Kimi K3.1 Pro and the 0.7-Point Gap: Reading the Release Sheet Before the Nov 4 Weights

Key takeaways

• Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0 — 0.7 points behind the 66.4% the vendor measured for Claude Opus 5.5 on the same harness, and 7.5 points ahead of GPT-6 Astra at 58.2% (Moonshot's launch post).
• The claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro on the vendor's stress bench — below Opus 5.5's 15.0% — with the stress-test caveat printed in the same paragraph (launch post).
• API pricing is verified on the platform's own page: $2.00/$10.00 per MTok with cached input at $0.20 — a published 90% cache discount — and cache writes billed separately by TTL at $2.00 (5-minute) or $4.00 (1-hour) (Kimi API pricing).
• Moonshot removed cyber-attack data from both training stages — the mid-training corpus and the RL environment mix — before the open-weight release (launch post, cybersecurity section).
• Full weights are vendor-confirmed for Nov 4; until then Artificial Analysis still labels the model proprietary at 2.8 trillion parameters, and the weights license is unnamed (Artificial Analysis model page).

Two numbers on the same launch page decide most of this analysis: 65.7 and 66.4. Kimi K3.1 Pro, the first model in Moonshot's K3.1 series and the company's "most capable 3T-class open-weight model to date" with a 1-million-token context window, scored 65.7% on Terminal-Bench 4.0 — and the same sheet prints Claude Opus 5.5 at 66.4%, measured on Kimi's own harness, with GPT-6 Astra at 58.2% (launch post). The launch leads with the row it narrowly loses. The sheet also prints a claim-level hallucination rate that fell from 23.4% for K3 to 8.3% for K3.1 Pro, a pricing page whose numbers check out line by line, and a Nov 4 date for the full weights. This article reads the release the way a buyer should: what the sheet already settles — capability class, price, workload fit — what only the weights can answer — license, calibration, replication — and the five-check audit to run on November 4, using the weights-day method this audit extends to a release that arrives sheet-first.

The comparator it prints against itself

The three Terminal-Bench 4.0 rows are the disclosure. K3.1 Pro at 65.7, Claude Opus 5.5 at 66.4 as measured by Moonshot on its own harness — the sheet's footnote attributes Opus's score to Anthropic's system card as the source of the comparators — and GPT-6 Astra at 58.2. A 0.7-point gap is inside any reasonable margin of noise, which is the point: no buyer should switch stacks on a gap this small, but the sheet removes the temptation to pretend the model is ahead of Anthropic's flagship. That distinction matters in procurement terms. A 7.5-point lead over OpenAI's agentic flagship is a real capability column; a claim of parity or better against Anthropic's frontier would have been a fabrication the independent index contradicts — Artificial Analysis' composite puts K3.1 Pro at an Intelligence Index of 55, fourth of 215 models, inside the same frontier cluster the proprietary leaders occupy. The vendor's own footnote pages cite its comparators to primary sources — the OpenAI page for Astra, Anthropic's system card for Opus 5.5 — so the sheet's provenance is checkable, not asserted.

The tool-attached scores are printed twice, and the gap is the lesson. On BenchCAD, which tests 3D reconstruction through CAD code generation, K3.1 Pro scores 92.7 with tools, near Opus 5.5's 96.2 and GPT-6 Astra's 95.9 — and 72.8 without tools, against Opus 5.5's 73.0 (launch post). Nineteen points of the published score depend on the tool harness. The same pattern holds on the hallucination bench, where the with-tools rate is 8.3% and the without-tools rate is 5.4%. A team that knows its own serving stack attaches tools to everything should read the with-tools column; a team evaluating a model's standalone behavior should read the other one. Printing both is the disclosure that makes the choice legible.

The independent column prices the work, and it is not free. Artificial Analysis' composite row — measured on Kimi's API — adds the side effects a capability table omits: $2.11 per Intelligence Index task, output speed of 57.5 tokens per second against an 82.3 median, and verbosity of 150M output tokens on the index run against an 87M median. The verbosity is a bill, the same mechanism the Sonnet 5.5 effort-economics analysis isolated: an agent that thinks in more tokens costs more to think. For always-on workloads the slower output speed is an interactive-latency tax; the 3.07s time-to-first-token is better than the 3.90s median, which softens it. A Moonshot answer to that tax is a second tier: K3.1 Pro Highspeed delivers 2–3× the speed of K3.1 Pro for the same interaction class, at $4.00/$20.00 per MTok on the pricing page.

The hallucination halving, stress-labeled

The honest numbers first, then the honest label. On the Kimi Hallucination Benchmark Hard set — claim-level hallucination rate assessed by fact-checking agents on deliberately error-prone, de-identified practical prompts — K3.1 Pro's rate fell from 23.4% for K3 to 8.3%, and the share of responses containing at least one factual error dropped to 42.4% of its predecessor's rate (launch post). On the with-tools column, K3.1 Pro's 8.3% sits below Opus 5.5's 15.0%, far below GLM-5.3's 22.6% and DeepSeek-V4.1-Flash-1017's 23.3%, with only GPT-6 Astra's 5.5% ahead. And the launch post carries its own caveat in the same section: the evaluation is a stress test for factual accuracy, the examples probe situations where models are prone to errors, and the reported rates "do not represent overall error rates in daily use." A buyer who quotes the 8.3% without that label is misquoting the vendor.

The operational read is still strong for document-heavy work. OfficeQA Pro v2 — fiscal-document analysis through retrieval, cross-document verification, and multi-step numeric reasoning — scores a stated state-of-the-art 64.4 for K3.1 Pro, and the vendor's internal finance benches improve by 7.2 points on financial modeling, 11.6 points on financial research, and a 4.2× multiple on long-horizon business operations over K3 (launch post). For a quote reconciliation or invoice audit agent, halving the claim-level error rate on a stress bench while printing the caveat is the difference between a model that needs a verification layer on every open-ended retrieval and one that needs it only where the prompts are adversarial. The always-on agent economics then price the rest.

What the pricing page commits

Every number a cost model needs is on the platform's own pricing page, checked this week. kimi-k3.1-pro bills at $2.00 per MTok input and $10.00 per MTok output, with cached input at $0.20; the Highspeed tier doubles both list prices and keeps the same $0.20 cached rate; all K3-series models carry a 1,048,576-token context window (Kimi API pricing). Billing is flat pay-as-you-go with no tiering by context length, and — the row that changes arithmetic — cache writes are billed separately by TTL: $2.00 per MTok for the 5-minute tier and $4.00 per MTok for the 1-hour tier. The API enables context caching automatically for repeated request prefixes at a default 5-minute TTL. A 1-hour cache write costs exactly twice the input price; a 5-minute write costs the same as input. Long-lived agent sessions should buy the cheap tier and re-derive, not assume one caching cost.

The cache rate is the third published row beside the flash tiers — at a different size class. MiniMax M3.1 Flash lists cached input at $0.015 per MTok (the launch-day read), and DeepSeek's 1017 snapshot computes to roughly $0.017 effective input at its observed cache-hit rate (the $0.017 Agent). Kimi's $0.20 is ten times higher in absolute terms — and it is a flagship-scale row (Artificial Analysis counts 2.8T total parameters) rather than a flash-tier product, with the published discount at 90%. The comparison a mid-market buyer should run is per-task at their workload's hit share, not the per-token list price.

The sheet also carries the migration clock. K2-series models close to new users on October 31, 2026, and the platform plans to retire all of them by the end of December (pricing notice) — every K2 deployment on the platform inherits a K3.1 migration with a deadline attached. One more cost lever is documented on the platform docs: reasoning effort is configured per request at low, high, or max (default max) and thinking can never be fully disabled — effort, not a toggle, is the per-task cost control (K3.1 Pro quickstart).

The release gate you can already read

The cyber-data removal is stated with its mechanism, not just as a slogan. Moonshot observed vulnerability-research capability rising as post-training scaled, traced cyber-related content embedded in the general mid-training corpus and cyber tasks in the RL environment mix, and — in preliminary experiments — found that "without data removal at either the mid-training or the RL stage, downstream cyber capability would increase substantially." For the open-weight release, the company removed cyber-related data from both training stages while retaining general-purpose coding tasks (launch post, cybersecurity section). That is the same gating pattern GLM-5.3 shipped in its staged release — weights after a safety pause (the staged-release precedent) — with one structural difference: K3.1 Pro published the entire detail sheet at launch and set a weights date, no staged pause in between.

October's open-weight releases now form three readable playbooks. MiniMax shipped weights four days after its API launch and let the license file carry the buying conditions — a $20 million revenue line with attribution. GLM-5.3 gated its release behind a safety pause. Kimi K3.1 Pro launched a sheet-first release with the safety gating applied pre-emptively and the weights set Nov 4. What varies is not the capability claim; it is which artifacts land first, and the buyer's audit has to follow the artifacts.

What stays open until Nov 4 — and the audit to run

The sheet answers capability, price, and workload fit. Four things it cannot answer, and the five-check weights-day method covers them the same way it covered MiniMax:

1. The license is unnamed. The MiniMax weights day wrote a revenue threshold and an attribution requirement into commercial self-hosting; K3.1 Pro's terms remain the top open item on November 4, and no self-hosting plan should be priced before the license file is read.
2. No independent calibration column exists. Artificial Analysis' composite measures capability ($2.11 per task at index 55), not the confidently-wrong failure mode — the AA-Omniscience-class row that showed MiniMax's flash tier at 92.8% confidently wrong (the weights-day audit). On open-ended factual retrieval with a verification layer absent, that column is the one the sheet cannot print about itself.
3. Replication under third-party harnesses. The composite index already runs Terminal-Bench 4.0 as one of its ten evaluations; when weights land, community harness reads will either reproduce the sheet's 65.7 or force a harness footnote — the same reconciliation MiniMax's table invited.
4. Card-level parameters and serving recipes. The "3T-class" framing and Artificial Analysis' 2.8T count need the model card's own statement; serving recipes for vLLM and SGLang-class stacks are what turn a download into a deployable.

Artificial Analysis still labels K3.1 Pro proprietary — a status the Nov 4 drop supersedes, and the kind of surface staleness a buyer should expect for days after any weights event.

The one-minute diagram separates the settled from the pending:

The two panels are the split the sheet draws: the amber side is what October 23 already settled with checkable numbers, the violet side is what only the November 4 weights can answer, and the coral strip is the event that closes the gap.

Related reading

• The $20 Million Line in the License: Weights Day Audits MiniMax M3.1 Flash — the sibling audit this sheet-first reading extends: the five-check method, run on a release where the weights landed first
• Inference Economics: Why Always-On Production Agents Are Now Affordable — the per-task cost frame that turns Kimi's published cache tiers into a workload decision
• The $0.017 Agent: DeepSeek's 1017 Snapshot Cuts Task Cost Without Cutting Prices — the other cache-economy row beside Kimi's $0.20: hit share, not the price sheet, sets effective agent cost

A representative build

A 400-person B2B software company runs a code-review and document-audit agent fleet that executed about 12,000 agent tasks last week on a mixed model pool, billed against its own MCP-fronted repository and ERP integrations. The release sheet moves two workloads this week: long-horizon coding tasks that Terminal-Bench 4.0 proxies route to K3.1 Pro at the with-tools 65.7 class, where the $0.20 cached-input rate and the 5-minute default cache TTL fit their repeated-prefix traffic; and financial-document audit runs that OfficeQA-class work proxies route the same way, with reasoning effort pinned to high instead of the default max after the team's own token-cost review. The 1-hour cache-write tier is rejected after the arithmetic — at twice the input price, their sessions re-derive cheaper than they hold. On November 4 the platform team runs the five-check audit before any self-hosting plan: license file first, then parameters, format and recipes, harness deltas against the sheet, and calibration — holding the pinned checkpoint even while serving continues on the vendor's API.

Request a scoped build. One-week discovery. You get a system inventory, workflow map, and fixed scope — whether or not you build with us.