Kimi K3.1 Pro: Moonshot overtakes GPT-6 Astra on Terminal-Bench 4.0 and opens up its 3 trillion parameters on November 4

Kimi K3.1 Pro: Moonshot overtakes GPT-6 Astra on Terminal-Bench 4.0 and opens up its 3 trillion parameters on November 4

🔎 An open-weight model corners the American camp on its own turf

On October 23, 2026, Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 series. A 3T-class open-weight model with a one-million-token context window, available now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. On paper, just another launch in a saturated market. On the benchmarks, something else entirely.

According to Unite.ai, K3.1 Pro reaches 65.7% on Terminal-Bench 4.0. GPT-6 Astra, OpenAI's flagship, stops at 58.2%. Only Claude Opus 5.5 does better, at 66.4%. In other words: an open-weight model, just 0.7 points behind the best closed model on the market on the industry's most closely watched agent benchmark.

And the rest is coming fast. The full weights are promised for November 4, 2026 at the latest, after the removal of offensive cybersecurity data from both training stages. According to Artificial Analysis, the gap between open weights and proprietary models has now narrowed from 12 to 3 points. For developers who self-host their models, the balance of power has just shifted.

The essentials

• Launched on October 23, 2026: Kimi K3.1 Pro, the first model in the K3.1 series, a 3T-class open-weight model, available on Kimi.ai, Kimi Work, Kimi Code, and the API.
• Architecture: 2.8 trillion total parameters, 104 billion active (MoE), multimodal, 1M-token context.
• Terminal-Bench 4.0: 65.7%, ahead of GPT-6 Astra (58.2%), behind Claude Opus 5.5 (66.4%).
• Intelligence Index: 4th place with a score of 55, up +11 points over Kimi K3 (44). 3rd on GDPval-AA with 1,846 Elo (+312).
• Hallucinations: 8.3% at the claim level on Kimi Hallu Bench Hard, versus 23.4% for K3.
• Long-horizon coding: 61.2% on Kimi Code Bench 3.0 Long-horizon, versus 55.95% for K3.
• API pricing: $2.00/M input tokens, $10.00/M output, $0.20 per cache hit (October 2026, verify on kimi.ai). Highspeed variant 2-3× faster.
• Open weights: announced for November 4, 2026, not yet released at the time of this article.

Recommended tools

K3.1 Pro is consumed in four different ways, and the right choice depends on your use case. Here's the map.

The numbers: 4th on the Intelligence Index, 0.7 points behind Claude Opus 5.5 on agents

Yes, K3.1 Pro enters directly into the very exclusive club of frontier models. Artificial Analysis ranks it 4th on its Intelligence Index with a score of 55, i.e. 11 points better than Kimi K3 (44 at most). On GDPval-AA, the model takes 3rd place with 1,846 Elo, a jump of 312 points.

The summary table, with figures from Kimi's official tech blog and third-party sources (October 2026):

One detail deserves to be highlighted, because it lends credibility to the whole thing: on BenchCAD with tools, Moonshot reports a score lower than the two compared models (92.7% versus 96.2% and 95.9%, according to Unite.ai). Publishing its own partial defeats is the sign of a press release that isn't really one.

My take: the Terminal-Bench 4.0 score is the real news. The rest confirms steady progress in the K series, but overtaking GPT-6 Astra on terminal agentic tasks is what the American labs will look at first.

Terminal-Bench 4.0: the benchmark that now obsesses the labs

Because it measures exactly what customers are paying for today: an agent capable of chaining together dozens of steps in a terminal, without human supervision, all the way to the result.

Terminal-Bench evaluates real command-line tasks: navigating codebases, running tools, debugging, deployment. It's the closest proxy to the daily work of a senior developer handed off to a machine. A score of 65.7% on this type of test says something that a multiple-choice score never will.

Important nuance: OpenAI mostly communicates about Terminal-Bench-Science, where GPT-6 Astra posts 64.6% — a different benchmark, more focused on hard sciences. On the standard Terminal-Bench 4.0, Unite.ai measures Astra at 58.2%. We covered this OpenAI communication in detail in our article on GPT-6 Astra and its reasoning that has become harder to monitor.

Meanwhile, OpenAI is playing another card: parallelism, with GPT-6 Astra Pro putting several agents to work in parallel on a single task. The battle has clearly shifted from "the model that answers" to "the agent that executes." To see clearly in this race, our comparison of the best LLMs for AI agents is updated continuously.

Hallucinations: from 23.4% to 8.3%, but read the footnote

The drop is real and spectacular — but Kimi itself points out that these rates don't reflect errors in everyday use.

The numbers, according to Kimi's official blog: claim-level hallucination drops from 23.4% (K3) to 8.3% on Kimi Hallu Bench Hard. The share of responses containing at least one factual error is reduced to 42.4% of K3's rate. Concretely, the model makes mistakes more than twice as rarely on factual claims.

On the coding side, the progress is more measured but real: 61.2% on Kimi Code Bench 3.0 Long-horizon, versus 55.95% for K3. "Long-horizon" is the key phrase: this is about sustaining a development goal over long sequences, not completing a single function. That's precisely the profile of the best LLMs for coding in 2026.

My take: these numbers are credible because they're measured on "hard" benchmarks and come with a public admission of weakness on BenchCAD. But be wary of any marketing transposition: a lab hallucination rate tells you nothing about the error rate on your business domain. Test on your own data before putting an agent into production.

Under the hood: 2.8 trillion parameters, 104 billion active, 1M context

K3.1 Pro is not a dense model: it's a 3T-class MoE (mixture of experts), with 2.8 trillion total parameters and only 104 billion active per token, according to Kimi's official account on X. The model is multimodal and accepts a one-million-token context.

Two technical choices deserve a closer look.

First, MoE at very large scale. It has become the industry's dominant trade-off: knowledge storage capacity close to that of a giant dense model, with inference cost reduced to the size of the active parameters. Moonshot is applying it to a size class — 3T — that no one had yet opened up with public weights.

Second, joint RL. Moonshot claims a reinforcement learning training conducted jointly across hundreds of domains, rather than domain by domain. The idea: force capabilities to transfer between one another instead of optimizing them in silos. It's a defensible research hypothesis, and the measured gains on both Kimi Code Bench and GDPval-AA give it weight.

Finally, the K3.1 Pro Highspeed variant promises 2 to 3 times higher throughput, with the API ids kimi-k3.1-pro and kimi-k3.1-pro-highspeed. For agents that chain dozens of tool calls, this option is not a luxury: cumulative latency is often what kills an agentic loop.

On November 4, open weight goes to scale — and self-hosted devs stand to benefit

For the first time, an open-weights model will play in the frontier league, with a 3T scale and scores that come within reach of the best closed models. This is no longer "cheap" open weight — this is frontier open weight.

According to Artificial Analysis, the gap between open weights and proprietary models has gone from 12 to 3 points. At that level, the question is no longer "is open weight good enough?" but "do I really need a closed API?". For a business: on-prem hosting, data that never leaves, fine-tuning on your own domain, zero dependence on a third party's pricing policy. For an indie dev: the ability to build on a frontier base without a billing cycle on every token.

Let's stay clear-headed about feasibility, though. 104 billion active parameters won't run on a laptop, no matter what the enthusiastic threads claim. You'll need serious GPU infrastructure, a dedicated server or a suitably sized VPS — Hostinger remains a reasonable entry point for experimenting (prices vary, check hostinger.com). If you're new to self-hosting, our guide on installing an LLM locally with Ollama or LM Studio remains the starting point, and our selection of the best LLMs to run locally will help you calibrate your expectations based on your hardware.

One point of uncertainty: November 4 is a promised date, not a confirmed one. Ontimebrief confirmed at the end of October that the weights had not yet been published. Wait to see the usage license and any potential restrictions before planning a migration.

The "hacking" cut: a security we must take on Moonshot's word

Moonshot removed offensive cybersecurity data from both training stages before the open weight release — but the evidence for this reduction rests on an unpublished internal test.

The facts, reconstructed from Kimi's blog and an implicator.ai analysis dated October 23, 2026: on xdev-bench-hard, the checkpoint without reduction measures — no cyber data removal, no RL environment filtering — solves 90.5% of the problems. The released version drops to 81.0%. The test backing this claim is a Moonshot internal benchmark that has not been made public.

On substance, the decision is defensible. Releasing an open weight model capable of solving 90% of a hacking benchmark would be an irresponsible move, and any serious lab would do the same. On form, it's more awkward: we have to take Moonshot at its word. A private benchmark, unauditable numbers, a 9.5-point delta that no third party can verify.

This is the exact mirror image of the closed-model problem — "trust us, our model is safe" — but this time on the open weight side, which specifically claimed to stand for verifiability. If the community cannot audit xdev-bench-hard, the line between real security and compliance marketing will stay blurry. My bet: this test will eventually leak or be reproduced, and it will be a significant moment for the credibility of the entire open weight ecosystem.

Pricing and access: $2 input, $10 output, a third less than K3

The API is already available, and cheaper than the previous generation — this is the second blow dealt to American pricing grids.

According to CellCog, K3.1 Pro costs $2.00 per million input tokens, $10.00 output, and $0.20 for a cache hit (October 2026, check on kimi.ai), i.e., a third less than K3. The model's reference page is also documented on LLM Gateway. The Highspeed variant is billed on the same grid.

For context: $2/$10 for a model that scores 65.7% on Terminal-Bench 4.0 is a price/performance ratio that closed frontier offerings simply don't offer today. For an agent consuming millions of tokens per day, the gap runs into thousands of euros per month. If your budget is zero, our roundup of the best free LLMs remains the entry point, but be aware that the frontier price floor has just dropped another notch.

My advice: start with the standard model to assess quality, then switch to Highspeed only if your agentic loops suffer from latency. No point paying for higher throughput on an overnight batch.

The underlying trend: the open-weight frontier is now Chinese

This launch is not an isolated accident. It is the continuation of a strategy already visible a few weeks ago.

Moonshot raised 2 billion dollars in October, at a time when Kimi K2.6 was already dominating the open-weight space and China was accelerating in the AI race. K3.1 Pro is the materialization of that capital: a 3T-class model, a 1M context window, frontier scores, and a weights release date held to within twelve days of the API launch. According to Ontimebrief, the gap with the best closed models is closing — that is exactly what Artificial Analysis measures.

And the American front is not only being attacked by Chinese labs. Alternative architectures are also entering the picture: Mercury 3, Inception's diffusion LLM, claims to beat GPT-6 Luna and Claude. When the frontier is being contested both by Chinese open weights and by new inference paradigms, the de facto monopoly of a few San Francisco Bay Area labs no longer makes much sense.

Practical consequence for you: our monthly comparison of the best LLMs must now be read with a systematic eye on the open-weight column. Six months ago, it was a curiosity. By the end of 2026, it is a full-fledged purchase option.

❌ Common Mistakes

Mistake 1: Taking the 8.3% hallucination rate as a guarantee

The problem: these figures are measured on Kimi Hallu Bench Hard, under test conditions. Kimi itself points out that they don't reflect errors in everyday use. The fix: treat this number as a trend signal (K3.1 Pro hallucinates far less than K3), then evaluate the actual error rate on your own use cases before any deployment.

Mistake 2: Believing you'll be running K3.1 Pro on your own machine on November 4

The problem: 104 billion active parameters require GPU infrastructure that few individuals possess. November 4 opens access to the weights, not the ability to run them anywhere. The fix: plan for a GPU VPS or dedicated cloud, or wait for community quantizations — keeping in mind that they always degrade the advertised scores somewhat.

Mistake 3: Comparing scores without checking the test conditions

The problem: Terminal-Bench-Science is not Terminal-Bench 4.0, a benchmark with tools is not a benchmark without tools, and the filtered checkpoint (81.0%) is not the raw checkpoint (90.5%) on xdev-bench-hard. The fix: before citing a number, check three things — the exact benchmark, whether tools were used, and the version of the model tested.

❓ Frequently Asked Questions

When will the open weights of Kimi K3.1 Pro be available?

By November 4, 2026 at the latest, according to Moonshot and CellCog. As of the date of this article (late October 2026), they have not yet been released. Moonshot removed the cyber offensive data from both training stages before release. Check the license on release day: it will determine what you can do in production.

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

$2.00 per million input tokens, $10.00 for output, and $0.20 for a cache hit (October 2026, check on kimi.ai), i.e., one third less than K3. Two ids are available: kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with the fast variant offering 2 to 3 times higher throughput at the same price point.

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

On Terminal-Bench 4.0, yes: 65.7% versus 58.2% according to Unite.ai. On Terminal-Bench-Science, OpenAI claims 64.6% for Astra — a different benchmark. Claude Opus 5.5 remains ahead on Terminal-Bench 4.0 with 66.4%. One benchmark does not make a verdict: cross-reference multiple sources before choosing.

What is the architecture of Kimi K3.1 Pro?

A class-3T MoE: 2,800 billion total parameters, 104 billion active per token, multimodal, with a one-million-token context. Moonshot claims joint RL across hundreds of domains. The Highspeed variant offers 2-3× higher throughput for latency-sensitive use cases.

Why did Moonshot reduce the model's hacking capabilities?

To make the open weight release responsible: offensive cybersecurity data was removed from both training stages. Verification relies on xdev-bench-hard, an unpublished internal test: 81.0% for the shipped version versus 90.5% for an unfiltered build. The decision is defensible; its verifiability remains contested.

Can you self-host Kimi K3.1 Pro at home?

After November 4, if the weights are released as promised — but with 104 billion active parameters, you need serious GPU infrastructure, far from a laptop. A realistic option: a VPS or a rented GPU server. Start with our guide to setting up a local LLM, then scale up according to your actual needs.

✅ Conclusion

With 65.7% on Terminal-Bench 4.0, hallucinations cut by nearly a third (divided by almost three), and frontier weights promised for November 4, 2026, Kimi K3.1 Pro moves open weight from the status of economical alternative to that of a frontier contender — and the balance of power between closed labs and self-hosting devs has just shifted. To find out where this model stands in the October 2026 hierarchy, check out our monthly comparison of the best LLMs.