Kimi K3 goes 2.8T params, and pricier than ever
What happened
Moonshot AI announced Kimi K3, its "most capable model to date" at 2.8 trillion parameters, calling it the first "open 3T-class model" (rounding up, generously). Open weights land by July 27, but it's already live via API, beating Claude Opus 4.8 max and GPT-5.5 on several benchmarks while trailing Claude Fable 5 and GPT-5.6 Sol.
Why this matters
K3 also comes with a major price hike — $3/$15 per million tokens, up from K2.6's $0.95/$4 — putting it in Claude Sonnet territory and making it the priciest release yet from a Chinese lab. That's a signal the 'cheap open-weight alternative' era might be quietly ending as labs chase frontier performance instead of frontier discounts.
The slightly cynical read
Everyone's racing to out-parameter each other while benchmark scores get squishier by the month — Simon Willison notes his famous 'pelican riding a bicycle' SVG test has basically stopped correlating with real model quality. Self-reported benchmarks from the lab that made the model are, shockingly, favorable to that model.
What to watch next
Watch whether the open-weight release (due July 27) actually matches the hosted API's benchmarks, and whether rivals like DeepSeek respond with their own scale-up. Also worth tracking: does the agentic tool-calling performance — the thing that actually matters now — hold up under real workloads?