}],
"reasoning_effort": "high"
}'
For the high-speed version, replace model with kimi-k3.1-pro-highspeed. You can also use the Anthropic-compatible endpoint; see the official quickstart for details.
OpenRouter
Model id:moonshotai/kimi-k3.1-pro.
AI Code With
As of 2026-10-24, the site catalog already includes kimi-k2.6,kimi-k2.7-code,kimi-k3,but kimi-k3.1-prois not listed yet. Once it is, we'll add the model id and usage notes here.
Source verification
Verified on:2026-10-24 (Asia/Shanghai)
