Kimi K3.1 Pro Is Live: $2 API, 4th on AI Index, Open Weights Nov 4

Moonshot AI has dropped its most capable model to date — and priced it like it wants everyone building with it. On October 23, 2026, the Beijing lab launched Kimi K3.1 Pro, the first model in its new K3.1 series, alongside a faster companion variant, Kimi K3.1 Pro Highspeed. The API price: $2 per million input tokens, $10 per million output, with cached input at $0.20 — 33% below K3. The full weights arrive by November 4, which would make K3.1 Pro the highest-scoring open-weights model on the AArtificial Analysis Intelligence Index>

Quick answer

• What it is: Kimi K3.1 Pro, Moonshot AI's first K3.1-series model — a 3T-class open-weight reasoning model with a 1-million-token context window, built as a post-trained version of Kimi K3.
• Launch: October 23, 2026 (coverage dates span Oct 23–24 across timezones), with a 2–3x-faster Highspeed variant at launch.
• How it ranks: 55 on the Artificial Analysis Intelligence Index — 4th overall, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56).
• The open-weights angle: full weights by November 4, 2026; once published, the gap between the leading open-weights and proprietary models narrows from 12 points to 3.
• The catches: nearly all the headline scores were run by Moonshot itself; the weights are still a promise; and thinking mode is mandatory — you cannot turn reasoning off.

What Kimi K3.1 Pro actually is

K3.1 Pro keeps the K3 architecture and spends its upgrade entirely on training. It is a post-trained version of Kimi K3, built with what Moonshot describes as substantially larger post-training compute — reinforcement learning run across hundreds of downstream domains, with the RL framework, long-context inference, sandbox infrastructure and rollout scheduling all re-optimized.

Artificial Analysis reports 2.8 trillion total parameters with 104 billion active per token (about 1.8% activation), a 1-million-token context window, and text and image input. One design choice stands out: K3.1 Pro only thinks. The Kimi Code documentation says requests with thinking disabled are rejected with an HTTP 400 error; reasoning effort can be low, high or max, with max as the default. Every score Moonshot reports was produced at max effort, which tends to flatter demos more than everyday usage.

The pricing: $2/$10, and what it actually buys

The rate card comes from Moonshot's own API pricing pagege, so it is vendor-stated but current. For kimi-k3.1-pro: input $2.00, output $10.00, cached input $0.20 per million tokens. For the Highspeed variant the price doubles on input and output ($4.00/$20.00), with cached input staying at $0.20. Under the hood, the Highspeed variant consumes roughly twice the quota.

The fine print, straight from the pricing page:

• Kimi K3 cost $3/$15/$0.30 — so K3.1 Pro is 33% cheaper on all three rates while scoring meaningfully higher.
• Billing is pay-as-you-go with no API subscription (separate from the Kimi membership and Kimi Code tiers). Prices exclude taxes, calculated at checkout by jurisdiction.
• Context caching is automatic on repeated request prefixes with a 5-minute default TTL; for K3-series models, cache writes are billed separately by TTL tier (5 min / 1 hour), while cached input itself is billed at the $0.20 cache-hit rate.
• File-related APIs — file content extraction and file storage — are temporarily free.

Independent corroboration exists: Artificial Analysis lists cost per Intelligence Index task at about $1.98, roughly equal to Kimi K3. The nuance is that K3.1 Pro used 55K output tokens per task — 14% more than K3 — while its token prices are 33% lower. The model is wordy: it spends more tokens to get the better answer, and the price cut offsets the verbosity.

The benchmarks: where it wins, where it trails

Moonshot's launch table puts K3.1 Pro level with or ahead of the closed frontier on many rows — but read the labels. Almost every score was run by Moonshot itself, inside Anthropic's Claude Code harness for the coding tests, with reasoning effort at max. Artificial Analysis ran independent checks on some axes. Both pictures are worth having.

Terminal coding — the headline win

On Terminal-Bench 4.0, Moonshot reports 65.7% — ahead of GPT-6 Astra at 58.2% and a hair behind Claude Opus 5.5 at 66.4%. Artificial Analysis independently measured 67% on the same bench. K3 managed only 13% there, so this is the single biggest jump in the release — though much of the coding evaluation was run inside Anthropic's Claude Code harness, an environment built around Anthropic's own models.

Agentic work — the real strength

Agentic benchmarks — tool use, multi-step execution — are where K3.1 Pro gains the most. On GDPval-AA, Artificial Analysis measured 1,846 Elo, up 312 from K3, ranking 3rd behind Opus 5.5 (1,866) and Muse Wonder Full (1,857). AutomationBench-AA rose from 58% to 73%. These are the strongest independent confirmations in the release.

Office, finance, 3D and the rest

On Moonshot's internal OfficeQA Pro v2, the model scored 64.4 — the top result Moonshot claims among tested models. Finance benchmarks moved +7.2 points on Financial Modeling, +11.6 on Financial Research, and roughly 4.2x on Business Operation versus K3, all company-reported. On BenchCAD (3D reconstruction via CAD code), it hit 92.7% with tools — close, but still behind Opus 5.5's 96.2% and GPT-6 Astra's 95.9%. A 3D Make feature is live on kimi.ai and the Kimi app: preview and edit 3D models online, download them ready for 3D printing.

Elsewhere the gains are moderate: Humanity's Last Exam 47% → 52%; HLE-Full 70.0%; BrowseComp 92.1%; GPQA-Diamond 94.2%; SWE Marathon 51.7%; DeepSWE 75.5%. On the internal Kimi Hallu Bench Hard stress test, claim-level hallucinations fell 23.4% → 8.3% — but Moonshot itself warns the test probes error-prone situations deliberately and doesn't represent daily-use error rates. Artificial Analysis's Omniscience check puts hallucination at 53% → 44% with accuracy 48% → 51%.

The cybersecurity section everyone is reading

The launch post spends a whole section on cybersecurity — unusual for a model launch, and the most distinctive part of the release. Moonshot says it noticed vulnerability-research capability climbing as post-training scaled, traced it to cyber-related content in the training corpus and RL environment mix, and removed cyber-related data from both training stages before the open-weight release. On the in-house xdev-bench-hard test, a checkpoint trained without the risk-reduction measures solves 90.5% at pass@3, while the shipped K3.1 Pro reaches 81.0% — and Moonshot says the model's cyber capability stays behind frontier proprietary models that have been publicly available for at least three months.

The blog frames two open-weight principles: never release open weights that would extend the absolute frontier of cyber capability, and progressively raise the cyber capability available in open weights anyway — arguing that non-discriminatory access gives more teams the means to audit the code they maintain. Take it as the company's stated policy, not a verified fact; independent auditors will decide whether 81.0% is the right line.

Where you can actually use it

Both models are live now across the Kimi product family:

• kimi.ai and the Kimi mobile apps (iOS, Android, HarmonyOS)
• Kimi Work desktop app, version 3.1.0 or later (Windows and Apple Silicon Macs)
• Kimi Code — select with the /model command
• Kimi API — model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with Zero Data Retention available for API customers
• Kimi Enterprise — admins can enable it; access is off by default at launch

The membership fine print matters for app users: under the new plans, k3.1-pro requires Plus and above, with the full 1M-token context reserved for Pro and above, while k3.1-pro-highspeed requires Pro and above (legacy plans: Moderato and Allegretto). API users bypass all of this — pure pay-as-you-go.

One practical note from the docs: start a new session rather than switching to the new models mid-session, and enter exact model IDs in third-party tools — a mistyped ID silently falls back to K2.8 Preview without thinking. And a deprecation clock is running: K2-series models stop being available to new users on October 31, 2026, and retire for everyone by end of December.

K3.1 Pro vs the frontier, compared

The comparison to note: Sonnet 5.5 and K3.1 Pro sit at the same $2/$10 price tier, with Sonnet one index point ahead — but K3.1 Pro will be the open-weights option. Once the weights drop, Artificial Analysis expects the open-vs-proprietary gap on the index to narrow from 12 points to 3. A word on the "highest-scoring open-weights" framing: it is AA's projection, contingent on the weights actually landing by November 4 and nothing else shipping in the meantime.

Pros and cons

Pros

• Genuine frontier-adjacent scores at a third less than its predecessor. 4th on the Intelligence Index at $2/$10 is a real value proposition for agent workloads.
• Agentic and terminal-coding strength. The independent GDPval-AA and Terminal-Bench numbers are the hardest scores to fake.
• Open weights by November 4. Self-host, fine-tune, or distill — the weights remove the vendor lock-in the price list implies, joining the open-release wave that includes Meta's downloadable Muse Spark Open and DeepSeek's open-source DeepInfer.
• 1M context with caching economics. $0.20 cached input makes long-context document and repo work affordable.

Cons

• The best scores are company-run. Moonshot ran nearly everything itself, at max reasoning effort, inside Anthropic's Claude Code harness. Independent verification is thin.
• It is verbose. 14% more output tokens per Intelligence Index task than K3 — the price cut partly offsets a habit that costs you tokens.
• Thinking cannot be disabled. You pay reasoning overhead on every request, including simple ones where a non-reasoning model would be faster and cheaper.
• The weights are a promise. "By November 4" is a commitment, not a release. Until then, it is an API product with frontier-level opacity.
• China-origin trust surface. Moonshot is a Beijing company; teams with data-residency or provenance policies should note the ZDR option covers the API path, not self-deployed geopolitics.

FAQ

How much does Kimi K3.1 Pro cost per million tokens?

$2 per million input tokens and $10 per million output tokens, with cached input at $0.20 — all pay-as-you-go, taxes extra. The Highspeed variant costs $4/$20. For comparison, Kimi K3 cost $3/$15.

When do the Kimi K3.1 Pro weights get released?

Moonshot says the full weights will be released by November 4, 2026. Once published, Artificial Analysis expects it to become the highest-scoring open-weights model on its Intelligence Index.

Is Kimi K3.1 Pro better than Claude Opus 5.5?

Not on the independent index: Opus 5.5 scores 58 vs K3.1 Pro's 55, and leads narrowly on most of Moonshot's own benchmark table. K3.1 Pro's edges are terminal-coding, agentic SaaS workflows, and price — one slot from the top, but not yet the top.

Can you turn off thinking mode in K3.1 Pro?

No. The model supports thinking mode only, and API requests with thinking disabled are rejected with an HTTP 400 error. Reasoning effort can be set to low, high or max (max is the default), but there is no non-reasoning mode.

Where can I use Kimi K3.1 Pro right now?

kimi.ai, the Kimi mobile apps, Kimi Work desktop 3.1.0+, Kimi Code via the /model command, and the Kimi API under the IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Enterprise admins can enable it, with access off by default. App-side access requires Plus or above; the 1M context and Highspeed variant need Pro.

Key takeaways

• Moonshot AI launched Kimi K3.1 Pro on October 23, 2026 — its first K3.1-series model, priced at $2/$10 per million tokens, 33% below K3.
• It ranks 4th on the Artificial Analysis Intelligence Index at 55, behind only Claude Opus 5.5, Sonnet 5.5 and Muse Wonder Full — with agentic and terminal-coding benchmarks as its strongest suit.
• The faster Highspeed variant doubles API price for 2–3x speed with identical capability; thinking mode is mandatory and cannot be turned off.
• Full weights land by November 4, 2026, which would make it the top-scoring open-weights model and shrink the open-vs-proprietary gap to 3 index points.
• Moonshot stripped cyber-related training data before the open-weight release — the rare launch where the safety section is the most interesting part.

October 24, 2026 — The pattern of the month is unmistakable: the frontier keeps getting cheaper, and the open-weights gap keeps shrinking. Whether K3.1 Pro holds its #4 spot depends on one date: November 4.