Kimi K3 Ships With a 3x Price Rise — and No Weights Yet

Moonshot AI released Kimi K3 on July 16: 2.8 trillion total parameters, a Stable LatentMoE architecture activating 16 of 896 experts, and a one-million- token context window. It took first place on Arena.ai's Frontend Code Arena at 1,679 points, climbing seventeen places from K2.6's eighteenth and passing Claude Fable 5. It leads six of seven frontend domains — the exception is Gaming, where Fable 5 still holds the top spot.

That is the story everyone ran. Two facts underneath it got almost no coverage, and both matter more than the leaderboard.

The weights are not out

Every headline this week called K3 the largest open-weights model ever released. The weights have not been released. Moonshot has promised them "by July 27," eleven days after the launch.

Until then, "open-weight" is a characterization, not an artifact. The Modified MIT license that most coverage asserted is also unconfirmed — it is inferred from the K2 precedent rather than stated. Both may well land exactly as promised. But a model you cannot download is not an open model yet, and July 27 is now a real test of how open this path actually is.

The discount is over

Moonshot raised prices, substantially, and in the same breath as claiming the open-weight crown.

| Tier | K2.6 | K3 | | ------------------ | ----- | ------ | | Input (cache hit) | $0.16 | $0.30 | | Input (cache miss) | $0.95 | $3.00 | | Output | $4.00 | $15.00 |

Per million tokens.

Uncached input rose roughly 3.2x. Output rose nearly 4x. Critically, there is no tier where K3 is cheaper than the model it replaces — even the cache-hit rate sits above what K2.6 charged for cached input. This lands K3 at rough parity with Western premium pricing.

For two years the strategic logic of Chinese open-weight models was simple: match capability, undercut price brutally, and commoditize the frontier labs' margin. The discount was the entire wedge. K3 abandons it.

That is not a pricing decision. It is a statement about market position — the behavior of a supplier that no longer believes it needs to buy share.

The benchmark numbers deserve more scepticism than they got

Moonshot's launch table puts K3 third on GDPval-AA v2 at 1,687, behind Claude Fable 5 Max and GPT-5.6 Sol Max and ahead of Claude Opus 4.8 at 1,600. We could not reach a primary leaderboard to confirm any of those figures, and a competing set is circulating — Artificial Analysis is reported showing K3 at 1,668 against Fable 5 at 1,760 on a similarly-named benchmark.

Notably, Opus 4.8 appears as exactly 1,600 in both sets while everything around it disagrees, which is the signature of one number propagating across otherwise inconsistent tables. Either two benchmark variants are being conflated, or someone is transcribing badly. Treat the precise values as vendor arithmetic. The Arena result is the one with a published primary source behind it.

The effective price is higher than the rate card

Simon Willison ran K3 against his standing pelican-on-a-bicycle SVG benchmark and watched it burn 13,241 reasoning tokens to produce 3,417 tokens of output — one drawing, twenty-five cents. A model that deliberates that hard costs more than its published rate implies.

His own caveat is sharper than the pricing point. The pelican, he notes, "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." He is naming the limits of his own test — but the criticism generalizes. Nothing in anyone's published benchmark set tells you what production systems actually depend on.

Why this matters

The open-weight discount was load-bearing for a lot of 2026 budget planning. If the leading Chinese labs have concluded they can charge frontier prices — in a domestic market that regulators have walled for them, on the same week Beijing approved a seven-name register of on-device AI suppliers and founded a 29-nation governance body in Shanghai — then the cheap floor everyone was budgeting against has moved up.

Watch July 27. Whether those weights actually land, and under what license, tells you more about the next two years than the leaderboard does.