Kimi K3.1 Pro is Moonshot’s open-weight flagship, available through its API since October 23, 2026 at $2 per million input tokens and $10 per million output tokens (USD). For most developers the API is the practical way in: the model has 2.8 trillion total parameters, so self-hosting is a data-center job. This is an analysis of the model card and published coverage, not a hands-on test.

What was announced
According to Let’s Data Science, Moonshot released Kimi K3.1 Pro in its API on October 23, 2026, with a 1-million-token context window and weights planned for November 4. A separate write-up from AICyber says the same and gives the same prices. A Hugging Face model card for the model also appears to be live, but I could not confirm the weights upload date, so treat the November 4 date as a schedule, not a verified release.
Key facts

| Item | Detail |
|---|---|
| API launch | October 23, 2026 |
| Weights | Scheduled for November 4, 2026; model card on Hugging Face |
| Size | 2.8T total parameters, 104B activated per token |
| Context | 1 million tokens |
| Standard API price (USD, per million tokens) | $2 input, $10 output |
| Highspeed API price (USD, per million tokens) | $4 input, $20 output |
| License | Kimi K3.1 Pro License |
Performance: vendor claims versus independent numbers
Moonshot reports 65.7% on Terminal-Bench 4.0. The independent figure available is from Artificial Analysis, whose Intelligence Index score for the model is 55, according to AICyber. The same write-up says Artificial Analysis measured 67% on Terminal-Bench 4.0 with a different harness, so the two numbers should not be compared directly. I found no broader independent testing yet.
Who it matters for
- Teams choosing an API model: at $2/$10 it is priced well below many frontier options, so it deserves a trial against your own tasks. Our look at GPT-6.1 Terra pricing uses the same cost-per-accepted-job lens.
- Organizations with data-control needs: downloadable weights allow self-hosting, but AICyber’s estimate is about 1.4 TB for the weights at 4-bit precision, far beyond a small GPU box. It frames these as theoretical figures, not measured file sizes.
- People comparing open models: see how Mistral Large 4’s license and hardware compare.
Limits and open questions
I have not read the license text, so check the Kimi K3.1 Pro License for commercial terms before building a product. The model card lists vLLM, SGLang and TokenSpeed as recommended inference engines but gives no hardware requirements. AICyber also reports that Moonshot removed cyber-related material from training and saw an internal benchmark drop from 90.5% to 81.0%; this is self-reported and unreproduced. Reports on pricing from aggregator sites conflict in places, so confirm on Moonshot’s own pricing page before budgeting.
My take
The headline is the combination of low API pricing and downloadable weights. The caution is that nearly all quality evidence is vendor-reported. Run a small batch of your own tasks, count retries, and use our comparison checklist to keep the trial fair.
Sources
- Let’s Data Science, October 23, 2026.
- AICyber, October 24, 2026.

