Moonshot’s product brand is still Kimi. The new SKUs are kimi-k3.1-pro and a faster twin, kimi-k3.1-pro-highspeed (~2–3× speed per the company). Both are described as post-trained from Kimi K3 with larger post-training and joint RL. Always-on thinking is on; reasoning_effort can be set to low / high / max (default max). Native visual understanding ships in the same model rather than as a bolted-on vision tower.
What changed for overseas builders is practical: you can hit the model on Kimi.ai, the apps, Kimi Work 3.1.0+, Kimi Code’s /model switch, and the public API today — without waiting for weights. International API base is api.moonshot.ai; China traffic also routes through moonshot.cn. OpenRouter mirrors the same model IDs and prices.
What it costs (flat USD, no context tiers)
Platform pricing on platform.kimi.ai (USD, per million tokens) is flat — no context-length tiers:
- Kimi K3.1 Pro — input $2 / output $10; Cache Write $2; Cache Hit $0.20
- Kimi K3.1 Pro Highspeed — input $4 / output $20; Cache Write $4; Cache Hit $0.20
Those are platform list prices as published with the launch; treat them as vendor pricing that can change. Enterprise access is off by default; Moonshot says API customers can get ZDR (zero data retention).
Illustration: what you can call today vs what is still a promise. Pricing and benches are company/platform figures.
Vendor benches — keep them in quotes
Moonshot’s launch blog posts a slate of internal numbers. Useful as direction-of-travel, not as third-party audits:
- Terminal-Bench 4.0 — 65.7% (company) vs GPT-6 Astra 58.2% / Claude Opus 5.5 66.4%
- Kimi Code Bench 3.0 Long-horizon — 61.2% vs K3’s 55.95%
- BenchCAD with tools — 92.7%
- OfficeQA Pro v2 — 64.4
- Kimi Hallu Bench Hard claim-level hallucination — 8.3% vs K3’s 23.4%
Every comparator name and percentage above is from the company blog. Independent reproductions are not in this piece.
Yes for API and apps today; not for self-host weights yet. Hit api.moonshot.ai or OpenRouter with the published USD rates, or use Kimi.ai / Work / Code. Full open weights are a Nov 4, 2026 promise — download day is not today. Enterprise is off by default; ask Moonshot if you need ZDR.
Open-weight safety narrative
For the coming open release, Moonshot says it removed cyber-related mid-training and RL data. Their own xdev-bench-hard pass@3 figure drops from 90.5% without those controls to 81.0% after — company narrative for a safer open drop, not an external red-team report. Read it as a deliberate capability trade they are advertising ahead of Nov 4.
Honest caveats
- Weights not out — full open weights promised by Nov 4, 2026; nothing to download for self-host today.
- Vendor benches only — Terminal-Bench / Code Bench / Hallu / CAD numbers are company-blog claims.
- “3T-class” — company sizing language; exact parameter breakdown is not independently verified here.
- Highspeed ~2–3× — company speed claim for the faster SKU.
- Cyber data strip — self-reported xdev-bench-hard tradeoff (81.0% vs 90.5%); not third-party audited.
- Pricing — platform list as of launch; confirm live on platform.kimi.ai / OpenRouter before budgeting.
Bottom line for overseas engineers: if you want a long-context, vision-capable, always-thinking Kimi agent today, the API path is open at flat $2/$10 (or $4/$20 Highspeed). If you want to run the weights yourself, the calendar date to watch is November 4, 2026 — and keep every leaderboard figure in the “vendor claim” column until someone else reproduces it.