Kimi K3 matches GPT-4 performance at half the cost

Moonshot AI's Kimi K3 benchmarks at frontier level while undercutting GPT pricing by 50%, pressuring inference economics

Moonshot AI's open-weight Kimi K3 model matches or exceeds performance of Claude Opus 4.8 and Claude Fable while pricing at half the cost of OpenAI's premium tier (GPT 5.6). The competitive pricing shift reflects a broader structural move in AI economics: as open-weight models close the capability gap with proprietary systems, margin pressure migrates from the model layer toward compute infrastructure and chip suppliers, potentially strengthening the infrastructure payback case despite not yet displacing enterprise adoption of proprietary models.

San Francisco cuts inclusionary housing requirement to 5%, exempts small buildings in bid to restart construction

The San Francisco Board of Supervisors voted 9-2 to slash the mandatory below-market housing requirement on new construction from 15% to 5%, completing a decline from 25% at the 2016 peak and exempting small buildings entirely. The move represents the city's most aggressive step to date to restart homebuilding, though the city's own economists have indicated that even a zero percent requirement does not make new construction economically viable under current conditions.

Kimi K3: 2.8T-Parameter Open-Weight Model with 1M Context and Frontier Coding Capabilities

Moonshot AI in Moonshot AI (Kimi) published a report titled "Kimi K3: 2.8T-Parameter Open-Weight Model with 1M Context and Frontier Coding Capabilities". Main takeaways:

• Kimi K3 achieves #3 on Artificial Analysis Intelligence Index (57 points), behind only Fable 5 and GPT-5.6 Sol, with 1668 Elo on agentic task benchmarks and 53% on AutomationBench-AA (ranked #1).
• Model demonstrates frontier long-horizon coding: autonomously optimized GPU kernels competitive with Fable 5, built MiniTriton compiler from scratch matching Triton performance, designed a functional chip in 48 hours using open-source EDA tools.
• 2.5× improvement in scaling efficiency vs. Kimi K2 through Delta Attention, Attention Residuals architecture, and scaled MoE sparsity (16 of 896 experts active); 21% fewer output tokens than K2.6 while achieving higher scores.
• Pricing at $3.00/$15.00 per 1M input/output tokens with 90% cache discount; cost per task ($0.94) similar to GPT-5.6 Sol, half of Opus 4.8; weights to release by July 27, 2026.