Kimi K3.1 Pro: Open Model Nearly Matches Claude Opus 5.5

Moonshot AI released Kimi K3.1 Pro on 23 October 2026, and it has done something no open-weights model has done before: within hours of launch, independent benchmarker Artificial Analysis placed it fourth on its Intelligence Index with a score of 55 — just three points behind Anthropic’s Claude Opus 5.5 at 58. The catch: the weights are not out yet. Moonshot has promised the full download by 4 November 2026, and until then the model only runs through Kimi’s own apps and API.

Quick answer: Kimi K3.1 Pro (23 Oct 2026): Moonshot's 2.8T open model scores 55 on Artificial Analysis's index — fourth, three points behind Claude Opus 5.5. Weights due 4 Nov.

This matters because the open-vs-closed gap has been measured at around 12 index points for years. If the weights land on schedule, Artificial Analysis expects the gap between the best open model and the best proprietary model to shrink from 12 points to three — the first time a soon-to-be-open model has reached this level against frontier systems.

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is not a brand-new base model — it is Kimi K3 with far more post-training. Moonshot describes it as a post-training of K3, pushed through far more reinforcement learning across what the company calls “hundreds of downstream domains”. The skeleton is unchanged from K3: a mixture-of-experts architecture with roughly 2.8 trillion total parameters (104 billion active at any one time), a 1-million-token context window, and text plus image inputs. Under the hood, WCCFTech reports the architecture has 69 layers, with a Gated Multi-Head Latent Attention layer placed after every three layers.

It is live right now in the Kimi app, the Kimi Work desktop app, the Kimi Code coding tool and the Kimi API — with a companion variant, Kimi K3.1 Pro Highspeed, that Moonshot says delivers the same abilities at two to three times the speed. The API identifiers are kimi-k3.1-pro and kimi-k3.1-pro-highspeed; on business plans, an admin has to enable the model first.

The honest caveat up front, per this site’s testing methodology: this article is desk-researched from dated reporting, verified on 24 October 2026. I have not run the model hands-on, and several of the most impressive numbers below are Moonshot’s own, not independently reproduced.

How close is it to the frontier, really?

It depends on which number you read — and who measured it. Moonshot’s own benchmark table and Artificial Analysis’s independent testing agree on the headline but differ on the details. Here is what both say:

Evaluation Moonshot’s own numbers (vendor-reported) Artificial Analysis (independent)
Artificial Analysis Intelligence Index — 55 — 4th overall, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56)
Terminal-Bench 4.0 (coding agents) 65.7% (vs 66.4% Claude Opus 5.5, 58.2% GPT-6 Astra) 67.2% (vs 13% for Kimi K3 — the single biggest jump)
GDPval-AA (economically valuable tasks, Elo) — 1,846 — 3rd, behind only Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857); up 312 over K3
AutomationBench-AA (agentic SaaS workflows) — 73%, up from 58% for K3
BenchCAD (3D-object CAD code, with tools) 92.7% (vs 96.2% Opus 5.5, 95.9% GPT-6 Astra, 92.6% Claude Fable 5.1) —
OfficeQA Pro v2 (financial-document test) 64.4, claimed state-of-the-art —
Internal hallucination stress test (claim-level error) 8.3%, down from 23.4% for K3 AA-Omniscience Index rises 20 to 30; hallucination rate 53% to 44%

Two independent readings make Moonshot’s own table easier to believe. Artificial Analysis tested K3.1 Pro at its default maximum effort within half an hour of launch and independently reproduced the story: +11 index points over K3’s 44, and +2 over GPT-6 Astra and Claude Fable 5.1 (both at 53) — and notably also +2 over Gemini 4 Argon at high (53). For context on the rival camp, see our Gemini 4 Argon vs Carbon comparison, which tracks how Google’s own model lineage is shaping up.

“TLDR on the kimi k3.1 pro benchmarks is that it looks like they figured out their RL, where the initial release was a very impressive base model with lighter post-training. China is very much keeping pace with capability gains.” — Nathan Lambert, post-training researcher, on X, 23 October 2026 (via Handy AI)

But the “awkward facts” checklist from mer.vin deserves equal weight: the weights are not out (so “open-weight” is a promise until 4 November), there is no exact parameter count, there is no separate technical report, rival scores in Moonshot’s table mix leaderboard maxima, system cards and Kimi’s own runs — and there is one glaring soft spot: on Terminal-Bench Science, Kimi K3.1 Pro scores 38.1 against 63.3 (Claude Opus 5.5) and 68.1 (GPT-6 Astra). The headline claim does not cover it. Treat vendor-run comparisons as directional, not definitive.

Why the “open” label is doing the heavy lifting

The most consequential decision Moonshot announced was not a benchmark — it was a deliberate weakening of its own model. Moonshot says it identified and removed cyber-related data from K3.1 Pro’s mid-training and reinforcement-learning stages after an unredacted version solved 90.5% of its internal Linux kernel, Node.js and Exim exploit tasks — and still scored 81% on the xdev-bench-hard test, according to WCCFTech.

The stated principle is blunt: do not release open weights that would extend the absolute frontier of cyber capability. K3.1 Pro’s offensive cyber capability is, by design, below what Moonshot’s pipeline can produce, and remains behind proprietary models public for at least three months. On Hacker News, the reception split exactly there: readers praised the 23.4% to 8.3% hallucination drop as “a very nice improvement” while others criticised the capped open weights (via Handy AI).

This is also why the open-weights promise matters structurally. For two years the working assumption has been that open weights trail the frontier by a generation and catch up six months later at a discount. K3.1 Pro compresses the lag to a single model version: a post-training pass, not a new architecture, closed 11 index points. If the weights ship on schedule, it becomes the highest-scoring open model available by a wide margin — and the price of running a near-frontier model stops being a function of what a US lab decides to charge.

Pricing: 33% cheaper than K3

The API is cheaper than K3’s across the board — and the cheaper model is the stronger one. Kimi K3.1 Pro is priced at $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens — roughly a 33% cut versus K3 on all three rates, and level with GPT-6.1 Sol, according to Skimwire and GenAInews. The Highspeed tier runs two to three times faster at $4/$20 — double the rate — while cached input costs the same in both tiers. Artificial Analysis puts cost per task roughly equal to Kimi K3 at $1.98, since K3.1 Pro uses 55K output tokens per index task (14% more than K3’s 48K) while first-party token prices are a third lower.

One caveat from mer.vin: global K3.1 Pro rates were not independently readable at launch time, so treat the published price list as Moonshot’s stated pricing until the billing pages can be verified. The model remains behind only three closed models on the index — and half the price of Claude Opus per one widely-cited launch-day post — which is exactly the discount-and-performance trade-off open models are supposed to offer.

How to try it now (and what to wait for)

You can use K3.1 Pro today without waiting for the weights — but you cannot run it on your own hardware yet. Three routes, pick by what you need:

Route Who it is for The honest note
Kimi app / Kimi Work desktop / Kimi Code Testing it like an end user or coding assistant The model picker already includes K3.1 Pro; speed was K3’s main complaint — early HN testers report nearly double the speed
Kimi API (kimi-k3.1-pro / kimi-k3.1-pro-highspeed) Developers plugging it into agent loops Use cached-input pricing ($0.20/1M) for repeated system prompts; Highspeed for latency-sensitive loops at double price
Wait for 4 November weights Self-hosting teams, auditors, fine-tuners License unspecified as of launch; expect the cyber-capped variant, not the unredacted pipeline output

My read: treat the next two weeks as an evaluation window, not a migration. Run your hardest agentic tasks against it through the app or API now, compare against your current model on your own work (not just public benchmarks), and watch three checkpoints: (1) whether the weights actually land by 4 November, (2) whether independent evaluators reproduce the 55 index score once the community can run its own tests, and (3) whether the Terminal-Bench Science soft spot (38.1 vs 63.3/68.1) matters for your use case. The long-context side also showed a small give-back — AA-LCR dipped from 89% to 86.3% — which matters if you are buying it for the 1M-token window.

Our take: the post-training proof point

Kimi K3.1 Pro’s real story is not “China catches up” — it is that post-training, not scale, did the catching up. The skeleton is the same 2.8T/104B-active model as K3; the gains sit almost entirely in agentic work — GDPval-AA up 312 Elo, AutomationBench 58% to 73%, Terminal-Bench 13% to 67.2% — exactly where multi-step execution and tool use live, while raw-knowledge evaluations barely moved (HLE 47% to 52.5%, SciCode 59% to 60.8%). That is a model taught to work rather than a model taught more, and it vindicates the RL-heavy approach Nathan Lambert flagged within hours of launch.

The honest limits: every eye-catching number except the Artificial Analysis index is vendor-run and unaudited; the science gap is real; and the weights — the thing that makes this an open-weights story at all — are a promise with a date, not a download. If the date holds and the community’s reproductions agree with AA’s 55, this is the most important open-model release of the year. If either slips, it is a strong API-only model with an asterisks section. Either way, it is the clearest signal yet that the frontier is no longer a closed club.

Frequently asked questions

When will Kimi K3.1 Pro’s weights be released?

Moonshot has promised the full weights by 4 November 2026, about eleven days after the 23 October launch — mirroring Kimi K3, whose weights followed the announcement by 11 days. They are not published yet, and Moonshot has not specified the license terms.

How much does the Kimi K3.1 Pro API cost?

$2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens — about 33% cheaper than Kimi K3 on all three rates. The Highspeed variant costs $4/$20 and runs two to three times faster, per Moonshot’s stated pricing.

Is Kimi K3.1 Pro better than GPT-6 Astra?

On the Artificial Analysis Intelligence Index, yes: 55 vs 53 for GPT-6 Astra — and on Moonshot’s own Terminal-Bench 4.0 numbers, 65.7% vs 58.2% (vendor-reported; rival scores from published reports). But K3.1 Pro trails Astra badly on Terminal-Bench Science (38.1 vs 68.1), so “better” depends entirely on the task.

Can I use Kimi K3.1 Pro today?

Yes — it is live in the Kimi app, the Kimi Work desktop app, the Kimi Code coding tool and the Kimi API under the identifiers kimi-k3.1-pro and kimi-k3.1-pro-highspeed. On business plans, an admin must enable it. Self-hosting requires the open weights, due by 4 November 2026.

Why did Moonshot remove cyber data from the model’s training?

Moonshot’s stated principle is not to release open weights that would extend the absolute frontier of cyber capability: after an unredacted version solved 90.5% of its internal exploit tasks, the lab removed cyber-related data from mid-training and reinforcement learning so the open release stays deliberately below what its pipeline can produce.

Sources and methodology

Claims in this article rest on reporting and independent evaluation published on 23–24 October 2026, verified on 24 October 2026: Artificial Analysis — “Kimi K3.1 Pro launches at 4th on the Intelligence Index” (independent testing, index v4.3.2); WCCFTech — launch report (architecture detail, cyber-data removal, exploit-task figures); Handy AI — “Model Drop: Kimi K3.1 Pro” (independent index figures, community reaction, Lambert quote); mer.vin — “Kimi K3.1 Pro and Highspeed: API, Pricing and Benchmarks” (critical checklist, Terminal-Bench Science gap); Skimwire — launch summary (pricing, availability); Intelligent Living — AA-measured comparison table. Vendor-reported benchmarks (Moonshot’s own tables) are labelled as such; rival scores in those tables come from leaderboard maxima, system cards or Moonshot’s own runs, not controlled comparisons. No hands-on testing was performed. Correlation (post-training effort vs agentic gains) is stated as the publication’s reading of the data, not a controlled finding.

Leave a Reply

Your email address will not be published. Required fields are marked *