THE SHORT ANSWER
● Index score — 55 against 44 - eleven points in one generation
● Price — $2 / $10 per million against $3 / $15 - a third off every rate
● Hallucinations — 8.3% against 23.4% claim-level, on Moonshot's internal hard benchmark
● Catch — The November open-weight build has cyber capability deliberately removed
Generational upgrades usually cost more. This one costs a third less and scores eleven points higher, which makes it worth laying out plainly.
| Measure | Kimi K3 | Kimi K3.1 Pro |
|---|
| AA Intelligence Index | 44 | 55 |
| Input per 1M tokens | $3.00 | $2.00 |
| Output per 1M tokens | $15.00 | $10.00 |
| Cache hits per 1M | $0.30 | $0.20 |
| Kimi Code Bench 3.0 long-horizon | 55.95 | 61.2 |
| Claim-level hallucinations (internal) | 23.4% | 8.3% |
| Context window | 1M tokens | 1M tokens |
| Parameters | not restated at launch | 2.8T total, 104B active |
The price cut is the real headline
Every rate fell by the same third: input $3 to $2, output $15 to $10, cache $0.30 to $0.20. At a million output tokens a day that is $150 a month saved on a model that also scores eleven points higher. There is no tier structure to navigate and no annual commitment attached. The Highspeed variant sits above it at $4 and $20.
The hallucination number deserves scrutiny
23.4% to 8.3% claim-level is a very large move, and it is measured on Moonshot's own Kimi Hallu Bench Hard rather than by a third party. Treat it as a direction rather than a figure until someone independent runs it. That said, if K3's hallucination rate was what stopped you deploying it, this is the specific thing Moonshot says it worked on.
What did not change
The 1 million token context window carries over. So does the overall shape - a mixture-of-experts served through the same API surface - so a migration is a model-string change rather than a rewrite.
The one reason not to upgrade blindly
If you are self-hosting K3 weights today, the November K3.1 Pro weights are not a clean superset. Moonshot removed cyber-related data from mid-training and reinforcement learning before the open-weight release, taking its internal xdev-bench-hard score from 90.5% to 81.0% at pass@3. That was a deliberate safety decision and it is defensible, but it means capability moved in both directions. If any part of your workload touches security tooling, test before you swap.