Moonshot Cut Kimi K3.1 Pro's Price by a Third. Independent Testing Says It Isn't Actually Cheaper.
Moonshot AI pushed Kimi K3.1 Pro live on its API on October 24, capping a rollout that had been leaking in pieces for weeks: the model identifier surfaced in Kimi Code's switcher on October 21, an official teaser followed on October 23, and the model itself arrived the next day. Moonshot's headline pitch is pricing, not capability: $2 per million input tokens and $10 per million output tokens, both about a third below what Kimi K3 charges. Full open weights are promised by November 4, continuing Moonshot's practice of shipping an API version first and open-sourcing later.
The release itself was a long time coming. A leaker first floated a 'K3.1' successor on X in late July, days after K3 shipped, and repeatedly moved the goalposts — August, then 'before the end of October' — while Moonshot offered nothing official until a cryptic digit string posted and deleted from its account in September. That pattern of vendor silence punctuated by unverified leaks is worth remembering when reading Moonshot's own numbers for the finished model, because there still aren't many independent ones to check them against.
The one independent scorecard tells a different cost story
Artificial Analysis, which runs its own standardized battery rather than relying on vendor-submitted numbers, puts K3.1 Pro's Intelligence Index at 55 — fourth of 215 models tracked, and genuinely strong. But the same evaluation flags two problems Moonshot's pricing page doesn't mention. Output speed comes in at 57.5 tokens per second, ranked 141st of 186 models and well below the 82.5 median — among the slower frontier-class models currently measured. And the model is unusually verbose: it burned through 150 million output tokens to complete the Intelligence Index suite, against an 87 million median, nearly double the typical token count for the same tasks.
Multiply a lower per-token price by nearly twice the tokens needed to finish a comparable task, and the arithmetic mostly cancels out. Artificial Analysis's blended cost-per-task metric puts K3.1 Pro at $2.11, ranking just 85th of 113 models on cost efficiency — squarely average, not the discount the sticker price implies. A developer switching from K3 to K3.1 Pro expecting a one-third cost reduction on real workloads would likely see something closer to flat spending, offset by noticeably slower responses in anything latency-sensitive.
What Moonshot isn't saying
Moonshot's own materials for K3.1 Pro lean on the per-token price cut and the '3T-class' open-weight framing, but conspicuously don't publish a benchmark table the way the company did for K3 in July — when it was candid enough to admit K3 trailed Claude and GPT-5.6 Sol on specific tests even while beating them on others. The absence of a comparable disclosure this time, paired with a verbosity problem that independent testing caught within the model's first 48 hours on the API, suggests the company is marketing the number most favorable to it and leaving the one that actually determines production cost to third parties to discover.
The bigger open question is whether any of this holds up once weights are actually public. Moonshot has missed self-imposed and leaked windows for this model for three months running; a November 4 open-weights date is a promise, not a fact yet. Until independent groups can run K3.1 Pro outside Moonshot's own API — on their own hardware, under their own cost accounting — both the intelligence ranking and the verbosity problem remain measurements of one company's hosted service, not of the model itself.