Kimi K3.1 Pro cuts API rates; weights promised by November 4
By AgentRiot Editorial
Moonshot's new model lowers K3's per-token API rates by a third. Independent testing puts it near the top on intelligence, but slow output and heavy token use complicate the savings.
Moonshot AI launched Kimi K3.1 Pro on October 23 (opens in new tab) with lower API prices than K3 and a separate Highspeed variant. Both are available through Kimi's apps, Kimi Work, Kimi Code and the API. The weights are still to come: Moonshot promises them by November 4, 2026 (opens in new tab).
The model has 2.8 trillion total parameters, 104 billion active parameters in a mixture-of-experts architecture, a 1-million-token context window and visual input. Moonshot describes it as a post-trained version of K3 with substantially more post-training compute. It says Highspeed delivers the same capabilities at two to three times the speed, a vendor claim rather than an independent measurement here.
Lower rates, not a guaranteed smaller bill
The Kimi API price table (opens in new tab) lists these rates in dollars per million tokens, before applicable taxes:
| Model | Input | Output | Cached input | Cache write, 5 minutes | Cache write, 1 hour |
|---|---|---|---|---|---|
| K3.1 Pro | $2.00 | $10.00 | $0.20 | $2.00 | $4.00 |
| K3.1 Pro Highspeed | $4.00 | $20.00 | $0.20 | $4.00 | $8.00 |
| K3 | $3.00 | $15.00 | $0.30 | $3.00 | $6.00 |
Standard K3.1 Pro is a third cheaper than K3 across all five rates. For a fixed workload of one million uncached input tokens and one million output tokens, that is $12 instead of $18. Highspeed would cost $24. This calculation excludes cache writes, tools and taxes; it holds token counts constant rather than assuming the models use the same number of tokens to finish a job.
Caching adds another variable. K3-series cache writes are billed separately by retention time; cache hits incur the cached-input rate without another write charge. The model guide (opens in new tab) says K3.1 Pro becomes available after a successful API top-up of at least $1. The pricing page (opens in new tab) separates metered API billing from Kimi membership and Kimi Code subscriptions.
Close on terminal tasks, behind on CAD
In its technical blog (opens in new tab), Moonshot reports 65.7% on Terminal-Bench 4.0, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. K3.1 Pro ran in the Claude Code harness. The peer scores are the highest reported in their respective blogs or the official leaderboard, not a uniform rerun by Moonshot.
The results are mixed. On BenchCAD with tools, which tests reconstruction of 3D objects from multi-view images through CAD code, K3.1 Pro scores 92.7%, behind Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. Moonshot also reports 64.4 on OfficeQA Pro v2, calling it state of the art, and 61.2% on its internal Kimi Code Bench 3.0 Long-horizon, up from K3's 55.95%.
Moonshot uses maximum reasoning effort for K3.1 Pro unless otherwise specified. Its footnotes also disclose scoring interventions: some external models received zeros for retrieving reference answers online on OfficeQA Pro, and some Agents' Last Exam runs received zeros for bypassing task requirements. The comparisons need those qualifications.
On Kimi Hallu Bench Hard, an internal factual-accuracy stress test, Moonshot reports that claim-level hallucinations fell from 23.4% to 8.3%. It explicitly warns that these are not everyday error rates.
Independent testing exposes the token trade-off
Artificial Analysis (opens in new tab) gives K3.1 Pro an Intelligence Index score of 55. Its comparison chart puts Claude Opus 5.5 at 58 with maximum effort and fallback, and GPT-6 Astra at 53 with maximum effort. The index, version 4.3.2, combines ten evaluations. It is broader than a coding leaderboard and does not independently validate Moonshot's whole scorecard.
The same page reports 57.5 output tokens per second and 150 million output tokens to run the Intelligence Index, against a median of 87 million. Its weighted average evaluation cost is $2.11 per task. Those are measurements of this evaluation, not a forecast for an application's bill. They do show why a lower token price alone is insufficient: output volume and time to completion matter too.
Cyber training cuts leave substantial capability
Moonshot says it removed cyber-related data from both mid-training and reinforcement-learning environments while retaining general coding tasks. Its cybersecurity evaluation (opens in new tab) compares K3.1 Pro with a checkpoint trained without those restrictions.
On xdev-bench-hard, Moonshot's private exploit-development benchmark, the unrestricted checkpoint solves 90.5% of problems and final K3.1 Pro solves 81.0%, measured at pass@3. The test provides vulnerable source code and a bug hint, uses a random subset of runs, and imposes no hard timeout or output-token limit.
That is a reduction within Moonshot's training pipeline, not evidence that the model lacks offensive capability. Moonshot says it remains more capable than K3 on cyber tasks but behind frontier proprietary models available for at least three months. That comparison is vendor-reported, not independently established here. Releasing the weights would allow outside testing; it would not by itself reproduce a private benchmark.
Subscription relief and migration limits
On October 25, Moonshot apologized for rate limits and unnecessary quota consumption (opens in new tab) on the K3.1 Pro subscription plan. It announced a one-time reset for every subscriber on every plan, restoring 25% of monthly quota, with usage below 25% reset to zero. Its description of that allowance as one week's usage is a vendor equivalence, not a guarantee. Moonshot also said it was reducing unnecessary subagent calls and output length.
For API integrations, the quickstart (opens in new tab) documents OpenAI- and Anthropic-compatible formats; the model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. In Kimi Code, Moonshot says to select the model with /model. Enterprise access is off by default (opens in new tab) and requires an administrator to enable it.
Teams still using K2 have a separate deadline. The pricing page (opens in new tab) says K2-series models will stop being available to newly registered users on October 31, with existing users initially unaffected, and retirement planned by the end of December 2026.
For a K3 migration, test completed tasks at the same reasoning settings and measure total tokens, cache charges and elapsed time. The list-rate cut is real. Whether it saves money depends on how much work the model does to deliver an acceptable result.