Moonshot AI published the weights of Kimi K3.1 Pro on November 4, so anyone can now download the model, run it on their own machines and build on it. It is an update to the company's July model, Kimi K3, and it costs a third less: $2 per million tokens read and $10 per million tokens written, against $3 and $15 before. That is half of what Claude Opus 5.5 costs and a fifth of what GPT-6 Astra costs.
What half a price buys depends on the kind of work.
It is built for long routine jobs, not hard puzzles
In the results Moonshot published, K3.1 Pro trails Opus 5.5 by under a point at working through a computer's command line, and it leads at office automation and at answering questions over long financial documents. It is far behind at writing whole programs from a specification, where it scores 31.5 against 65.0 for Opus, and at hard science problems that must be solved in one go.
That split is the useful part. A model that drives tools through a long job does well, and a model that has to reason alone through one hard problem does not. Most of what companies buy agents for is the first kind. For the repetitive work that fills a back office, the cheaper model is probably good enough, and the expensive one earns its price on the rare hard problem.
What half the price looks like on one job
For example, a finance team at a hospital group might have an agent that reads a month's supplier invoices and matches each one to a purchase order. Say one run reads 2 million tokens and writes 200,000. If each model uses the same number of tokens, that run costs $6 on Kimi K3.1 Pro, $12 on Opus 5.5 and $30 on GPT-6 Astra. At 100 runs a month, that is $600, $1,200 or $3,000.
For one team, that is a pleasant saving and not a decisive one. It becomes decisive for software vendors, who charge per seat for AI features and are pricing against the dearest models. Once their customers know a near-equal costs half as much, the vendors will pass some of the saving on or add more work for the same fee. I expect the AI line on your software invoices to shrink or to buy more within a year.
Why most businesses will not download it, and why it matters anyway
The files take up 1.56 TB, and running the model takes a server with eight of Nvidia's newest data-centre chips. Each of those servers costs roughly $550,000 to $650,000, and a company that depends on the model needs two. A hospital group will not casually run that in its own back office.
What the weights change is that no single company controls the model any more. Moonshot can charge what it likes, but it cannot take the model back, and anyone with enough hardware can offer it. A business with strict rules about where patient or customer data may go can run it on its own machines, and the licence lets it do that for free. Only a company that earns more than $20 million a year and resells access to the model itself needs a deal with Moonshot.
So the cheap option is real, but the benchmark table only tells you where to look. Pick a job you already run every week, send it through both models, and compare what each costs per correct result. If the cheaper one gets it right, you have found the half of that bill you can stop paying.