Kimi K3.1 Pro: Moonshot Cuts the Price by a Third and Lands a Point Behind Claude Opus 5.5
Moonshot shipped Kimi K3.1 Pro with a third off the per-token price and 65.7% on Terminal-Bench 4.0, seven tenths behind Claude Opus 5.5. The useful question, though, is not what a token costs but what a finished task costs.
What Kimi K3.1 Pro Is and What Changed From K3
On October 23, 2026 Moonshot AI introduced Kimi K3.1 Pro and its Highspeed variant. This is not a model built from scratch: it is a post-trained version of Kimi K3, with a much larger joint reinforcement learning effort across domains. The company published fourteen evaluations alongside the announcement, spread across coding, science and knowledge work.
A Post-Trained K3, Not a Model Built From Scratch
The architecture stays the same: mixture of experts on Kimi Delta Attention and Attention Residuals, with native vision for image input. What changes is the post-training, which explains why the numbers move up without the model structure changing. If you already ran K3, migrating is a configuration change, not a new integration.
2.8 Trillion Total Parameters, 104 Billion Active, 1M Context
The model claims 2.8 trillion total parameters with 104 billion active per token, the figure that actually matters for cost and latency. Context reaches one million tokens, the same as K3. Worth pinning down: some coverage rounds it to three trillion, and that is not accurate.
Read also
Two Versions: Pro and Highspeed
Moonshot offers K3.1 Pro and K3.1 Pro Highspeed, which the company says is two to three times faster. That is a vendor figure with no independent measurement published, so treat it as a hint rather than a latency guarantee.
The Price: A Third Less Per Token
$2 and $10 per Million, With Cache Reads at $0.20
On Moonshot's API, K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20. Kimi K3 launched in July at $3 and $15, with cache at $0.30. The cut is exactly one third across all three rates, and OpenRouter lists the same figures. Third-party prices vary, so any comparison should say where each number comes from.
Where You Can Use It
The model is live on Kimi.ai, Kimi Work, Kimi Code and the API from launch day. It is not an open model yet: today you consume it as a product or a service, not as a download.
Moonshot's Benchmarks and How to Read Them
Terminal-Bench 4.0: 65.7 Against GPT-6 Astra's 58.2 and Opus 5.5's 66.4
On the terminal-agent test Moonshot publishes, K3.1 Pro claims 65.7%. That puts it ahead of GPT-6 Astra at 58.2% and seven tenths behind Claude Opus 5.5 at 66.4%. These are vendor numbers measured in the vendor's own harness, and should be read as such: it neither beats Opus 5.5 nor sits far from it.
BrowseComp, HLE With Tools and SWE-Marathon
On the other three headline tests, the company reports its best-in-row value: 92.1 on BrowseComp, 70.0 on Humanity's Last Exam with tools and 51.7 on SWE-Marathon v1.1. They measure web search, reasoning with tool access and long engineering work, three different things that a single score cannot summarise.
What Independent Tests Say: 55 Points and Fourth Place
Artificial Analysis gives it 55 points on its Intelligence Index, against a median of 27 for comparable models, placing it fourth. That is an independent measurement and considerably more sober than the vendor table, which makes it useful for putting the numbers in context.
The Detail Almost Nobody Checks: Cost Per Task
More Verbose: 55,000 Output Tokens Per Task
The same independent analysis describes the model as notably slow and very verbose. It generated 150 million tokens across the index evaluation, against a median of 87 million, and burns around 55,000 output tokens per task: 14% more than K3. That is where the rest of the story lives.
When It Pays Off and When It Does Not
With those figures, cost per finished task lands near $1.98, practically the same as K3, even though the per-token price is 33% lower. Because the output rate dominates in agents that reason and call tools, the saving shrinks on its own. It pays off for long-reading or huge-context work, where cheap input matters; it does not pay off for live autocomplete, where latency shows up more than a benchmark.
Open Weights on November 4
What Moonshot Promised and What Is Still Unconfirmed
Moonshot promised to publish the full weights before November 4, 2026. Until then it is not an open-source model: it is an API and product model. The licence is unconfirmed too. K3 shipped under the Kimi K3 License, described as a modified MIT variant, but nothing has been announced for K3.1 Pro.
What It Means for Self-Hosting
Even when the weights land, size decides everything: K3 already demanded serious infrastructure, and 2.8 trillion parameters do not run on a workstation. What to watch on November 4 is the actual licence, the download size and the minimum hardware. If self-hosting is the plan, those three facts make the decision, not the announcement.
Conclusion
Kimi K3.1 Pro is good engineering news and a reminder that list price is not your invoice. Before switching models, measure your own output tokens per task, not the leaderboard's: if your agent writes a lot, a third off the rate can end up as almost nothing. If you follow the open-model market, the Mistral Large 4 story follows the same launch-by-API pattern with weights later, and the GLM-5.3-Flash release shows the other Chinese contender from last week. To see where a model like this actually gets used, there is the AI code editor comparison. And on this blog we keep measuring with data, not headlines.
