Moonshot AI released Kimi K3 on Thursday, July 16 at roughly 15:00 Beijing time. The model is a mixture-of-experts architecture at 2.8 trillion total parameters with a native one-million-token context window and native video and image inputs. Moonshot published the launch pricing at $3 per million input tokens and $15 per million output tokens through its own API and via partners including Together AI, Fireworks and OpenRouter, with the full open weights targeted for release on Hugging Face on July 27 and an intermediate developer preview drop on July 22. The API opened at scale immediately.
The number that drove the week landed within eight hours. Arena.ai's Frontend Code leaderboard, the benchmark professional developers have converged on for code-agent evaluation, opened Kimi K3 at Elo 1,679. Claude Fable 5 sat at 1,631 and GPT-5.6 Sol at 1,618. Bank of America's coverage note pegged Kimi K3's effective per-task cost at roughly half of Sol's for equivalent output length once the shorter completion times were factored in. The Elo lead is narrow in absolute terms and Fable 5 remains ahead on longer-horizon reasoning suites, but Frontend Code is the leaderboard the industry uses to price coding-agent seats. The number moved the pricing conversation immediately.
The Friday repricing
The market reaction lagged by roughly 24 hours as Wednesday's Kimi headlines were absorbed. Friday, July 17 opened with the Asian sessions selling first. TSMC in Taipei closed down 7 percent despite having reported a 77 percent year-over-year profit surge to NT$706.56 billion on Thursday morning and lifting 2026 capex guidance to $60-64 billion. SoftBank in Tokyo shed 9 percent as investors reassessed the Stargate spend thesis in the shadow of a viable open-weight frontier. SK Hynix in Seoul lost another 11.5 percent in what became the fourth consecutive down session for the memory sector, dragging Samsung more than 8 percent lower. The Kospi closed the session down 4.4 percent, its worst single day of 2026.
New York followed. Alphabet closed down 4.4 percent to $353.81 on renewed questions about the delayed Gemini 3.5 Pro cycle. Meta gave up 2.4 percent. Nvidia lost 1.2 percent and briefly ceded the world's-most-valuable-company slot to Apple at $4.88 trillion intraday before recapturing the top spot by the close at $4.92 trillion. The Hong Kong-listed Chinese AI peers took the hardest damage on the theory that Kimi K3 makes their own model layers less differentiated: Z.ai crashed 27 percent, MiniMax 16 percent, and Baichuan 11 percent. Bitcoin and the major cryptocurrencies followed the AI complex lower as crypto traded as an AI-capex proxy through the day.
"This is the second DeepSeek moment in 18 months. The first one was a shock. The second one erases the assumption that the frontier is a US-only construction, which is a different kind of pricing event."— Dylan Patel, SemiAnalysis, quoted in Bloomberg (2026-07-18)
Anthropic capitulates inside 48 hours
The clearest downstream signal came from Anthropic. Claude Fable 5 had launched in June at a metered-only $10 input and $50 output per million tokens with a rolling 5-hour bundle inside Max and Team Premium seat plans. On Sunday, July 19 at 20:04 UTC, Anthropic announced that Fable 5 would be permanently bundled into Max and Team Premium plans starting Monday, July 20, at 50 percent of the previous usage limits with no metered surcharge for in-bundle consumption. Pro and Team Standard reverted to API-only access with a one-time $100 credit. The-Decoder's coverage read the move as an acknowledgement that a mid-tier professional developer buying a subscription now had a plausible non-Anthropic path to a comparably-priced frontier coding model.
The Anthropic reversal has a specific inside-baseball read. Since Claude Fable 5 shipped in June, Anthropic had held the position that metered pricing was the future because the model was expensive to serve at any bundle price the market would accept. Kimi K3 broke that argument by proving that a comparable coding-agent capability could be served at a $3/$15 price point without a metered wrapper. The reversal is a real-time repricing of what a Fable-class token is worth on the enterprise seat side.
What Kimi K3 actually is
Simon Willison's post, which was the first English-language technical write-up to fetch numbers directly from Moonshot's model card, described the architecture in detail. Kimi K3 uses 896 experts with 16 active per token, a Kimi Delta Attention hybrid layer, and MXFP4 quantization-aware training with 32-bit master weights. The Frontend Code Arena result comes with the caveats that Arena is a preference benchmark with known selection effects and that the training data cutoff is not fully disclosed. Willison called the compact-context inference numbers "extraordinary for a Chinese open-weight release" and flagged the July 27 weights drop as the event to watch, since API-only frontier models cannot be independently evaluated the way open-weights ones can.
"The K3 model card claims a 40 percent reduction in cost per correct answer against Sol on their internal FrontierBench. The number I want to see is the same figure on the SWE-Bench Pro suite after the July 27 weights drop, when we can run it locally."— Simon Willison, 2026-07-16
The DIGITIMES Friday analysis, syndicated by Japan Times, framed Kimi K3 as the second DeepSeek moment in 18 months, this time confirming rather than surprising. The National ran a Fortune-sourced profile of Yang Zhilin, Moonshot's 32-year-old founder, whose Carnegie Mellon PhD advisor Russ Salakhutdinov confirmed that Yang had multiple US tech job offers including Apple as he finished his 2019 doctorate but chose to return to China. American Bazaar's essay ran under the headline "AI's Linux moment" and argued that the trillion-dollar closed-model stack now has to price against something.