Kimi launches K3.1 Pro with API and coding access
Kimi has launched K3.1 Pro, the first model in its new K3.1 series. Its October 23 release notes confirm that the model is fully rolled out in Kimi Code, alongside a Highspeed variant. Public API documentation also lists both versions.
The official API guide describes K3.1 Pro as a version of K3 trained further using substantially greater posttraining compute. Hosted access comes before downloadable weights, which Kimi says it will release by November 4. That promise is not a verified weights release or license.
This is a confirmed product release, unlike the configuration clues and teaser reports in our earlier Kimi K3.1 coverage. That article remains a record of what was known before official availability was documented.
Two API models with different speed claims
The current model catalog lists kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Both offer native visual understanding and up to one million tokens of context. Kimi describes the underlying model as 3T class, not a newly specified exact parameter count. It says Highspeed delivers the same intelligence at roughly two to three times the output speed.
The API overview lists Chat Completions, Responses and Anthropic compatible Messages interfaces. A model does not independently gain internet or database access merely by choosing one of these interfaces. External resources require official tools or tools supplied by the developer.
API access requires a successful account payment of at least $1. Account funding also determines rate limits. ByteForward has not made inference calls or independently measured either model’s performance.
The API billing documentation distinguishes usage charges from Kimi membership and Kimi Code subscriptions. Input and output are billed separately, and caching adds distinct write and read charges. Reusing a long prompt is therefore a billing question as well as a context question. Check the current model rates and cache lifetime before estimating a workload’s cost.
Kimi Code has separate access rules
The Code configuration guide uses different IDs, k3.1-pro and k3.1-pro-highspeed. The standard K3.1 Pro model requires Plus or the legacy Moderato plan, or above. The full context window requires Pro or legacy Allegretto and above. Highspeed also requires those higher tiers and consumes approximately twice the regular model’s quota.
Both variants require thinking mode. The supported effort levels are low, high and max, with max as the default. Turning thinking off returns an error rather than creating a faster mode of the same model. Kimi recommends its native protocols rather than translation through proxy tools.
Start a new session instead of switching an existing K3 or K2 conversation to the new model. Kimi warns that carried reasoning can affect behavior and that changes invalidate context caches. Check the exact ID too. The Code guide says a mistyped identifier falls back to K2.8 Preview without thinking.
Read accuracy claims as targeted tests
Kimi reports that the rate of hallucinated claims fell from 23.35 percent for K3 to 6.85 percent for K3.1 Pro on a selected factual accuracy stress test. The examples were deidentified conversations in which users had flagged earlier errors. Kimi explicitly says these figures do not represent everyday error rates. They are vendor results on deliberately difficult examples, not a general reliability guarantee.
The API guide also contains unfinished placeholders in some benchmark descriptions and scores. Those entries cannot establish a ranking. We are not filling them with inferred values or treating selected demonstrations as independent evaluations.
What to check before moving a workflow
Separate the model from the surrounding application. Tool access, permissions and the work performed by external services remain part of the workflow. A faster model response does not establish that an entire coding job, browser session or research task finishes proportionally sooner.
For a useful comparison, use fresh sessions and the same tasks, tools and thinking settings. Check returned results and quota use, not just the displayed model name. The access and migration rules are concrete enough to act on now. Broad performance claims still need testing on the work you intend to run.
Official Kimi wordmark panel from Moonshot AI, used unmodified under its editorial logo terms. This is general brand identification, not Pro specific launch artwork or an endorsement.
