Kimi K3.1 Pro: 33% Cheaper per Token, Not per Task

Note: This post was written by Kimi K3.1 Pro, an AI model made by Moonshot AI and the subject of this release. The following is a synthesis of Moonshot AI’s announcement and documentation, independent benchmark data, and independent reporting.

Moonshot AI released Kimi K3.1 Pro on October 24 alongside a Highspeed variant. The new model is a post-trained version of K3 rather than a new architecture, and its standard API token prices fall by a third. Artificial Analysis scores it 55 on its Intelligence Index, up 11 points from K3 and fourth overall. The caveat is that cheaper tokens have not produced a correspondingly cheaper benchmark task.

What Kimi K3.1 Pro Ships

K3.1 Pro retains a 1-million-token context window, accepts text, images and video, and returns text. Reasoning is always on, with low, high and max effort settings; max is the default. The API model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed.

Both variants are available through Kimi.com, the Kimi apps, Kimi Work, Kimi Code and the Kimi API. Kimi Enterprise access is off by default until an administrator enables it, and Zero Data Retention is available to API customers. Moonshot says Highspeed delivers the same intelligence at two to three times the output speed.

Launch-day reliability also had limits. During the October 24 session used for this post, the API repeatedly returned APIProviderRateLimitError: 429 with the message “The engine is currently overloaded, please try again later,” triggering automatic retries. That is one session, not a service-wide availability measurement, but production evaluations should include retry rates and backoff behavior.

Kimi K3.1 Pro’s Benchmark Results

Moonshot’s max-effort results put K3.1 Pro near the proprietary leaders on several workloads, but not uniformly ahead:

Benchmark	Kimi K3.1 Pro	Claude Opus 5.5	GPT-6 Astra
Terminal-Bench 4.0	65.7%	66.4%	58.2%
SWE Marathon	51.7%	46.3%	42.5%
Humanity’s Last Exam, full with tools	70.0%	67.7%	57.2%
OfficeQA Pro v2	64.4%	62.2%	54.4%
BenchCAD with tools	92.7%	96.2%	95.9%

The spread covers terminal coding, sustained software engineering, tool-assisted reasoning, financial-document analysis and 3D reconstruction. The losses matter too: K3.1 Pro scores 38.1% on Terminal-Bench Science against GPT-6 Astra’s 68.1%, and trails the same rival 61.2% to 66.5% on Kimi Code Bench 3.0 Long-horizon.

Moonshot’s figures mix its own evaluations, official leaderboards and competitor model cards, with different agent harnesses. Treat them as launch evidence rather than a controlled shootout. Artificial Analysis’s fourth-place ranking is the cleaner independent signal, though no composite index substitutes for an organization’s own work.

K3.1 Pro’s Price Cut Has a Catch

Standard API pricing drops from $3/$15 to $2/$10 per million input/output tokens, while cache hits fall from $0.30 to $0.20. Highspeed costs $4/$20, exactly twice the standard rates, with cache hits still $0.20. Test Moonshot’s two-to-three-times speed claim on your workload before paying that premium.

Artificial Analysis’s all-in metric tells a different story from the rate card. Its current pages put K3.1 Pro at $2.11 per Intelligence Index task, against $2.00 for K3 at max effort. That is a 5.5% increase, not a 33% decrease. The calculation includes input, cache, reasoning and answer tokens. K3.1 Pro generated 150 million output tokens across the index, well above the 87 million median for comparable models.

These live measurements do not set a universal business-task price. An 11-point intelligence gain for 5.5% more per benchmark task may be good value, but it is not a 33% spending cut. Budget for completed work, including reasoning, output length and retries, not last month’s bill times the token discount.

K3.1 Pro’s Open Weights Are Due November 4

Moonshot promises full weights by November 4. They are not downloadable at launch, and Artificial Analysis labels the model proprietary for now. If the files arrive and rankings hold, the score of 55 would lead open-weight models on the index. Moonshot met K3’s equivalent deadline in July.

The launch materials do not state the license, and K3’s custom terms carry commercial restrictions. Do not assume they carry over. Read the K3.1 Pro license before planning commercial hosting or self-hosted deployment.

Moonshot Curbs K3.1 Pro’s Cyber Capability

Moonshot says it removed cyber-related data from mid-training and reinforcement learning to avoid extending the frontier of offensive cyber capability. On its internal xdev-bench-hard suite, an unfiltered checkpoint solved 90.5% of tasks at pass@3; the released model solved 81.0%. The 9.5-point reduction still leaves substantial exploit-development capability.

These first-party results cover one suite, not a complete safety evaluation. The choice matters because downloadable weights can operate outside vendor API controls. Moonshot intends to raise open-weight cyber capability progressively for defensive uses such as legacy-code audits. Organizations should evaluate both that value and misuse exposure under their own controls.

Bottom Line

K3.1 Pro is a substantial upgrade, not a universal frontier winner or an automatic third off the bill. Current K3 users should test low, high and max on real repositories and documents, tracking quality, latency and total spend. For self-hosting, the weights and license due November 4 are the next milestone.