Kimi K3.1 Pro vs Kimi K3: Same Backbone, Lower Price, and a Post-Training Bill That Came Due

Kimi K3.1 Pro is not a new model in the way Kimi K3 was a new model. Moonshot AI's October 24, 2026 release is explicitly a post-trained version of Kimi K3 with substantially larger post-training compute — the same architecture, the same 1,048,576-token context, the same native vision, and a different set of behaviours bolted on by reinforcement learning across hundreds of downstream domains. What makes the comparison interesting is not the delta in the benchmark table. It is that the newer, better model costs one-third less per token than the older one.

That inversion is rare enough to be worth examining carefully. Normally a lab ships a stronger model at the same price and lets the old one drift downmarket. Moonshot did the opposite: it kept the old price for the old model and cut the new one to $2.00 input and $10.00 output per million tokens, against Kimi K3's $3.00 and $15.00. Understanding why requires separating what actually changed in the model from what changed in the business around it.

The architecture did not move

Kimi K3 arrived on July 16, 2026 as the first open model in the 2.8-trillion-parameter class: a mixture-of-experts design with roughly 104 billion active parameters, 896 experts and 16 activated per token, a 1M-token window, native image and video understanding, tool calling, prompt caching and built-in reasoning. Its weights followed on July 27 under a licence that permits use, modification, distribution and fine-tuning, with a separate agreement required for model-as-a-service providers earning more than $20 million a year and attribution required for very large commercial products.

Kimi K3.1 Pro keeps all of it. Moonshot calls the new model "3T-class", which is positioning language; reporting from the launch puts it at the same 2.8 trillion total and 104 billion active. The context ceiling, the vision support and the 1M-token output limit are unchanged. If you were hoping for a cheaper active-parameter count or a longer window, this is not that release.

What did change is the training after pre-training. Moonshot describes large-scale joint reinforcement learning across hundreds of domains, supported by reworked sandbox infrastructure, rollout scheduling and environment-quality checks. In plain terms, the company spent a great deal of compute teaching the existing model to behave better on agentic and professional tasks rather than teaching it more facts. That is why the gains cluster where they do.

What the numbers say, and where the comparison is thin

The vendor's benchmark sheet is broad — 14 evaluations — but it is not a clean paired comparison. Moonshot published Kimi K3 figures for only a subset of the new tests, so some of the most dramatic deltas are against its own internal suites rather than against a public benchmark both models have taken.

• Long-horizon coding — Kimi Code Bench 3.0 Long-horizon: 61.2% for Kimi K3.1 Pro against 55.95% for Kimi K3. This is the cleanest predecessor number in the launch, and also the most vendor-controlled, since both the benchmark and the harness belong to Moonshot.

• Hallucination — Kimi Hallu Bench Hard claim-level rate fell from 23.4% on Kimi K3 to 8.3% with tools on Kimi K3.1 Pro. Without tools the new model scores 5.4%. Moonshot also says responses containing at least one factual error are now 42.4% of the predecessor's rate. For document-heavy work this is the most consequential improvement in the release.

• Finance and business operations — internal Financial Modeling 57.6 against 50.4, Financial Research 70.4 against 58.8, Business Operation 37 against 8.9. The last is a fourfold improvement on a task where the old model was barely functional. These are Moonshot's own evaluations, so treat the magnitude as directional.

• 3D and CAD — BenchCAD with tools 92.7, without tools 72.8. Moonshot did not publish the Kimi K3 figures for this test, so the gain is unmeasured in the launch post even though the absolute score is strong.

• Adversarial coding — xdev-bench-hard pass@3 is 81.0% for Kimi K3.1 Pro, against 33.3% for Kimi K3 and 85.7% for DeepSeek-V4.1-Flash-1017. A checkpoint trained without the company's risk-reduction measures scored 90.5%, and Moonshot attributes that 9.5-point gap to safety work and training-data controls.

• Frontier standing — Terminal-Bench 4.0 at 65.7%, behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. BenchCAD with tools at 92.7 trails Opus at 96.2 and Astra at 95.9. The pattern is consistent: close to the frontier, rarely on top of it.

None of this is independently verified. Kimi K3 entered the world with an Artificial Analysis Intelligence Index of 57, third globally at the time behind Claude Fable 5 and GPT-5.6 Sol, and no equivalent third-party score for K3.1 Pro existed on launch day. The harness problem is not hypothetical: Terminal-Bench scores swing by tens of points depending on scaffolding, which is why a 65.7% and a 58.2% produced in different labs are not a measurement of a 7.5-point gap. The direction is credible; the magnitude is not yet.

The price inversion is the real story

Kimi K3 launched as the most expensive Chinese model on the market, at $3.00 input and $15.00 output per million tokens, roughly three to four times the K2 generation. That pricing was a deliberate bet that capability, not cost, would win the frontier workload. Kimi K3.1 Pro reverses the direction on the newer model while leaving the older price intact.

• Input — $2.00 per million on K3.1 Pro against $3.00 on K3, a 33% cut.

• Output — $10.00 against $15.00, also 33%.

• Cache hit — $0.20 against $0.30, a 33% cut on the line that dominates agent workloads.

• Cache writes — $2.00 and $4.00 at five-minute and one-hour TTL, against $3.00 and $6.00.

• Highspeed variant — $4.00 and $20.00, with the same $0.20 cache-hit price; two to three times the output speed, roughly double the quota in Kimi Code.

The arithmetic is simple enough to do on a napkin. 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. An agent that re-reads a large repository every turn and hits cache 70% of the time pays about $4.11 on K3 and $2.74 on K3.1 Pro for the same shape of call. If the new model also finishes in fewer turns — which is the entire point of the long-horizon training — the cost per completed task falls by more than the token price suggests.

Why cut? The likeliest explanation is competitive rather than charitable. A 2.8T-parameter model is expensive to serve, and the market for frontier-adjacent models has spent 2026 converging on price. Moonshot's own K2 retirement schedule adds pressure: K2 series models stop being available to new users on October 31, 2026 and are due to be retired at the end of December. Pricing K3.1 Pro below K3 gives every K2 customer a reason to migrate upward rather than shop elsewhere. The older K3 price stays where it is because nobody who has already budgeted for it needs a discount to stay.

Where Kimi K3 is still the right answer

Three cases survive the comparison. The first is self-hosting today: Kimi K3's weights have been public since July 27, while Kimi K3.1 Pro's are promised only by November 4, 2026. If your deployment cannot call a hosted API, the newer model is not an option yet, whatever the benchmark sheet says.

The second is behavioural stability. A model that has been in production for three months has a known failure profile. Kimi K3.1 Pro changes reasoning behaviour by design, and its always-on thinking mode with a default of max effort is not a drop-in for every pipeline. Teams with tight evaluation suites should expect to re-baseline, not just swap a string.

The third is sampling control. Kimi K3.1 Pro fixes temperature at 1.0 and top_p at 0.95 and rejects other values; changing reasoning effort mid-conversation also invalidates prefix-cache hits. If your application depends on low-temperature determinism, check whether the older model gives you the control the newer one does not before assuming the upgrade is free.

The migration decision, without the hype

For coding agents, long-horizon tool use, financial research and document verification, the case for moving is strong: better vendor-reported scores in exactly those areas, a lower price on every billing line, and a hallucination rate that fell by roughly two-thirds. The risk is that all of it is self-reported, so the sensible path is a shadow evaluation on your own tasks before cutting over production traffic.

For batch generation and long-context retrieval, the cache-hit price is the number that matters, and it favors the new model by a third. For anyone pinned to open weights, the calendar decides: November 4 is the earliest date the comparison becomes real.

OrcaRouter's part in this is deliberately unglamorous. Our catalogue carries kimi/kimi-k3 today at Moonshot's list price with 0% markup, on the same key as the rest of your stack, so the old side of this comparison is already callable without a second contract. Kimi K3.1 Pro is not in the catalogue at the time of writing. When it lands, the pass-through model means the $2/$10 above is what you pay here rather than a reseller's version of it, and switching generations is a model-string change inside a routing policy rather than a migration project.

Verdict

Kimi K3.1 Pro is the better model and the cheaper one, which makes the recommendation easy and the evidence thinner than it looks. Moonshot has improved the behaviours that matter for agentic and professional work, cut every price line by a third, and disclosed a safety-related regression that most launch posts would have buried. What it has not done is submit the result to an independent harness. If you were already on Kimi K3, the upgrade pays for itself on the invoice before it pays for itself on the benchmark. If you were waiting for proof, the weights and the first outside evaluations are both due within two weeks — and that is when this comparison stops being a vendor claim and becomes a measurement.