Moonshot Releases Kimi K3.1 Pro: Intelligence Index Rises to Fourth, API Price Cut 33%
According to an October 23, 2026 report by The Decoder, Moonshot AI released Kimi K3.1 Pro, which scored 55 on the Artificial Analysis Intelligence Index, 11 points higher than K3, ranking fourth. The report also says the standard version's per-token price was cut by about 33%, that a Highspeed variant exists, and that open weights are planned for release by November 4.
What changed
According to an October 23, 2026 report by The Decoder, Moonshot AI released Kimi K3.1 Pro, a post-trained version of the previous Kimi K3 with significantly more post-training compute. The model scores 55 on the Artificial Analysis Intelligence Index, an 11-point improvement over K3, placing it fourth; the report says only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56) rank higher. The model retains 2.8 trillion total parameters, 104 billion activated parameters, and a 1 million token context, and supports text and images. The report says its agentic capabilities improved most, with GDPval-AA reaching 1846 Elo, Terminal-Bench 4.0 at 65.7%, and hallucination rate falling from 23.4% to 8.3% in an internal stress test. A Highspeed variant that runs two to three times faster was also launched, and both are already available through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The report also says the standard version's per-token API price has been cut by about 33%, and that open weights are planned for release by November 4, 2026.
What it means for developers
The reported price cut and performance jump directly change cost per task rather than unit price alone. According to Artificial Analysis data cited by The Decoder, the cost per Intelligence Index task is about $1.98, nearly flat against K3's $2.00, because the new model generates about 14% more output tokens; this means developers who compare unit prices only on the input side may misjudge total cost. For output-heavy workloads such as long-chain agents, terminal operation, and office document automation, K3.1 Pro's relative advantage may be smaller than the leaderboard score gap suggests. If open weights ship as reported, self-hosting and offline audit scenarios will gain a new option, but until the weights are released, all availability claims apply only to Moonshot's official channels and the Kimi API, not third-party platforms.
Integration and evaluation
The following are evaluation suggestions based on the report, not first-hand testing conclusions from this site. First, given the price decline but higher output token count, compare total cost per successful task before deciding whether to migrate. For long-output agent workflows, take the same real requests and call both your current model and the vendor-provided K3.1 Pro, recording input, output, cache hits, retries, and task completion rate. Scale traffic only if total cost per successful task still falls while meeting your defined quality threshold; if your workload is mainly short Q&A or long input, measure it as a separate group rather than applying the reported average cost. Second, for the Highspeed variant, first determine whether latency affects your business before deciding whether to pay a higher unit price. The report says Highspeed's input and output unit prices are both twice the standard version's, and the vendor claims two to three times the speed, but this does not prove your own end-to-end pipeline will speed up by the same margin. For interactive requests with response deadlines, hold prompts, tools, and retry policies fixed, and compare the two versions on time to first token, 95th percentile overall latency, success rate, and cost per successful task; choose Highspeed only if the speedup meets your business deadline and the cost stays within budget. For offline batch processing without deadline pressure, evaluate the standard version first.
Limits and open questions
Kimi K3.1 Pro and the Highspeed variant are not currently listed in KimiSeek's live catalog, and the API availability and pricing in this article refer to Moonshot's official channels. The benchmark results are as relayed by The Decoder from Artificial Analysis or Moonshot, and have not been retested by this site; an internal hallucination stress test cannot be treated as a general error rate, and a leaderboard advantage does not guarantee real workflow gains. The report notes that the vendor's benchmark table uses max reasoning effort, and readers still need to validate with their own tasks. Prices may change, so the vendor's current pricing page should be treated as authoritative. As of the report date, open weights remain a plan; until they actually ship, license, deployment conditions, and availability timing cannot be confirmed on that basis.
Illustrations from the original report
