Moonshot ships Kimi K3 at 2.8T parameters, Sonnet pricing

Moonshot AI announced Kimi K3 this week and the interesting thing is not that it topped any leaderboard, it is that a Chinese open-weight lab priced itself up to Anthropic's Claude Sonnet tier. Simon Willison's write-up records the numbers: $3 per million input tokens and $15 per million output tokens, versus the earlier Kimi K2.6 rates of $0.95 and $4. Moonshot is also calling this the first open 3T-class model, rounding 2.8 trillion parameters up, with an open weight release promised by July 27, 2026.

On Artificial Analysis's private long-horizon knowledge work evaluation, K3 lands at an Elo of 1547, up 732 points from K2.6 and behind only Claude Fable 5. Cost per task comes in around $0.94, similar to GPT-5.6 Sol at $1.04 and roughly half of Opus 4.8 at $1.80, with 21% fewer output tokens than K2.6.

Why any of that matters if you are not running a Chinese lab: the shape of the market is changing. For the last few cycles the working assumption was that Chinese open-weight releases would keep undercutting the frontier on price while trailing on capability. K3 breaks half of that, matching Sonnet's list prices at the same time it closes on the top of the leaderboard. If your product plan quietly relied on Kimi as the cheap fallback, that thesis needs a rewrite.

The honest caveat is one Willison flags himself. The pelican-riding-a-bicycle prompt he ran cost 25 cents on K3, because the model spent 13,241 reasoning tokens to output 3,417 tokens of response, and he notes the test 'doesn't touch at all on the thing that matters most for today's model: agentic tool calling and the ability to operate tools reliably as conversations grow in length.' What the piece does not give you is any agentic benchmark result, any real-world tool-use test, or a read on how the fixed 'max' reasoning setting behaves under production latency budgets.

The forward-looking part is still worth watching. If Moonshot lands the open weights on July 27 as promised, teams priced out of hosted frontier APIs get a 3T-class model to run on their own hardware, and the benchmark community gets a very fresh, very token-hungry target to build the next generation of agentic evaluations against.