Kimi K3.1 Pro Launches Cheaper Than Its Own Predecessor: $2/$10 and a Rougher Benchmark Sheet

Moonshot AI released Kimi K3.1 Pro on October 24, 2026, and the most consequential number in the launch is not on the benchmark chart. It is on the rate card: $2.00 per million input tokens and $10.00 per million output tokens, against $3.00 and $15.00 for the Kimi K3 it is built from. A successor that is one-third cheaper than its predecessor is unusual enough to lead with, because unlike a benchmark score it does not depend on whose harness ran the test. If you already pay for Kimi K3, the upgrade case starts with a smaller invoice.

The second number is the one Moonshot wants you to notice: 65.7% on Terminal-Bench 4.0, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. That puts the new model second on the vendor's own table, not first. It is an honest thing to publish on launch day, and it tells you something about how this release is positioned. Kimi K3.1 Pro is not being sold as the strongest model in the world. It is being sold as a cheaper one that is close enough to the frontier to win on cost per finished task.

What actually changed between K3 and K3.1 Pro

Moonshot is explicit that this is a post-training release, not a new architecture. The company calls Kimi K3.1 Pro a post-trained version of Kimi K3 trained with substantially larger post-training compute, including large-scale joint reinforcement learning across hundreds of downstream domains. The context window is unchanged at 1,048,576 tokens. Vision and video input are unchanged. The 1M-token output ceiling is unchanged.

What that means in practice is that the base model's knowledge and reasoning capacity are the same assets as before; the difference is in behaviour. Moonshot says it reworked the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and environment-quality checks to make that post-training run possible. Those are engineering claims about how the model was made, and they are hard to verify from outside. The behavioural claims are easier to test, and that is where the launch numbers live.

Parameter count is where the vendor language gets slippery. Moonshot calls it a 3T-class open-weight model. Reporting from the launch puts the backbone at 2.8 trillion total parameters with roughly 104 billion active, which is the same mixture-of-experts configuration as Kimi K3. Both can be true: "3T-class" is a positioning phrase, and 2.8T rounds into it. What matters for your bill and your latency is the active count, and there is no evidence the active count moved.

The benchmark sheet, and how much of it to trust

Moonshot published results across 14 coding, science and knowledge-work evaluations. The pattern is consistent: large gains on the tasks the company trains for, near-parity with the closed frontier on the rest, and one deliberate disclosure of a regression.

• Coding at the frontier — Terminal-Bench 4.0: Kimi K3.1 Pro 65.7%, Claude Opus 5.5 66.4%, GPT-6 Astra 58.2%. Vendor-run.

• Long-horizon coding — Kimi Code Bench 3.0 Long-horizon: 61.2% for K3.1 Pro against 55.95% for Kimi K3. This is the cleanest predecessor comparison in the set, and it is also the most vendor-controlled, since the benchmark and the harness are both Moonshot's.

• 3D and CAD — BenchCAD with tools: 92.7 for K3.1 Pro, 96.2 for Opus 5.5, 95.9 for GPT-6 Astra, 92.6 for Claude Fable 5.1. Without tools: 72.8, against 73.0 for Opus 5.5 and 60.6 for Fable 5.1. The with-tools spread is narrow enough that the ordering could move on a rerun.

• Hallucination — Kimi Hallu Bench Hard: claim-level rate of 8.3% with tools, down from 23.4% for Kimi K3. GPT-6 Astra scores 5.5% and Opus 5.5 15.0%. Without tools the rate is 5.4%, against 4.2% for Astra and 6.8% for Opus.

• Finance and operations — Frontier Finance 58.6; internal Financial Modeling 57.6 against 50.4 for K3; internal Financial Research 70.4 against 58.8; internal Business Operation 37 against 8.9. Those last three are Moonshot's own evaluations, so they measure improvement on the company's chosen tasks, not standing in a public arena.

• The regression — xdev-bench-hard pass@3 is 81.0%. A checkpoint trained without the company's risk-reduction measures scored 90.5% on the same test. Moonshot published that gap and attributed it to targeted safety work and training-data controls. Whatever you think of the trade, disclosing a 9.5-point capability cost is not how a launch post usually reads.

The honest summary is that these are vendor-reported numbers with no independent replication yet. Kimi K3 had an Artificial Analysis Intelligence Index of 57 at its July launch, and no equivalent third-party score for K3.1 Pro existed on release day. The harness matters enormously here: Terminal-Bench results move by tens of points depending on scaffolding, so a 65.7% and a 58.2% produced by different labs on different harnesses are not a clean measurement of a 7.5-point gap. Treat the table as a claim about direction, not magnitude.

What the price cut actually does to a workload

The API price is flat across the full 1M-token window, with no long-context surcharge. The full card, in US dollars per million tokens:

• Kimi K3.1 Pro — $2.00 input, $10.00 output, $0.20 cache hit, $2.00 cache write at five-minute TTL, $4.00 at one-hour TTL.

• Kimi K3.1 Pro Highspeed — $4.00 input, $20.00 output, $0.20 cache hit, $4.00 and $8.00 cache writes. Same model, two to three times the output speed, roughly twice the quota consumption in Kimi Code.

• Kimi K3 — $3.00 input, $15.00 output, $0.30 cache hit, $3.00 and $6.00 cache writes.

A request that reads 1M input tokens and writes 200K output costs $6.00 on Kimi K3 and $4.00 on Kimi K3.1 Pro. Add realistic prompt caching and the gap holds at about one-third, because the cache-hit price fell from $0.30 to $0.20 as well. For an agent that re-reads a large repository or a long financial document on every turn, that is the difference between a workflow that pays for itself and one that does not. Moonshot has effectively priced the upgrade as a margin cut rather than a premium, which is a strange move for a model that just posted better numbers almost everywhere — and a welcome one for anyone whose K3 bill was the reason they were shopping.

There is a catch in the mechanics. Kimi K3.1 Pro always thinks: reasoning cannot be turned off, and the reasoning_effort field accepts low, high or max, defaulting to max. Sampling parameters are fixed at temperature 1.0 and top_p 0.95, and passing other values returns an error. Changing reasoning effort mid-conversation invalidates prefix-cache hits, which quietly raises the effective cost of the tuning loop you were planning to run. The cheaper token price is real; the cheaper workflow depends on how often you change your mind mid-context.

Where you can run it, and the date that matters more than the launch

Kimi K3.1 Pro is live in the Kimi consumer app on web, iOS, Android and HarmonyOS, in Kimi Work desktop version 3.1.0 or later, and in Kimi Code through the model selector. On the API it is available as kimi-k3.1-pro and kimi-k3.1-pro-highspeed through Moonshot's own platform, with zero data retention offered and enterprise access off by default at launch. It is also appearing through third-party inference platforms, which is normal for a model this size and says nothing about how well any of them serve it.

The more important date is November 4, 2026. Moonshot says full model weights will be released by then. That is what makes the "open-weight" description more than marketing, and it is also the point at which the cost story changes again: self-hosting a 2.8T-parameter mixture-of-experts model is not cheap, but a permissive licence removes the per-token meter for anyone with the GPUs. Until the weights land, the price comparison above is the whole comparison.

One more operational note for teams still on the K2 generation: Moonshot's pricing page says K2 series models stop being available to newly registered users on October 31, 2026, and are scheduled for retirement at the end of December. If you are on K2.6 or K2.7 Code, the migration decision is being made for you, and K3.1 Pro is the intended destination.

Should you switch on day one?

If your workload is coding, agentic tool use or financial document work, the combination of a lower price and a better vendor-reported score is a straightforward case for a trial. If it is long-context retrieval with stable prompts, the cache-hit price is the number to model, and it favors the new model too.

If your production path depends on reproducible sampling, read the fixed-parameter constraint before you migrate. Temperature 1.0 is not a neutral default for every task, and a model that refuses to accept a different value is a different kind of dependency than one that does.

And if you need independent evidence, wait. Nobody outside Moonshot has published a K3.1 Pro evaluation yet. The launch numbers are internally consistent and unusually candid about the safety-related regression, but they are still the vendor grading its own homework. A model this cheap will get tested quickly; the first independent harness results are the ones that will tell you whether the 65.7% survives contact with someone else's scaffolding.

For teams already routing through OrcaRouter, the practical change is smaller than the launch suggests. Our catalogue carries kimi/kimi-k3 today at Moonshot's list price with 0% markup, so the K3 side of this comparison is live on the same key as the rest of your stack. Kimi K3.1 Pro is not in the catalogue at the time of writing; when it is added, moving from one generation to the other is a model-string change rather than a second contract and a new SDK. The price cut is the reason to care, and the pass-through model is the reason it will show up on our side at the same number Moonshot publishes, not a marked-up one.

The launch in one paragraph

Kimi K3.1 Pro is a post-training upgrade to Kimi K3, sold at two-thirds of the older model's price, with a strong vendor benchmark sheet, a real gain in hallucination control, a disclosed safety-related regression on one adversarial coding test, and no independent verification yet. The launch is genuinely news because of the price and the weights commitment, not because the model is a new architecture. The date to circle is November 4, when the weights are due and the first serious outside evaluations should start arriving.