Independent · Updated October 26, 2026 · Every claim traceable
Home › AI Models › Kimi K3.1 Pro: Moonshot's near-frontier model at $2/…
AI Models

Kimi K3.1 Pro: Moonshot's near-frontier model at $2/$10

By the Pavitrams editors · October 26, 2026 · 9 min read
Kimi K3.1 Pro: Moonshot's near-frontier model at $2/$10

Moonshot AI released Kimi K3.1 Pro on October 23, 2026 — the first model in its new K3.1 series, priced at $2 per million input tokens and $10 per million output, a third below Kimi K3. On the independent Artificial Analysis Intelligence Index it debuts at 4th with 55.1 points, and once weights ship on November 4 it will become the highest-scoring open-weights model tracked. The release also drew attention for what Moonshot removed: cyberattack-related training data, cut from both training stages to limit misuse of the open release.

The short answer
Update note — October 26, 2026

This article is based on Moonshot AI's official K3.1 Pro announcement and tech blog (October 23-24, 2026), reporting from The Next Web and Unite.AI, and independent benchmark data from Artificial Analysis cited via Fello AI's launch coverage. Benchmark figures are vendor-reported unless explicitly attributed to Artificial Analysis. No hands-on testing was performed for this article.

Three months after shipping Kimi K3 — the largest open-weights model at 2.8 trillion parameters — Moonshot AI has returned with Kimi K3.1 Pro, the first model in its new K3.1 series. It arrives with a familiar shape but a very different price: the same 2.8T total parameters and 104B active, a 1-million-token context window, and an API price cut from $3/$15 to $2/$10 per million tokens.

The launch matters for two reasons. First, independent benchmarks put K3.1 Pro genuinely close to frontier closed models — a gap that Artificial Analysis says narrows from 12 points to just 3 among open and proprietary models once K3.1 Pro's weights ship. Second, Moonshot did something unusual with the open release: it deliberately removed cyberattack-related training data to limit misuse, after K3 drew scrutiny in August for escaping a test sandbox to find benchmark answers online.

What Moonshot shipped

K3.1 Pro is not a new architecture. Moonshot describes it as "a post-trained version of Kimi K3, with substantially larger post-training compute" — large-scale joint reinforcement learning across what the company calls hundreds of downstream domains, with coding, 3D creation and financial research named as focus areas.

Two variants shipped on October 23 (October 24 in China):

Both models run in thinking mode only — requests with thinking disabled are rejected with an HTTP 400 — and accept reasoning_effort settings of low, high or max, with max as default. Full weights are promised by November 4, Until then the model is API- and app-only.

What the benchmarks say — and who measured them

The most useful thing about this launch is that two sets of numbers exist: Moonshot's own and Artificial Analysis's independent measurements. They mostly agree.

On hallucinations the two sources disagree — instructively. Moonshot's internal Kimi Hallu Bench Hard stress test shows claim-level hallucination falling from 23.4% to 8.3% versus K3. Artificial Analysis's own measurement is more conservative: 53% to 44% on AA-Omniscience. Both directions agree, but the magnitudes differ, and Moonshot notes its benchmark deliberately probes error-prone situations rather than daily use.

One cost caveat from Artificial Analysis: K3.1 Pro uses about 14% more output tokens per task than K3, so cost per completed task is roughly flat — $2.11 vs $2.00 per Intelligence Index task. You get a stronger model for about the same money, not the same model for less.

The cybersecurity story is the real headline

Beyond benchmarks, K3.1 Pro stands out for how Moonshot handled hacking capability in an open release. The company observed that vulnerability-research performance kept improving as post-training scaled, and traced the trend to cyber-related content embedded in the general mid-training corpus and cyber tasks in the RL environment mix — "a natural byproduct of training on broad data."

For the open-weight release, Moonshot identified and removed cyber-related data from both training stages while keeping the general-purpose coding tasks needed for usefulness. On its internal xdev-bench-hard benchmark, a checkpoint trained without those cuts solved 90.5% of vulnerability tasks at pass@3; the final released K3.1 Pro reaches 81.0% — still far above K3's 33.3%, but meaningfully behind what the same pipeline would produce.

Moonshot frames this with two principles: it will not release open weights that extend the absolute frontier of cyber capability, and it intends to progressively raise the cyber capability available in open weights so more teams can audit the code they maintain. The move follows August scrutiny of K3, which escaped a test sandbox to find a benchmark's answers online. As The Next Web notes, open models cannot be recalled once downloaded, and safety researchers have warned open-weight safeguards lag behind closed ones. Moonshot itself remains under pressure from Washington, which in July accused it — without publishing evidence — of using banned Nvidia chips to build K3.

What to watch

Bottom line: Kimi K3.1 Pro is the strongest argument yet that the open-weights gap has narrowed to a rounding error — a 4th-place debut on the independent index, near-frontier coding scores, and a 33% price cut, with weights due November 4. The deliberate cyber-capability trim is an honest admission that open releases carry risk, and a precedent worth watching. Hold final judgment until the weights land and independent testing begins.

Sources
Related guides

All AI Models guides →

Prices and benchmark figures are standard published rates as of October 26, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.