Kimi K3.1 Pro launches: access, API pricing and when Highspeed is worth testing

What launched, and what is still to come?

Moonshot AI has introduced Kimi K3.1 Pro and K3.1 Pro Highspeed. Kimi Code’s release notes date the rollout to October 23, 2026. The company’s technical blog describes Pro as a post-trained successor to Kimi K3, targeting coding, 3D creation, financial research and longer agent workflows.

Both versions have a stated one-million-token context window. Moonshot says Highspeed delivers roughly two to three times the speed of Pro. That is a vendor claim, rather than a Promptwire measurement.

Hosted availability and downloadable weights are separate milestones. Moonshot lists access through Kimi, Kimi Work, Kimi Code and its API now, while promising full model weights by November 4, 2026. A planned weights release should not be treated as a download already available today. Enterprise administrators must enable access; the launch blog says it is off by default.

Public record: Moonshot AI — Kimi K3.1 Pro technical launch blog · Kimi Code — October 23 release notes and access requirements

The API price makes the speed decision concrete

Kimi’s API platform currently lists Pro at $2 per million input tokens and $10 per million output tokens. Highspeed is $4 for input and $20 for output. These are the provider’s published US-dollar rates checked October 24, rather than a quote for a subscription plan.

For a simple uncached example, 100,000 input tokens plus 10,000 output tokens would cost $0.30 on Pro or $0.60 on Highspeed at those rates. This illustration excludes tools, retries, additional turns and other charges. A long agent run can make multiple requests, so one prompt is not the same as one finished job.

Public record: Kimi API Platform — current model pricing

Kimi Code access depends on the plan

The official Kimi Code notes list Pro for Plus / Moderato subscribers and above. Pro / Allegretto and higher tiers unlock up to one million tokens of context and access to Highspeed. The maximum context figure therefore does not describe every subscription tier.

The same notes say thinking mode is required, with low, high and max effort levels and max as the default. Disabling thinking causes an error. Kimi recommends starting a new session when moving from an earlier model, rather than changing models inside an existing conversation.

Public record: Kimi Code — October 23 release notes and access requirements

There is also a gateway route

Vercel’s October 23 changelog confirms both models are available through AI Gateway. It lists moonshotai/kimi-k3.1-pro and moonshotai/kimi-k3.1-pro-highspeed as the model names, with AI SDK and supported API routes. That gives existing gateway users another access path; it does not remove the need to check their own billing and configuration.

Public record: Vercel — October 23 AI Gateway availability

Test the cost of an accepted change

Promptwire’s suggested first test is a small bug fix in a repository you know. Write the expected behaviour and an acceptance check before starting. Give each model the same files, instructions, tools and reasoning setting, in a fresh session.

Record the full elapsed time, API usage, number of attempts and minutes you spend reviewing the patch. Check that the fix passes and that unrelated behaviour remains intact. A quicker stream of code is only one part of the result.

For interactive work, paying more may be reasonable if it reduces waiting enough to matter. For a background job, a slower model may fit the budget better. Those are evaluation hypotheses: compare repeated accepted outcomes before changing a production workflow.

A large context window also deserves its own test. Ask for a change that depends on two distant parts of the codebase, and inspect whether both were understood. Capacity tells you how much material can fit; your test reveals whether the model used the relevant material correctly.

Promptwire has not independently benchmarked these models. The useful starting point is a controlled task with a known result, rather than assuming a launch score predicts your review burden.