JOURNAL / 2026.10.25

Kimi K3.1 Pro improves agents and cuts token prices, with weights still pending

Moonshot offers an evolution of K3 with competitive results and lower token prices. Integration, reasoning usage and the future weights release will determine how much it changes real work.

Moonshot has launched Kimi K3.1 Pro, a version of K3 with more post-training, available through its products and API. It promises full model weights by November 4. For people building agents, there is a new service to evaluate today; the ability to retain and operate it independently remains pending.

The original K3 announcement already described native vision and a million-token context window. K3.1 does not introduce those capabilities. The relevant question is how much better it becomes at finishing work with them: navigating a project, running tools and correcting errors until it delivers something usable.

An improvement worth testing

There is evidence beyond the launch presentation. Artificial Analysis places K3.1 Pro at 55 points and fourth on its Intelligence Index when checked today. The index combines ten evaluations, spanning tool-based work, knowledge, science and long context. It signals broad competitiveness; its aggregate score cannot establish reliability in a particular repository or a success probability for any assignment.

Moonshot reports 65.7% on Terminal-Bench 4.0, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. These are terminal tasks with specific tests, not a measure of all programming. Its notes combine in-house results with external reports and leaderboards; this is not one controlled comparison. Nor does its entire table favor Kimi. I find that sufficient to include it in a serious evaluation, insufficient to declare a universal winner.

A less conspicuous part of the launch could determine that evaluation's outcome. The Kimi Code documentation requires reasoning to remain enabled and recommends starting a new session when switching to the model. It also warns of variations from protocol translation. Simply reusing an old conversation could end up measuring a faulty integration.

In that product, Pro is available from Plus/Moderato, while the full context and Highspeed require Pro/Allegretto or above. The model name and maximum window alone do not describe what each account includes.

Less per token, less per result?

The official API platform charges $2 per million input tokens and $10 per million output tokens for Pro, against $3 and $15 for K3. Highspeed costs $4 and $20; cache reads have a separate rate. Moonshot advertises faster serving for that variant, but we have not measured it. Choosing it depends on the value of the waiting time it can save.

There is a useful arithmetic consequence: at base rates, the same mix of input and output costs one-third less than with K3. If both volumes grew by 50%, the bill would return to its starting point. This conceptual example excludes caching, tools and taxes; it does not describe the model's observed usage.

That is why I would compare the cost of accepted changes, including failed attempts and review, before the price of a response. A more capable agent could save money even while writing more; another could consume the savings in reasoning and corrections. The rate card supports scenario calculations. Choosing requires measuring the entire workflow.

My reading is that K3.1 Pro expands the alternatives for demanding work and makes competition on outcomes more interesting. I would keep two decisions on separate dates: evaluate the service now, then examine the license, artifacts and requirements when the promised weights arrive. A good API trial helps establish whether the capability matters; it does not yet demonstrate that operating it independently is viable.

Sources consulted on October 25, 2026. We reviewed documentation and published results; we have not run the model or reproduced its evaluations.

← Back to journal