Kimi K3.1 Pro deserves a place on an enterprise AI shortlist. It combines lower published token rates than Kimi K3 with stronger independent launch results. That is a reason to test it, not proof that it will lower your cost per accepted piece of work.
We have added the release to Aeon's Model Economics Lab, including the default master plot, evidence notes and downloadable dataset. Sources were checked on October 23, 2026. This is a source-based analysis, not an Aeon-run benchmark.
Available through the API, with weights still to come
Moonshot's announcement makes K3.1 Pro and its Highspeed variant available across its app, Work, Code and API products. The model has a 1,048,576-token context window. Moonshot plans to release full weights by November 4. API availability today does not mean downloadable weights are available today.
The announcement reports a 65.7% Terminal-Bench 4.0 result using a Claude Code harness and describes stronger coding, finance and 3D workflows. These are vendor-reported results. They are not Aeon's measurements or a guarantee for a customer's repositories, documents or controls. Moonshot also describes training-time measures to reduce cyber misuse; that does not make an agent safe to run with unrestricted tools.
For buyers considering private deployment, wait for the actual weights, licence and serving requirements. Then price the infrastructure, operations and governance needed for the intended workload. Do not book a self-hosting saving against an unreleased deployment option.
What the published API rates mean
The Kimi API pricing documentation lists the following USD rates per one million tokens, excluding applicable taxes:
| Billing category | K3.1 Pro | K3.1 Pro Highspeed |
|---|---|---|
| Input, without a cache hit or write | $2.00 | $4.00 |
| Output | $10.00 | $20.00 |
| Cached input read | $0.20 | $0.20 |
| Cache write, 5-minute TTL | $2.00 | $4.00 |
| Cache write, 1-hour TTL | $4.00 | $8.00 |
The platform's current model catalogue lists Kimi K3 at $3 input and $15 output per million tokens. K3.1 Pro is one-third lower on those two rates. That comparison holds the number and category of tokens constant. It does not hold an agent's retries, reasoning length or review effort constant.
For an illustrative workload with one million ordinary input tokens and 100,000 output tokens, Pro's input/output bill would be $3 and Highspeed's $6. This is rate arithmetic, not a measured agent task or a quote. It excludes cache writes, tools, retries, infrastructure and tax.
Cache reads, cache writes and ordinary input are separate billing categories. Do not charge the same input tokens twice when estimating a cache write. The default cache lifetime is five minutes; longer retention has a different write rate. A useful cost model needs the actual mix, not an assumption that every repeated token gets the cache-hit price.
Moonshot describes Highspeed as two to three times faster. Treat that as a provider claim, not a guaranteed reduction in end-to-end task time. Tools, tests, network waits and human review may dominate the workflow. API usage is also separate from Kimi app or Code subscription allowances.
Independent results: keep the benchmark and its bill together
Artificial Analysis's launch analysis reports an Intelligence Index score of 55, versus 44 for Kimi K3 at maximum effort. Its launch snapshot reports about $1.98 per Intelligence Index task and 55,000 weighted output tokens per task, compared with 48,000 for K3. The evaluator describes task cost as roughly unchanged despite lower token rates.
That is the central economics lesson: more capable, cheaper tokens can still produce a similar task bill when the system uses more of them. The launch results support a serious evaluation, not an automatic production routing change.
The Intelligence Index is not the Coding Agent Index. The live model page already reports a different weighted Intelligence Index task-cost snapshot, $2.11. We retain the dated launch figures together rather than splice the live cost into the launch token count. Neither snapshot is a cost per successful business outcome.
Where K3.1 Pro appears in the master plot
At this check, the Coding Agent Index had no complete K3.1 Pro coding-agent row with matched score, task cost and token use. Its current v2.0 suite also differs from the Lab's retained historical cohorts. We will not turn an Intelligence Index score of 55 into a coding score.
The Lab's default includes every new release plus the maintained shortlist. K3.1 Pro therefore appears as a clearly labelled K3 predecessor-proxy diamond, using the retained K3 observation: 61.3354 on AA Coding Agent Index v1.3, $3.175187 per evaluated task and 88,806 output tokens. These are K3's historical coordinates, not K3.1 measurements, forecasts or a claimed location on today's coding frontier. The proxy's Max label belongs to the predecessor observation.
Kimi K3 remains visible separately. The default now contains 17 models; All / history retains 22 families and 39 operating points. The inspector and exports identify the proxy basis. A future complete K3.1 coding run can replace that placeholder with a dated, source-linked observation without rewriting the predecessor's history.
A buying pilot that can change a decision
Start with one workflow and a fixed acceptance standard. For coding, use representative changes with known tests and review criteria. For document-heavy work, use a bounded source pack with an answer key, citation checks and explicit treatment of missing evidence. Keep customer or regulated data out until the selected account, contract and handling controls have been approved.
Compare ordinary Pro with the route already in use. Test Highspeed separately only where turnaround time has value. Keep task scope, permissions and acceptance tests consistent, and record the actual settings and provider route.
Measure accepted outputs, failed attempts, reviewer minutes, wall-clock completion time, total billed usage and material errors. Set the decision threshold before the run. A saving should survive review and rework, while a faster configuration should improve a meaningful service level rather than just tokens per second.
Aeon helps organizations evaluate models against their own workflows, then implement the routing, access controls and human approval points needed for dependable use. Explore the Model Economics Lab, start an AI Control and ROI Assessment, or contact Aeon to scope a workload-specific comparison.