Moonshot AI releases Kimi K3, a 2.8 trillion parameter open-weight model with API pricing and delayed weights availability
Model release ● Confirmed 95% confidence first seen
Moonshot AI released Kimi K3, a 2.8 trillion parameter model available via API at $3 per million input tokens and $15 per million output tokens, with open weights promised by July 27, 2026. The model demonstrates comparable performance to Claude and GPT-4 level systems but at significantly different cost-performance tradeoffs compared to competing Chinese models like DeepSeek V4 Pro and GLM-5.2.
Decision brief
- What changed
- Moonshot AI released Kimi K3, a 2.8 trillion parameter model accessible via API at $3/million input and $15/million output tokens, with open weights not scheduled for release until July 27, 2026. It performs comparably to Claude Opus/Sonnet-class and GPT-4/5.6-class models on coding and reasoning tasks, but is priced far above other Chinese open-weight competitors like DeepSeek V4 Pro (available now, MIT license, ~$0.04/task) and GLM-5.2.
- Why it matters
- Chinese labs are now producing frontier-comparable coding and reasoning models at a fraction of Western API costs (K3 is one-third to one-fifth the price of Claude/Opus tiers on some tasks), intensifying pressure on Anthropic and OpenAI's pricing power and margins. However, K3 itself is markedly slower (4x) and more expensive than sibling Chinese models, so procurement decisions should weigh latency and total task cost, not just headline token price, and multiple sources argue current US model pricing reflects compute scarcity rather than true cost structure—meaning the competitive threat may be more about infrastructure economics than model quality per se.
- Evidence
- Seven independent outlets (Simon Willison, MarkTechPost, Exponential View, TLDR Dev, The New Stack, TLDR, IEEE Spectrum) consistently report the same pricing figures ($3/$15 per million tokens) and comparable-to-frontier performance claims, with The New Stack providing direct benchmarked cost/speed comparisons against Claude Fable 5 and MarkTechPost providing head-to-head benchmarks against DeepSeek V4 Pro and GLM-5.2.
- What remains uncertain
- Benchmark rankings vary by source (K3 ranks third on Artificial Analysis Intelligence Index behind cheaper rivals), and it's unclear how representative the limited coding-task comparisons are of broader enterprise workloads; the claim that US pricing reflects 'artificial scarcity' rather than genuine cost advantage is asserted by only one source (TLDR) and not independently verified. The 8-month delay to open weights (July 2026) is a stated commitment, not a guaranteed outcome.
- Monitor next
- Watch whether Moonshot AI delivers open weights on schedule by July 27, 2026, and whether Anthropic or OpenAI adjust pricing in response to sustained cost/performance pressure from Chinese open-weight models.
Analytical support, not advice — assumptions and open questions stated above.
Source coverage
Kimi K3, and what we can still learn from the pelican benchmark
Simon Willison
Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2: Open Trillion-Scale MoE Models Compared on Benchmarks, License, and Serving Cost
MarkTechPost
🔮 Kimi K3 surprise & AI economics; the solar paradox; AI's right to learn, cancer vaccine & junior jobs++
Exponential View
The Kimi K3 Moment
TLDR Dev
newsletter
Claude Fable 5 vs. Kimi K3: Same results, one-third the cost, 4x slower
The New Stack
Who's Afraid of Chinese Models?
TLDR
newsletter
China’s Low-Priced Z.ai Model Is Exposing Costly Coder Habits
IEEE Spectrum AI