Kimi K3.1 Pro: 0.7 points behind Claude Opus 5.5, with weights promised by November 4
Moonshot AI released Kimi K3.1 Pro on October 23, 2026, a mixture-of-experts model with about 2.8 trillion total and 104 billion active parameters and a 1-million-token context. By Moonshot's count it scores 65.7% on Terminal-Bench 4.0, 0.7 points behind Claude Opus 5.5, and costs $2/$10 per million tokens; it is API-only for now, with full weights promised by November 4
The three lines
- Scale — ~2.8T total, ~104B active, 1M context, image and video input
- Standing — 65.7% Terminal-Bench 4.0; 55 on Artificial Analysis, second to Opus 5.5
- Terms — $2/$10, a third cheaper than K3; weights pledged by November 4
Key questions
- Kimi K3.1 Pro benchmarks
- **Within a point of Claude Opus 5.5 on Moonshot's coding-agent test.** | Test | K3.1 Pro | Comparison | |---|---|---| | Terminal-Bench 4.0 | 65.7% | Opus 5.5 66.4%, GPT-6 Astra 58.2% | | BenchCAD (tools) | 92.7 | Opus 5.5 96.2 | | Hallucination rate | 8.3% | K3 23.4% | | Artificial Analysis index | 55 | Opus 5.5 58 |
- Kimi K3.1 Pro pricing
- **$2 input and $10 output per million tokens.** | Model | Input | Output | Cache hit | |---|---|---|---| | K3.1 Pro | $2 | $10 | $0.20 | | K3.1 Pro Highspeed | $4 | $20 | $0.20 | | K3 | $3 | $15 | $0.30 |
- Is Kimi K3.1 Pro open weights
- **Not yet. Moonshot says full weights will be released by November 4, 2026.** | Item | Status (Oct 24) | |---|---| | API and Kimi apps | Live | | Weights | Pending, due by Nov 4 | | License | Not named | | Cyber data | Removed from training, Moonshot says |
A Chinese lab now sits within a point of the best US model on a major coding-agent test. Moonshot AI's Kimi K3.1 Pro, released October 23, 2026, scored 65.7% on Terminal-Bench 4.0 by the company's count, against 66.4% for Anthropic's Claude Opus 5.5 and 58.2% for OpenAI's GPT-6 Astra.
The bigger news is the promise attached: Moonshot says it will publish full weights by November 4. If it does, the closest thing yet to a frontier model becomes downloadable. For now, K3.1 Pro is available only through the Kimi apps and API.
1. What shipped
K3.1 Pro is a post-trained upgrade of K3 rather than a new architecture. Moonshot calls it "3T-class."
| Item | Value | Why it matters |
|---|---|---|
| Total parameters | ~2.8T | What you must store |
| Active per token | ~104B | What you compute (~3.7%) |
| Context | 1,048,576 tokens | Output ceiling also 1M |
| Input | Text, image, video | Multimodal |
| Reasoning | Always on; low, high, max | Defaults to max |
| Sampling | Fixed at temperature 1.0, top_p 0.95 | Other values return an error |
Even once the weights are out, this is not a laptop model. For scale, Meta's 816-billion-parameter Muse Spark Open is about 1.5TB in 16-bit form; K3.1 Pro is more than three times larger. Our explainer "What is model quantization" covers how such models get squeezed onto servers.
2. The scorecard
| Test | K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra | Source |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 58.2% | Moonshot |
| BenchCAD (tools) | 92.7 | 96.2 | 95.9 | Moonshot |
| Hallucination, with tools (lower is better) | 8.3% | 15.0% | 5.5% | Moonshot |
| Artificial Analysis index | 55 | 58 | 53 | Independent |
On the independent Artificial Analysis index, K3.1 Pro ranks second at 55, behind Opus 5.5 at 58 and ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon at 53 each. Simon Willison called it the first time a soon-to-be open-weight model has scored this close to the frontier.
| Price per 1M tokens | Input | Output | Cache hit |
|---|---|---|---|
| K3.1 Pro | $2 | $10 | $0.20 |
| K3.1 Pro Highspeed (2–3x faster) | $4 | $20 | $0.20 |
| K3 | $3 | $15 | $0.30 |
That is a third cheaper than K3, matching GPT-6.1 Sol and half of Opus 5.5 by Willison's comparison, with no long-context surcharge.
3. Hacking skills, deliberately trimmed
Because released weights cannot be recalled, Moonshot says it stripped cyberattack material from training before the open release.
| Exploit-real-bugs test (Moonshot's own) | Success |
|---|---|
| Checkpoint before data removal | 90.5% |
| DeepSeek V4.1-Flash (Moonshot's count) | 85.7% |
| Released K3.1 Pro | 81.0% |
| K3 | 33.3% |
Even trimmed, it is more than double K3. Why that matters is illustrated by the Claude Mythos 5 incident analysis published the same day.
4. Open questions
- Weights: The November 4 date and the license are unconfirmed.
- Self-reported scores: Most figures above are Moonshot's; outside reproduction will settle the "0.7 points" claim.
- Allegations: In July the White House accused Moonshot of using banned Nvidia chips and copying an Anthropic model, without publishing evidence. Our explainer on distillation attacks covers the background.
- K2 retirement: K2-series models close to new sign-ups on October 31 and retire at the end of December.