We’re demonstrating how frontier models have continued to improve in realistic mental health conversations with MentalHealthBench. This new open benchmark was built with input from more than 80 mental health clinicians. We’re releasing it openly so other researchers can examine the methods, run their own evaluations, and build on the work.

Fun plot of uv performance over the last 28 releases. Resolution and installation (on real, complex examples) is >2x faster now. Software should get faster, not slower, over time!

In preview, uv will now omit package metadata from the lockfile. (This is separate from the resolved dependency graph, which we retain.) Internally, this reduced lockfile size by up to 50% (more like 20% on average). Fewer conflicts too.

gpt-6-luna (flex, low thinking) is phenomenal for text extraction. Half the cost of the next cheapest and just as accurate as gpt-5.6-luna. I switched from gemini-2.5-flash-lite when I started getting refusals for text extraction (?). Much better quality with luna in general.

The best thing about GPT-6-Sol is efficiency. In my testing the difference between GPT-5.6-Sol and GPT-6-Sol is the new version consumes 1/2 the tokens and takes 1/5 of the time to complete the task. On top of the 50% cost reduction, it should be quite a nice daily driver

GPT-6 Sol and Luna are out, and they are better AND 50% cheaper than 5.6. Luna is now $0.10 input / $0.50 output per 1M tokens. This is on top of the 80% price cut to Luna we made at the end of July. Output went from $6 -> $0.50 within two months.

BREAKING: @OpenAI just dropped GPT-6 Sol. It’s my new daily driver in Codex: not quite Astra, but close enough for much of my everyday work, faster, and 50% cheaper than 5.6 Sol. We tested it across the work we actually do at @every and ran it head to head versus Opus 5.5. Here’s my vibe check: • Writing: On a paragraph-writing task drawn from my real work, Sol scored close to Astra, which is still my top model for writing. It writes clean, minimal prose and puts the important idea first. Opus 5.5 is pleasant to work with, but its drafts still tend to bury the point. • Computer use: If you love Astra’s computer use, you’ll like Sol. An earlier Sol preview matched Astra on 17 of 18 attempts across six of our simpler Hands tasks, at a much lower token price. • Coding: Sol improves on GPT-5.6—including better Ruby code in @kieranklaassen tests—but Opus 5.5 has the higher ceiling for long, autonomous builds. One frustration: Codex’s new security classifier repeatedly stopped work we’d already authorized to ask for approval. That’s friction from the classifier, not necessarily a limitation of Sol, but it made long runs harder to leave alone. Overall: Sol feels like an S-class iPhone release: It will give you much of Astra’s power at about a fifth of Astra’s price. Opus 5.5 is the bigger surprise. In some of our tests, it matches or even beats Fable 5.1 on many tasks while also being cheaper than Opus 5. A few people on our team who were Codex converts have started to wobble with Opus 5.5. If you’re already in the Claude ecosystem you’re going to love this model. If you’re a Codex user, it’s worth a look especially for your top-end coding tasks. You’ll like Sol 6 if you spend your day in Codex reading, writing, and getting things done. It’s fast, much cheaper than Astra, and it’s the model I keep reaching for. You’ll like Opus 5.5 if you want to hand an agent a hard coding or visual project and see how far it can take it. Its best work went further than Sol’s in our tests—enough to pull some of our team back toward Claude. You can go deeper on all of our evals including each real-world task and how they're scored at the links below: Writing: Knowledge work: Reading: Full vibe checks on @every coming soon!

Especially compared by per-task pricing, which is the metric that should matter, I don't think there is anything competitive anywhere in the market. We want people to be able to use tons of AI; it is important to being able to explore this new renaissance in front of us.

We have been focusing on efficiency and intelligence for all. Very proud of the team. Only possible when you have incredible models at the top end of the capability that you can then use to make a big difference in everything else.

One of my favorite pastime: I closed like ~30 ty issues yesterday by asking Codex to identify stale issues (i.e., issues that were already fixed, deliberately or inadvertently, on main)

GitHub Actions + Rust's Miri can leak your secrets in CI 🧵 If you run Miri in CI: - upgrade to the 2026-09-22 nightly - clear caches - rotate any secrets that the `cargo miri` CI job had access to GPT-6 Astra helped find this 🦀