Kimi K3.1 Pro: Within a Point of Claude Opus 5.5 on Terminal-Bench, $2 per Million Input Tokens, Open Weights by 4 November
Moonshot's new flagship is a post-trained Kimi K3 with a 1M-token context and a Highspeed version that runs 2-3 times faster. It costs about a third less than K3 through the API, and Moonshot deliberately cut its hacking ability before the open release.
TL;DR
Moonshot AI released Kimi K3.1 Pro on 23 October 2026, a post-trained version of Kimi K3 with a 1-million-token context window, 2.8 trillion total parameters and 104 billion active. It scores 65.7% on Terminal-Bench 4.0 by Moonshot's count, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra, and Artificial Analysis puts it at 55 on its Intelligence Index against Opus 5.5's 58. API pricing is $2 per million input tokens and $10 per million output tokens, about a third less than K3, and full open weights are promised by 4 November. Moonshot removed cyberattack-related training data before release, dropping its internal exploit benchmark score from 90.5% to 81%.
Kimi K3 launched in July as Moonshot's open-weight flagship. K3.1 Pro is the same base with much more post-training, and Moonshot is again promising the weights. The headline benchmark is close to Claude Opus 5.5. The details that matter for buyers are the price cut, who can actually use it in which app, and what Moonshot removed before the open release.
The scores
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 58.2% |
| Artificial Analysis Intelligence Index (independent) | 55 | 58 | - |
| BenchCAD with tools (Voxel IoU) | 92.7 | 96.2 | 95.9 |
| Hallu Bench, no tools (hallucination rate) | 5.4 | 6.8 | 4.2 |
| Hallu Bench Hard, with tools | 8.3 | 15.0 | 5.5 |
Against its own predecessor, Moonshot reports the claim-level hallucination rate on its hardest internal test fell from 23.4% on K3 to 8.3%, and its long-horizon Kimi Code Bench score rose from 55.95% to 61.2%. It also reports 64.4 on OfficeQA Pro v2, which it calls state of the art.
Price and access
- API: $2 per million input tokens and $10 per million output tokens, about a third less than Kimi K3, under the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Zero data retention is available for API customers.
- Apps: Kimi.ai, the iOS, Android and HarmonyOS apps, Kimi Work on Windows and Apple silicon Macs (version 3.1.0 or later), and Kimi Code.
- Free plan: Notebookcheck reports limited K3.1 Pro usage on the web and in Kimi Work. Unite.AI reports that in Kimi Code the model needs the Plus plan or higher. Both may be true for different apps, so check before you plan around the free tier.
- Highspeed: 2-3 times faster than K3.1 Pro and limited to the Pro plan and above.
- Business: off by default in Kimi Business; administrators have to enable it.
What Moonshot removed before the open release
Moonshot says it found and removed cyber-related data from both training stages for the open-weight release while keeping general coding. On its internal xdev-bench-hard test, which asks a model to exploit recently disclosed vulnerabilities, a checkpoint trained without that step solved 90.5% of tasks and the final K3.1 Pro solved 81%. The original K3 scored 33.3%, so the released model is still far more capable here than its predecessor. None of these figures has been independently checked, and Notebookcheck notes that one group on Hugging Face has already announced plans to strip the safeguards once the weights are out.
What isn't known yet
- Licence: neither the announcement nor coverage names the licence for the open weights.
- Weights: due by 4 November. As of 25 October no K3.1 version was listed on Hugging Face.
- Hardware: at 2.8 trillion parameters, it is far too large for a PC or laptop. Most people will use Moonshot's servers or a hosted provider.
- Release date: Moonshot's page and most trackers say 23 October; Unite.AI's text says 24 October.
What this means if you're choosing a model
- For agentic coding on a budget, K3.1 Pro is the closest open-weight option to Opus 5.5 on Terminal-Bench, at a much lower API price. Test it on your own tasks first, since the headline score is Moonshot's.
- If hallucination matters most, GPT-6 Astra still scores lower on Moonshot's own tests, but K3.1 Pro beats Opus 5.5 there.
- If you plan to self-host, wait for the weights and the licence on 4 November before committing.
- Kimi Business admins have to switch it on themselves, since it is off by default at launch.
Sources
AI Industry Reporter
Priya covers model releases, industry announcements, and the gap between what labs claim and what independent evaluators actually find. She reads the primary source - the paper, the system card, the benchmark org's own statement - before writing a word.
More on AI Chatbots & Assistants
Microsoft Starts Capping Free Copilot Chat for Companies With 2,000+ Seats, and Won't Say Where the Cap Is
Priya Nair · 4 min
Thinking Machines' Inkling 2 Small: Scores 31 on Artificial Analysis, Ships Closed, Promises Apache 2.0 Weights Next Week
Priya Nair · 5 min
Anthropic's Amended IPO Filing: $16.96 Billion in Q3 Revenue, a $567 Million Operating Loss and $5 Billion of Short-Term Compute
Priya Nair · 5 min