Kimi K3.1 Pro Explained: The Open-Weight Model One Point Behind Claude Opus

On October 23, 2026, Beijing's Moonshot AI released Kimi K3.1 Pro — and made the open-source AI race considerably more interesting. According to Moonshot's own numbers, this soon-to-be open-weight model now sits within a single percentage point of Claude Opus 5.5, the world's top-ranked closed model, on one of the toughest coding-agent tests in the industry.

The headline matters because the gap between open and closed AI models has been the industry's most watched metric. If Moonshot's claims hold — and the weights really ship on November 4 — that gap shrinks to almost nothing.

Here is what Moonshot actually launched, what the numbers really say, and what is still unproven.

What Moonshot actually launched

Kimi K3.1 Pro is not a brand-new base model. Moonshot describes it as a post-trained version its 3-trillion-parameter-class predecessor released in July 2026, put through far more reinforcement learning across what the company calls "hundreds of downstream domains." The scale is enormous:

• Architecture: mixture-of-experts, roughly 2.8 trillion total parameters with 104 billion active at any moment
• Context window: 1 million tokens — enough to hold an entire codebase or a stack of long documents
• Inputs: text and image; it outputs text
• Availability: live now on Kimi.ai, mobile apps, Kimi Work, Kimi Code, and the Kimi API (model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed); Kimi Enterprise admins must enable it first

Alongside it comes Kimi K3.1 Pro Highspeed, which Moonshot says delivers the same capabilities at two to three times the speed.

The benchmarks — and the caveats that come with them

The number everyone is quoting is Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks with strict pass/fail checks. Moonshot reports:

| Model | Terminal-Bench 4.0 | |---|---| | Claude Opus 5.5 | 66.4% | | Kimi K3.1 Pro | 65.7% | | GPT-6 Astra | 58.2% |

Important context, which Moonshot itself discloses: the company ran its own model inside the Claude Code harness for these tests, while rival scores were taken from public leaderboards or the rivals' own reports. These are Moonshot's compilations, not an independent head-to-head.

There is, however, a genuinely independent data point: Artificial Analysis scored K3.1 Pro at 55 on its Intelligence Index, placing it 4th of 215 models — behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Meta's Muse Wonder Full (56). It gained 11 points over K3 on the same index. Independent evaluator Nathan Lambert, who writes the Interconnects newsletter on open models, said the scores "look to be clearly the strongest current open weight model, if the weights are released as they claim by November 4th" — an appropriately conditional endorsement.

Other Moonshot-reported figures worth noting honestly:

• BenchCAD (3D CAD coding with tools): 92.7% for K3.1 Pro, behind Opus 5.5 (96.2%) and GPT-6 Astra (95.9%) — close, but not a win
• OfficeQA Pro v2 (financial document analysis): 64.4, which Moonshot claims is a state-of-the-art score
• GDPval-AA (agentic): 1,846 Elo, up 312 points from K3
• Hallucination rate: fell from 23.4% for K3 to 8.3% on Moonshot's own hardest factuality test

Why the November 4 date is the real story

Right now, K3.1 Pro is a closed model in practice: you can only reach it through Moonshot's apps and API. The promise is that the full model weights will be published by November 4, 2026, which would make it the highest-scoring open-weight model on the Artificial Analysis Index — and narrow the open-versus-proprietary gap on that index to just 3 points.

Note on terminology: "open-weight" is not "open source" — the training data, code, and recipe stay proprietary. But weights are what matter for the community: once they ship, researchers and developers can run the model on their own infrastructure, audit its behavior, fine-tune it, and benchmark it independently. That independent testing is also what will finally verify — or puncture — Moonshot's benchmark claims.

One detail to watch: Moonshot has not yet announced the license for the weights. Some "open" model licenses restrict commercial use or require attribution. Until the license is published, businesses should not budget around self-hosting it.

The safety subplot: Moonshot stripped the hacking data

One of the more unusual parts of the announcement concerns cybersecurity. Ahead of publishing open weights — which let anyone run the model without Moonshot's oversight — the company says it removed cyber-related data from mid-training and from the reinforcement-learning environments.

Moonshot quantified the cost of that decision on its own unpublished exploit test, xdev-bench-hard: a checkpoint trained without the safety filtering solved 90.5% of the test's problems, while the final model solves 81.0%. That 9.5-point gap is the company's own measure of what the safety step took away.

Be honest about the limits here: the test is unpublished, so nobody outside Moonshot can reproduce it — and once weights are public, fine-tuning can potentially restore the removed capability. As a gesture it is meaningful; as a guarantee it is not.

The pricing: $2 in, $10 out

The Kimi API lists K3.1 Pro at $2 per million input tokens and $10 per million output tokens — about 33% cheaper than K3 across input, output, and cache-hit rates, and (per coverage comparing it to Opus) roughly half the price of Anthropic's flagship. The Highspeed tier costs roughly double the standard token rate but keeps cached-input pricing the same, making it the natural choice for latency-sensitive agent loops that re-read long system prompts.

That undercuts the flagship closed models while matching them on the hardest agent tests — and self-hosting after November 4 could push inference costs lower still for teams with their own GPU capacity.

What remains uncertain

• Independent verification: every capability number above comes from Moonshot or outlets relaying Moonshot's numbers, except the Artificial Analysis score. The real test starts when the weights ship and third parties run their own benchmarks.
• The license: still unannounced; it determines whether this is genuinely open for business use. (Launch-day social media also produced the usual unverified viral claims — treat unsourced posts as noise, not evidence.)
• The predecessor's cautionary tale: this release follows the same playbook as MiniMax's recent M3.1 Flash launch — aggressive pricing, company-reported benchmarks, weights "coming soon." Read the breakdown of how that one played out. Kimi's advantage is that the open-weights promise makes verification automatic.
• Highspeed's real-world value: 2–3x faster is Moonshot's claim; actual latency and reliability under load remain untested by third parties.

The honest bottom line

Kimi K3.1 Pro is a real, available model with a credible (if company-reported) near-frontier result, a 1M-token context, and pricing that undercuts the flagship closed models. The launch is genuinely significant for one reason: if the weights ship on November 4 as promised, the open-versus-closed gap becomes a rounding error — and independent testing can finally confirm what Moonshot is claiming.

Try it cheaply through the API now, watch what the November 4 weights release actually includes (especially the license), and let independent benchmarks — not launch-day posts — decide whether this is the moment the open-weight race effectively ended. For the other open-weight contender aimed at agent workloads, see DeepSeek V4.1-Flash-1017 — MIT-licensed weights, already available.

Sources: SiliconANGLE · GenAInews · Oossa · Madrobot · WebProNews (cyber-data removal) · ExplainX