Kimi K3.1 Pro launches at $2/$10 per million tokens, weights promised by 4 November

The release is live, but the open-weight claim has a date attached

At Kimi's published US-dollar API rates, a request with 100,000 new input tokens and 10,000 output tokens costs $0.30 on K3.1 Pro. The same token counts cost $0.45 on K3 and $0.60 on K3.1 Pro Highspeed. That is a concrete price cut for a large-context agent workflow, calculated from Kimi's current tariff. It does not prove that K3.1 Pro completes the same job more cheaply: different models can use different numbers of thinking, tool and retry tokens.

The larger news is a model release with two clocks. Kimi says K3.1 Pro and a faster Highspeed variant became available on 23 October through Kimi.ai, Kimi Work, Kimi Code and the API. It calls Pro its most capable 3 trillion-class model and advertises a 1,048,576-token context window. But the launch post says the full weights will arrive by 4 November 2026. Until they and their licence appear, this is a hosted release with a planned open-weight release, not a model that a laboratory can download and audit today.

What shipped on 23 October, and what remains a promise

Kimi describes K3.1 Pro as a post-trained K3, with substantially more post-training compute and joint reinforcement learning across hundreds of downstream domains. It names improvements to its reinforcement-learning framework, long-context inference, rollout scheduling and sandbox infrastructure. The post does not publish a separate K3.1 architecture paper or exact active-parameter count. It would be unsound to transfer every parameter or expert-routing figure from the existing K3 model card to K3.1 without a new disclosure.

Release status from Kimi's 23 October launch post and API documentation.

Item	Stated and available	Not established yet
Hosted access	K3.1 Pro and Highspeed in Kimi products and API; model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed.	Long-term availability, rate limits for each customer tier and measured service-level latency.
Context	K3 series API documentation states 1,048,576 tokens.	That one request can use the full window at useful speed or predictable total cost.
Weights and licence	Kimi promises full model weights by 4 November.	Downloadable K3.1 files, weight formats, hashes, licence terms and local deployment requirements.
Training and safety	Kimi describes more post-training RL and cyber-data filtering.	Exact K3.1 training compute, full data composition or independent K3.1 cyber assessment.

The distinction matters to procurement and research teams. An API gives access to Kimi's hosted snapshot and controls; released weights would enable independent local inspection and deployment, subject to the licence and infrastructure demands. Kimi has not yet published K3.1's licence. The existing K3 repository is useful predecessor context, not a licence or hardware specification for K3.1 Pro. A million-token input allowance is a capacity ceiling, not a latency guarantee or evidence that the whole prompt is equally useful to the model.

Where the benchmark chart is strong, and where it cannot settle a ranking

The most striking coding number in the Kimi launch post is 65.7% on Terminal-Bench 4.0 for K3.1 Pro, compared in Kimi's chart with 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. On its own internal Kimi Code Bench 3.0 Long-horizon, Kimi reports 61.2%, up from 55.95% for K3, a 5.25 percentage-point gain. A percentage-point change is the right unit here; calling that a 5.25% relative gain would describe a different calculation.

The source matters as much as the score. Kimi says its Terminal-Bench run used a Claude-Code harness, while competitor values were taken from the higher of release-blog or official-leaderboard numbers. Its footnotes say other comparisons can involve different agent harnesses, model effort settings, benchmark subsets, external reports and possible fallback behaviour. Some tests are Kimi's own. That mixture makes the chart useful for locating promising tasks, but insufficient for a single universal 'best model' verdict.

Selected figures in Kimi's launch post. They are vendor-assembled comparisons, not one independent matched-condition test.

Task and measure	K3.1 Pro	Comparator shown by Kimi	How to read it
Terminal-Bench 4.0, agent task success	65.7%	Opus 5.5 66.4%; GPT-6 Astra 58.2%	Kimi run and external figures assembled under differing reporting routes.
Kimi Code Bench 3.0 Long-horizon	61.2%	K3 55.95%	Internal benchmark; same family trend is more informative than cross-vendor rank.
BenchCAD with tools, CAD reconstruction	92.7%	Opus 5.5 96.2%; GPT-6 Astra 95.9%	Kimi used a 1,000-case sample for open models; competitors came from reports.
OfficeQA Pro v2, document reasoning	64.4	Opus 5.5 62.2	A task score, not a measured error rate for financial advice.

An independent check is available, but it answers a different question. On 23 October, Artificial Analysis listed K3.1 Pro at 55 on its Intelligence Index, 77.1 output tokens per second, and $2.11 per Intelligence Index task for the configuration it measured. The last number is benchmark-workload cost, incorporating usage on those tasks; it is not Kimi's list price per million tokens. Artificial Analysis's page currently labels the model proprietary, consistent with the weights not yet being public. Its index cannot validate every CAD, finance or cyber claim in Kimi's chart, and its measured speed should not be assigned to the separately priced Highspeed variant without a separate result.

The price is competitive, but the bill depends on token mix and cache

Kimi prices API use by fresh input, output, cache-hit input and, for K3 models, cache writes with a chosen time-to-live. Its English pricing page says a million tokens means 1,000,000 tokens, excludes applicable tax and is separate from Kimi membership or Kimi Code subscriptions. The default prefix-cache life is five minutes. A repeated prefix can reduce input charges only when it actually hits the cache; the original write is billed separately. Kimi's API guide says Pro access requires at least a $1 top-up and rate limits depend on cumulative top-up. It says thinking is always enabled, with low, high and max effort settings, and max is the default. Those settings can change output-token use and the real bill.

Public US-dollar standard API list rates checked 24 October 2026. Fresh input and output are per 1 million tokens; models and workloads are not quality-matched.

Model	Fresh input	Output	Qualification
Kimi K3.1 Pro	$2	$10	Cache hit $0.20; 5-minute write $2.
Kimi K3.1 Pro Highspeed	$4	$20	Kimi claims 2 to 3 times speed; no matched external latency test here.
Kimi K3	$3	$15	Predecessor; 1,048,576-token API context.
Claude Opus 5.5	$4	$20	1M context; Anthropic cache read $0.20.
Claude Sonnet 5.5	$2	$10	1M context; Anthropic cache read $0.10.
Gemini 4 Argon	$2	$10	Google introductory rate through 30 November; $4/$20 from 1 December.

The direct comparison is sharper than a blanket bargain claim. K3.1 Pro's fresh input and output rates are one-third below K3, and half of Opus 5.5. They match Sonnet 5.5 and Google's temporary Gemini 4 Argon rate. Those are tariff comparisons only; they do not align quality, tool calls, thinking effort, cache policies, latency, rate limits or tax. Anthropic's Opus and Sonnet model pages and Google's Argon pricing note supply the competitor rates.

Zoom

The chart uses one formula for every row: (100,000 ÷ 1,000,000 × fresh-input rate) + (10,000 ÷ 1,000,000 × output rate). For K3.1 Pro that is $0.20 + $0.10 = $0.30. It is a price comparison for a fixed token count, not a measure of work completed or response quality.

For a genuinely long prompt, the arithmetic changes scale: 800,000 fresh input plus 100,000 output tokens costs $2.60 on K3.1 Pro, before any extras. A later request with 600,000 cache-hit input, 200,000 fresh input and 100,000 output would cost $1.52 at Kimi Pro's stated rates: $0.12 plus $0.40 plus $1. That second figure excludes the first cache-write charge and assumes a qualifying, unexpired prefix hit. An agent's total bill also depends on how many requests, failed tool calls and output reasoning tokens the task consumes. The honest procurement metric is cost per successful task on your own workload, not the cheapest row of a token tariff.

3D and finance gains deserve task-level tests

Kimi also positions K3.1 Pro as a production agent for 3D work. It reports 92.7% on BenchCAD with tools, below the Opus 5.5 and GPT-6 Astra figures shown in its chart, and demonstrates scripts that build and revise assets in Blender and CAD tools. Kimi says its open-model BenchCAD test used 1,000 cases sampled from a 17,900-case dataset, while the named proprietary comparators came from their own published material. The 3D Make product can preview and export models, but an attractive demonstration is not evidence that every CAD output is dimensionally correct or ready for engineering use. Test the generated geometry, tolerances and export path on the intended toolchain.

The finance section offers another useful distinction between a task benchmark and real-world reliability. Kimi reports 64.4 on OfficeQA Pro v2 and internal gains against K3 of 7.2 points in financial modelling and 11.6 points in financial research. Its internal Business Operation simulation rises to about 4.2 times K3's score. Different scoring rules and Kimi-run tests make these signals, not a single composite measure of investment accuracy.

Kimi's internal Hallu Bench Hard puts claim-level error at 8.3% for K3.1 Pro, down from 23.4% for K3. The difference is 15.1 percentage points; the relative reduction is about 64.5%, derived as 15.1 divided by 23.4. Kimi explicitly warns that this is a stress test with deliberately difficult prompts, not the error rate of ordinary use. The operational control remains source-grounded extraction, reconciliation against original documents and human review before financial decisions.

The cyber safety result is a controlled comparison, not a safety certificate

Kimi says it removed cyber-related content from mid-training and filtered cyber tasks from its reinforcement-learning mix before the planned weight release. On its internal xdev-bench-hard test, an untreated checkpoint solved 90.5% at pass@3, while the final K3.1 Pro reached 81.0%. That is a 9.5 percentage-point reduction against Kimi's own untreated checkpoint. Pass@3 allows up to three attempts, and Kimi says the random subset had no hard time or output-token limit. This is evidence of a reduction in that setup, not evidence that the released model cannot assist with offensive work.

The UK AI Security Institute and US CAISI assessment from July tested K3, not K3.1 Pro. It found K3 below the leading US closed-weight models on its preliminary cyber evaluations, but also found that K3's safeguards did not stop attempts at exploit development. In one simulated 32-step enterprise network range, K3 reached step 17 on average and completed one of ten attempts within a 100 million-token limit. The range was intentionally vulnerable and lacked active defenders. That earlier independent result gives context for why a K3.1 evaluation matters; it cannot be transferred numerically to the new model.

What changes for developers and the model market

The immediate market pressure is on hosted frontier-model pricing. Kimi is pairing a million-token context allowance with a $2/$10 tariff and a separately priced Highspeed option, while showing credible-looking task gains. But $2/$10 is already a contested tier, not a new floor: Sonnet 5.5 sits there, and Google's Argon introductory offer does too. The differentiation Kimi has to prove is successful work per dollar and per minute on real agent jobs, not the nominal token price alone.

For teams considering local control, the consequential event has not happened yet. Kimi calls K3.1 a 3T-class open-weight model, but until the weight files, model card, checksums and licence are public, a buyer cannot calculate a trustworthy deployment bill or confirm permissible uses. Even after release, a trillion-class model should not be described as a desktop or home-lab option without published formats, memory requirements, throughput and software support. Context-window size is also not a substitute for those measurements.

For UK organisations, the practical decision is narrower: run a small evaluation with real documents, code or 3D jobs, record full token usage, cache hit rate, answer quality, latency and human correction time, and compare that total against the alternatives under the same task and governance controls. Kimi says Zero Data Retention is available to API customers and enterprise administrators must opt in to K3.1 Pro at launch. A security or procurement team should verify the actual account terms and retention setting before sending sensitive material, rather than treating a product label as a contractual control.

A deployment check before moving production work

The question after this launch is not whether Kimi can display a strong benchmark bar. It is whether the same model, under the same tool and token budget as its rivals, completes a buyer's real work accurately and cheaply, and whether the promised weights arrive with terms and deployment evidence that make independent scrutiny possible.

Key facts and source record

Sources read on 23 October 2026: Kimi's technical launch; Kimi API prices and cache rules; Kimi API access and effort guide; Artificial Analysis K3.1 Pro snapshot; UK AISI and US CAISI's K3 assessment; Claude Opus 5.5; Claude Sonnet 5.5; and Google's Gemini 4 Argon pricing note. Kimi's benchmark scores and safety comparison are attributed to Kimi. Artificial Analysis and AISI are independent sources, but the AISI report covers the predecessor K3.