What happened
Simon Willison released LLM 0.32, the largest update since the project's launch. The CLI now displays reasoning traces from reasoning models, supports GPT-5.6 as the default model (GPT-5.6 Luna), and integrates server-side tools from OpenAI (CodeInterpreter, WebSearch) and Anthropic (WebSearch, WebFetch, CodeExecution, AnthropicMCP). The Python API now accepts a model.prompt(messages=[]) parameter and exposes stream_events() to handle mixed response types (reasoning, text, tool calls, images).
Why it matters
The update bridges the gap between how LLMs actually work (request-per-turn with full message history) and how developers have been coding against them. Server-side tools eliminate the need to manage code execution or search separately—the model can call them directly as part of a single API request. The new logging uses a content-addressable store (modeled after Git) to avoid duplicate JSON when message sequences are appended across turns, making audit trails leaner for production systems.
What to watch
The llm-anthropic plugin (now at 0.26) adds support for the Claude 5 family, and llm-gemini, llm-openrouter, and llm-mistral updates are coming soon. Existing plugins will continue to work, but model-providing plugins need upgrading to 0.32 to use the new streaming events system. The release notes, including detailed guides for plugin implementation with Structured messages and streaming events, are published on the project's documentation.
Summaries like this, in your inbox every morning.
LLM 0.32 represents a shift in how the toolkit models the developer experience around language models. The earlier abstraction—where users created a conversation and sent messages one at a time—hid the true nature of LLMs: each request carries a complete message history. As models evolved to return mixed response types (reasoning, text, tool calls, images), that abstraction became a bottleneck. The new model.prompt(messages=[]) parameter and stream_events() method expose this reality directly, giving developers fine-grained control over how they handle each chunk of output.
The integration of server-side tools is equally significant. Rather than forcing developers to manage code execution or search as separate concerns, tools like OpenAI's CodeInterpreter and Anthropic's MCP integration let models call them directly within a single API roundtrip. For developers building systems that loop tools to achieve goals—what Willison now calls "agents"—this eliminates orchestration overhead.
The content-addressable logging system deserves attention for production use. By storing only unique message fragments and referencing them by hash (like Git), the new schema drastically reduces duplicate JSON when messages are appended across turns. This matters for systems like Datasette Agent that maintain long conversation histories and need auditable logs.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
In Q2, billionaires or their funds sold six 'Magnificent Seven' AI stocks
Nvidia was rising early Friday, even as the wider chip sector kept falling, Barron's reported
ElevenLabs completed a $300 million employee tender at a $22 billion valuation, and co-founder Mati Staniszews…
On October 15, OpenAI announced the Rosalind Academic Plan, giving approved US university labs free access to…
Neostellar Capital principal Evan Schlossman said CoreWeave's $103.7 billion in remaining performance obligati…
A Qiita article compared automated red-teaming methods for LLM apps — GCG, PAIR, and TAP — using mock implemen…
