Simon Willison (@simonwillison.net): My notes on Kimi K3, plus some thoughts on what we can still learn from the pelican benchmark even while it becomes further detached from how good the models are at the things that matter (like agenti

Sam Altman announces GPT-5.6 sol, a new OpenAI model that offers significant cost and efficiency improvements over competing models. The model is priced at half the cost and delivers approximately twice the token efficiency compared to 'fable' for accomplishing the same tasks. This represents a major price-performance breakthrough in the AI model market.

Early testing indicates that Sol and Fable models represent significant performance jumps over previous AI models, creating a substantial gap between them and the next-best alternatives. These two models are positioned as the top choices for any work where superior intelligence matters, though users may have individual preferences between them.

SpaceXAI has launched Grok 4.5, an Opus-class model developed in partnership with Cursor that's positioned as comparable to GPT 5.5/Opus 4.7 but with better speed and cost efficiency. The model represents a new weight class focused specifically on coding and agent applications, marking Cursor's first expansion beyond pure software engineering models. It's immediately available through multiple platforms including Grok API, Cursor, and Hermes Agent.

Thinking Machines Lab has released Inkling, a new open-weights multimodal foundation model family with a 975B parameter Mixture-of-Experts architecture (41B active), 1M token context, and training on 45 trillion tokens of text, images, audio, and video. The launch includes Inkling-Small (276B total, 12B active) for lower-cost deployment, with weights available on Hugging Face and immediate fine-tuning support on their Tinker platform. The company positions this as a customizable multimodal base model for practical use rather than a benchmark-chasing flagship.

Moonshot AI has launched Kimi K3 2.8T-A50B, claiming it's the largest open-weight model ever released with 2.8 trillion parameters, 1M-token context, and multimodal capabilities. Early benchmarks suggest its intelligence is comparable to top closed models like Claude Opus 4.8 and GPT-5.5, though it reportedly still lags behind Claude Fable 5 and GPT-5.6 Sol. The model is positioned for agentic coding and self-evolving workflows, with open weights promised to be released by July 27, 2026.