Moonshot AI released Kimi K3.1 Pro on 23 October 2026, or 24 October in China, Fello AI reported. The Chinese company’s new model is a post-trained version of Kimi K3, with lower API token prices and a fourth-place ranking on the Artificial Analysis Intelligence Index, an independent evaluation suite. For companies building with Kimi, the release offers a stronger model through Moonshot’s hosted services while downloadable weights remain a promised next step.
Key points
- Kimi K3.1 Pro ranks fourth on the Artificial Analysis Intelligence Index, behind three proprietary models.
- The model retains K3’s 2.8 trillion total parameters and 104 billion active parameters.
- API input and output prices have fallen from $3 and $15 to $2 and $10 per million tokens.
- Moonshot says it removed cyber-related training data before the planned open-weight release.
Kimi K3.1 Pro reaches fourth on the index
K3.1 Pro scored 55 points on the Artificial Analysis Intelligence Index, 11 points above K3, The Decoder reported. Its fourth-place position puts it behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 and Muse Wonder Full at 56 each. Fello AI gives the more precise figures of 55.1 for K3.1 Pro and 57.6 for Claude Opus 5.5, with K3.1 Pro scoring 11.5 points higher than K3 on its figures. Those are scores on the evaluation suite, which combines multiple benchmarks, rather than percentages of tasks a model can complete in ordinary use.
The gains also appear on tests involving tools. Moonshot reports a 65.7 percent score for K3.1 Pro on Terminal-Bench 4.0, which tests coding agents on terminal tasks with limited resources, against 58.2 percent for GPT-6 Astra and 66.4 percent for Claude Opus 5.5, Unite.AI reported. Artificial Analysis measured 67 percent for K3.1 Pro on Terminal-Bench 4.0, according to Fello AI. The two K3.1 Pro scores come from separate runs of the same named benchmark.
On BenchCAD, a 3D reconstruction benchmark involving code generated from multiple views of an object, the company reports 92.7 percent for K3.1 Pro when tools are available. Moonshot also reports 61.2 percent on its internal Kimi Code Bench 3.0 Long-horizon test, against 55.95 percent for K3. That test covers work extending across multiple steps, including tool use and completing a problem from start to finish.
K3’s parameters, more post-training and lower prices
Moonshot kept K3’s 2.8 trillion total parameters and 104 billion active parameters in K3.1 Pro. K3 is a mixture-of-experts model: its parameters are arranged so that only a portion is activated for each piece of text processed, rather than bringing the entire model into every calculation. Moonshot says it spent substantially more computing effort on post-training K3.1 Pro, including reinforcement learning across hundreds of downstream domains. AI Affairs previously reported that K3 launched on Amazon Bedrock.
K3.1 Pro accepts text and images, and its documented context limit is 1,048,576 tokens, Kingy AI reported. A context window is the amount of material that can fit into a request, much as a larger desk can hold more pages at once. Kingy AI cautions that the capacity alone does not measure how reliably the model uses material placed throughout that window.
Checking figures in a long document could benefit from being able to supply more of it in one request. Whether the answer would be dependable would still turn on how accurately the model finds and uses the relevant passages, rather than on the amount of material it can accept.
The Kimi API charges $2 per million input tokens and $10 per million output tokens for K3.1 Pro, down from $3 and $15 for K3. Moonshot also offers K3.1 Pro Highspeed, which it says has identical model capabilities and runs two to three times faster. Its API prices are $4 per million input tokens and $20 per million output tokens, according to The Decoder. The lower standard-model token rates concern the price of processing text; K3.1 Pro generates 14 percent more output tokens per Artificial Analysis task than K3.
Moonshot’s cyber test and 4 November weights
Moonshot says it removed cyber-related material from both mid-training data and reinforcement-learning tasks in preparation for an open-weight release. On its in-house xdev-bench-hard test, which uses source code for vulnerability discovery and exploitation tasks, a checkpoint trained without those measures solved 90.5 percent of problems at pass@3, while the final K3.1 Pro reached 81.0 percent. Moonshot says the comparison reflects a reduction in cyber capability relative to the version its training process would otherwise have produced.
Factual-error tests have different scopes. Moonshot reports an 8.3 percent claim-level hallucination rate on its internal Kimi Hallu Bench Hard, down from 23.4 percent for K3, and describes the test as deliberately focused on error-prone situations rather than everyday use. Artificial Analysis measured 44 percent for K3.1 Pro on its separate AA-Omniscience test, down from 53 percent for K3.
Moonshot has made K3.1 Pro and its Highspeed variant available through its Kimi products and API. Moonshot promises to release the full model weights by 4 November 2026.
