Simon Willison shipped LLM 0.32 this week with a complete database redesign that finally handles conversations the way modern chat APIs work. If you’ve been using LLM to talk to language models from the command line, this is a big upgrade.
The headline change is content-addressable message storage. Instead of linear chat logs, LLM now stores individual messages with hash-based IDs and can represent forked conversations as trees. That means you can branch off from any point in a conversation without duplicating the entire history.
More importantly, it means LLM can now handle OpenAI-style chat completions properly. You know how every API call to ChatGPT includes the full message history? LLM can now deduplicate that automatically instead of storing the same messages over and over.
Willison also released a companion tool called llm-chat-completions-server that exposes local models through an OpenAI-compatible endpoint. Point it at a local model running in LLM and you get a /v1/chat/completions endpoint that works with any client expecting OpenAI’s API. Useful if you’re testing code against different providers or want to swap in a local model without changing your application code.
The new schema is a breaking change, so Willison released it as 0.32rc1 and rc2 first. RC2 fixed some dependency issues and switched the default model from GPT-4o mini to GPT-5.6 Luna. Luna costs slightly more ($0.20 input, $1.20 output per million tokens vs. $0.15/$0.60 for 4o mini), but it’s a much better model. You can switch back with llm models default gpt-4o-mini if you want.
Who should care: If you use LLM regularly, this is worth upgrading for. The conversation tree support is particularly useful for experimenting with different prompt variations. If you’ve never used LLM, it’s a command-line tool for running prompts against any LLM provider from a single interface. Works with OpenAI, Anthropic, local models through Ollama, and dozens of others through plugins.
Speaking of Luna, OpenAI cut prices on GPT-5.6 models this week. Luna is now $0.20 per million input tokens (down from $0.30) and $1.20 output (down from $1.80). Terra dropped to $1.25 input and $5.00 output, down from $2.00 and $8.00.
These aren’t huge cuts, but they continue the pattern of OpenAI dropping prices every few months as they get better at running inference. Luna in particular is now cheap enough that it makes sense as a default model for most applications. It’s smarter than GPT-4o and costs less than GPT-4o mini did at launch.
The pricing page also shows batch processing discounts of 50% if you can wait up to 24 hours for results. That puts Luna at $0.10 input and $0.60 output for batch jobs, which is legitimately cheap for that level of capability.
DeepSeek updated their V4 Flash model on July 31st, though the release notes are sparse. According to Artificial Analysis benchmarks, the new version shows improved performance on several benchmarks while maintaining the same pricing.
DeepSeek models are interesting primarily because they’re extremely cheap. V4 Flash runs about $0.30 per million tokens for input and output combined, which is roughly half what most Western providers charge for comparable models. Quality has historically been mixed, but the V3 and V4 generations have been competitive with mid-tier models from OpenAI and Anthropic.
If you’re processing large volumes of text where you don’t need top-tier reasoning, DeepSeek is worth testing. The API is OpenAI-compatible, so swapping it in is usually straightforward.
GitHub published a blog post about stacked sessions and pull requests in the Copilot app. The feature lets you work on multiple related changes in parallel, each in its own session, and then create a stack of dependent pull requests.
This is basically bringing the stacked diffs workflow that some teams use into the Copilot interface. Instead of one giant PR that refactors a bunch of code and adds a feature, you can break it into a series of small PRs that build on each other. Each one is easier to review, and you can merge them independently as they’re approved.
The implementation is specific to the GitHub Copilot desktop app, which is still in preview. If you’re already using that app and you work with stacked PRs, this will probably be useful. If you’re happy with your current workflow, it’s not a reason to switch.
Who can ignore this: Anyone not using the GitHub Copilot desktop app, which is most developers. The feature might eventually make it into other interfaces, but for now it’s desktop-app only.
One email at dawn. The five stories that mattered, with the bits removed and the meaning kept. Free, for now.