Moonshot’s Kimi K3.1 Pro Nearly Matches Claude Opus 5.5 on Coding, but the Weights Arrive on Nov. 4

Moonshot AI has released Kimi K3.1 Pro, the first model in its new K3.1 series, and the headline number is the kind open source communities dream about. On Terminal Bench 4.0, a test of coding agents in constrained terminal environments, the model scored 65.7 percent. That is 7.5 points ahead of GPT-6 Astra at 58.2 percent and 0.7 points behind Claude Opus 5.5 at 66.4 percent. Moonshot calls it the most capable open weight model it has built, with 2.8 trillion total parameters, 104 billion active parameters per pass, and a context window of 1,048,576 tokens.

The model is live today on Kimi.ai, the Kimi mobile apps, Kimi Work, Kimi Code and the Kimi API. There is also a Highspeed variant that the company says runs two to three times faster. The full model weights, the part that makes the open weight claim real, are promised by November 4.

That gap between available now and open is where the interesting details sit. Moonshot produced its coding numbers with the Claude Code harness, its finance and hallucination results come from internal tests, and its safety chapter describes removing capability the model had already acquired. None of that makes the claims false. It does mean most of them still need a second pair of eyes.

What Moonshot actually shipped

Kimi K3.1 Pro is a post trained version of Kimi K3, not a new architecture. Moonshot says the work went into scale: joint reinforcement learning across hundreds of downstream domains, plus changes across the RL framework, long context inference, training efficiency, sandbox infrastructure and rollout scheduling. The pitch is that every extra unit of training compute now returns a predictable gain, and that the gains held across coding, 3D modelling, finance and agentic work.

The API model names are kimi-k3.1-pro and kimi-k3.1-pro-highspeed, per the Kimi API documentation. Moonshot says zero data retention is available for API customers, and that Kimi Enterprise administrators have to switch the model on for their organisation because it ships off by default.

One number matters more than the rest for anyone planning to self host: November 4. Until that date, K3.1 Pro is a hosted model with an open weight promise. That is a different product from an open weight model you can download, inspect and run on your own hardware.

The benchmark table, and the fine print around it

The release post is dense with scores. On Terminal Bench 4.0, K3.1 Pro reached 65.7 percent against 66.4 percent for Claude Opus 5.5 and 58.2 percent for GPT-6 Astra. On Kimi Code Bench 3.0 Long horizon, an internal test of multi step engineering work, it scored 61.2 percent against 55.95 percent for Kimi K3.

The 3D results are the ones that will raise eyebrows. On BenchCAD, which asks a model to rebuild 3D objects from multi view renders through CAD code, K3.1 Pro scored 92.7 percent with tools, against 96.2 percent for Opus 5.5 and 95.9 percent for GPT-6 Astra. On OfficeQA Pro v2, a financial document reasoning test, it posted 64.4.

Moonshot also published a long horizon demo that is more interesting than the headline scores. It set the model loose on a human walking dynamics simulator built on human-atlas, an open anatomical model with 296 skeletal components and 402 muscles. Across 167 missions and 420 independent reviews, the model diagnosed failures, found 36 errors in shoulder muscle moment arm calculations and retrained the controller. The simulated human went from collapsing within seconds to walking for 11.2 seconds.

Context matters here. Moonshot ran its own Terminal Bench evaluation with Claude Code, and it took rival scores from either the model maker’s blog or the official leaderboard, whichever was higher. That is a defensible method, and it is also one that can flatter a new model. A 0.7 point gap is inside the range where harness settings, retry policy and reasoning effort start to matter. The same pattern showed up in OpenAI’s COBOL migration claim earlier this month.

Pricing: real, and cheaper than the frontier

Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million. Moonshot frames that as 50 percent below Claude Opus 5.5 at $4 and $20, and 80 percent below GPT-6 Astra at $10 and $50. Independent listings on OpenRouter show the same $2 and $10.

The independent picture is not only about price. Artificial Analysis placed K3.1 Pro fourth on its intelligence index with a score of 55, up from 44 for Kimi K3 at maximum effort. The three models above it were Claude Opus 5.5, Claude Sonnet 5.5 and Meta’s Muse Wonder Full, all in their maximum effort configurations. On the code arena’s WebDev board, K3.1 Pro opened at number eight with 1,692 points, which the arena team described as matching Claude Opus 5 at roughly 70 percent lower cost.

For teams running long agent loops, the input price is the number that compounds. A coding agent that reads a large repository on every turn spends far more on input than output, so a 50 percent cut there is worth more than a bigger discount on a small output bill.

What is confirmed, and what is not

The safety section deserves its own note. Moonshot says vulnerability research performance climbed steadily as post training scaled, and that cyber content was sitting inside the general mid training corpus and the reinforcement learning mix. For the open weight release, it removed that data from both stages. In its own xdev bench hard test, a checkpoint without those controls solved 90.5 percent of exploitation tasks, and the final model solved 81.0 percent. Moonshot also says the final cyber capability stays behind frontier proprietary models that have been public for at least three months.

That is a candid disclosure, and it is also a reminder that the capability existed before the controls. The open weights, when they arrive, will decide whether the final number holds.

A comparison worth keeping

What it means for Indian developers

For Indian teams building coding agents, the hosted price is the immediate win. Two dollars per million input tokens halves the cost of feeding a repository into Claude Opus 5.5, and the $0.20 cache read price rewards the repeated context that agent loops generate. Startups billing Indian clients in rupees while paying for inference in dollars will notice the difference on every invoice.

The bigger shift lands on November 4, when the weights are meant to be downloadable. Self hosting keeps code and customer data inside a company’s own environment, which matters for regulated clients in banking and healthcare. It also means hardware. A 2.8 trillion parameter mixture of experts still activates 104 billion parameters per token, so this is not a laptop model, and an eight GPU node is optimistic.

The practical advice is boring and correct. Test K3.1 Pro on your own tasks before moving a production workload. Vendor benchmark tables, including the ones covered here, describe a model’s ceiling on curated tests, not its behaviour on your messy monorepo. The same caution applies to the DeepSeek inference stack and to the llm-d cost claims covered earlier on this site.

There is also a timing question. A model that matches the frontier in October may trail it by December, when OpenAI and Anthropic ship again. Moonshot’s own release cadence, a major upgrade roughly every three months this year, suggests K3.1 Pro will be superseded quickly. Build for portability, not loyalty.

Frequently asked questions

When will the open weights for Kimi K3.1 Pro be released?

Moonshot says the full model weights will be released by November 4, 2026. Until then the model is available only through the hosted Kimi API, Kimi.ai, the Kimi apps, Kimi Work and Kimi Code.

How close is Kimi K3.1 Pro to Claude Opus 5.5?

On Terminal Bench 4.0 it scored 65.7 percent against 66.4 percent for Opus 5.5, a gap of 0.7 points. Artificial Analysis ranks it fourth overall on its intelligence index, behind Opus 5.5, Sonnet 5.5 and Muse Wonder Full.

Is the $2 per million token price real?

Yes. Moonshot’s Kimi API charges $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million. OpenRouter lists the same rates.

Did Moonshot fix the hallucination problem?

On its internal Kimi Hallu Bench Hard, the claim level hallucination rate fell to 8.3 percent from 23.4 percent for Kimi K3. The test is a stress set built to provoke errors, so the absolute number should not be read as a general error rate.

What did Moonshot change for safety before release?

It removed cyber related data from mid training and filtered the reinforcement learning environment mix. In its own exploit benchmark, the model dropped from 90.5 percent without those controls to 81.0 percent with them. Moonshot says the final capability still trails frontier proprietary models.

For now, K3.1 Pro is the strongest argument yet that open weight models have caught up to within a rounding error of the closed frontier on coding. The test that matters arrives on November 4, when anyone can download the weights and run the benchmarks themselves.