Models→Reported→Simon Willison's Weblog
Chinese AI lab Moonshot AI has announced Kimi K3, a model with 2.8 trillion parameters, describing it as their most capable to date. The model is available via website and API, with open weights promised by July 27, 2026. Self-reported benchmarks suggest Kimi K3 outperforms Claude Opus 4.8 and GPT-5.5 on some tasks, and pricing is set at $3 per million input tokens and $15 per million output tokens, making it the most expensive Chinese model so far.
Why it matters: Kimi K3 represents a new scale for open-weight AI models, intensifying competition with leading proprietary systems.
Jul 16, 2026
Models→Reported→Simon Willison's Weblog
OpenAI's GPT-5.6 models are now generally available in three sizes: Luna, Terra, and Sol. They feature a 1 million token context window, 128,000 output tokens, and claim superior agentic performance on Agents' Last Exam, with Sol scoring 53.6, beating Claude Fable 5 by 13.1 points. However, on SWE-Bench Pro, Sol scored 64.6% compared to Fable 5's 80%, and OpenAI has criticized that benchmark as having approximately 30% broken tasks.
Why it matters: The GPT-5.6 family introduces tiered pricing and efficiency claims that could reshape competition in the AI model market, especially for long-running agentic tasks.
Jul 10, 2026
Models→Reported→Simon Willison's Weblog
Simon Willison released sqlite-utils 4.0rc2, developed primarily using Claude Fable for about $149.25. The AI identified critical bugs, including a data loss issue in delete_where() that left connections in a broken state. Over 37 prompts and 34 commits, the team resolved all release blockers and made design improvements.
Why it matters: This demonstrates the growing capability of AI coding agents to perform complex software maintenance tasks, including identifying subtle bugs and managing major version releases.
Jul 10, 2026
Open Source→Reported→Simon Willison's Weblog
Simon Willison released shot-scraper 1.10, introducing a new 'shot-scraper video' command that uses a storyboard.yml file to define routines for web apps and records video demos using Playwright. The tool is intended to help coding agents produce demonstrations of their work, with an example showing bulk CSV import into Datasette.
Why it matters: This tool enables AI coding agents to automatically generate video demonstrations of their work, improving communication and review processes.
Jul 10, 2026
Open Source→Reported→Simon Willison's Weblog
Simon Willison ported the Moebius 0.2B lightweight image inpainting model to run in the browser using WebGPU, removing the need for PyTorch and CUDA. The demo is available at simonw.github.io/moebius-web/. The project was completed as a side task while waiting for an AI coding agent to finish a larger refactor.
Why it matters: This shows that efficient AI models can be made accessible directly in the browser, lowering hardware barriers and broadening access to AI image editing tools.
Jul 10, 2026
Models→Reported→Simon Willison's Weblog
Chinese AI lab Z.ai has released GLM-5.2, a 753B-parameter Mixture-of-Experts model with 40 active parameters, under an MIT license. GLM-5.2 leads the Artificial Analysis Intelligence Index among open-weight models and ranks second on the Code Arena WebDev leaderboard, behind only Claude Fable 5. The model features a 1 million token context window and is available via OpenRouter at competitive pricing.
Why it matters: GLM-5.2 sets a new performance standard for open-weight text-only LLMs, rivaling proprietary models in coding and general intelligence benchmarks.
Jul 10, 2026
Open Source→Reported→Simon Willison's Weblog
Pyodide 314.0 now allows Python packages built for Pyodide to be published directly to PyPI, eliminating the previous need for maintainers to manually build and host over 300 packages. Package maintainers can now build and publish Pyodide wheels to PyPI in the same way as native wheels, as demonstrated by Simon Willison with the luau-wasm package.
Why it matters: This change reduces the maintenance burden on Pyodide maintainers and enables easier distribution of Python packages compiled to WebAssembly.
Jul 10, 2026
Models→Reported→Simon Willison's Weblog
Simon Willison reports that Claude Fable 5 autonomously used browser automation and custom screenshot techniques to debug a UI glitch. The model opened browser windows, iterated through macOS windows, and used Python with pyobjc-framework-Quartz to capture screenshots without explicit instruction. Willison describes the model as 'relentlessly proactive' in pursuing its goals.
Why it matters: This demonstrates a significant leap in AI agent autonomy, where a model independently devises and executes multi-step tool use strategies beyond its explicit instructions.
Jul 10, 2026
Models→Reported→Simon Willison's Weblog
Anthropic has launched Claude Fable 5, a new frontier model with strict safety guardrails, and Claude Mythos 5, which shares its capabilities but lacks the safety classifiers. Both models feature a 1 million token context window, 128,000 maximum output tokens, and are priced at $10 per million input tokens and $50 per million output tokens. Early impressions from Simon Willison describe Fable 5 as slow, expensive, but highly capable.
Why it matters: Claude Fable 5 introduces enhanced safety mechanisms for advanced AI models, including automatic fallback when guardrails are triggered.
Jul 10, 2026
Models→Reported→Simon Willison's Weblog
Anthropic has released Claude Opus 4.8, described as a modest but tangible improvement over its predecessor. The model is noted for its increased honesty, being around four times less likely to allow flaws in code to pass unremarked, and achieving the lowest incorrect rate on hallucination benchmarks by abstaining on uncertain questions. Pricing remains at $5/million input and $25/million output, with a new fast mode at double the price for research preview organizations.
Why it matters: This release signals a shift in AI development priorities toward honesty and reliability over raw capability, with measurable reductions in hallucination and unsupported claims.
Jul 10, 2026
Policy Safety→Reported→Simon Willison's Weblog
Pope Leo XIV has released an encyclical titled 'Magnifica Humanitas' focused on safeguarding the human person in the era of artificial intelligence. The document addresses ethical challenges posed by AI, including issues of interpretability in large language models, and draws parallels to the Church's response to the industrial revolution. The Pope chose his name in honor of Leo XIII, who addressed similar societal shifts in his 1891 encyclical Rerum novarum.
Why it matters: This encyclical marks a significant institutional statement from the Vatican on the ethical implications of AI, framing it as a transformative force akin to the industrial revolution.
Jul 10, 2026