GPT-6 Astra is now GA in Copilot, and OpenAI is publishing its own homework
Issue #032 · 4 min read
GPT-6 Astra is now GA in Copilot, and OpenAI is publishing its own homework
Plus a coding agent that merges PRs almost as well as a human, and one dev's 5,000-line unreviewed PR problem.
The big story
Matt Clifford is quitting his job running ARIA, the UK's advanced research agency, before he even starts a role at Anthropic.
He's leaving now, specifically so the conflict of interest never has to be explained later. That's the story here — not a scandal that happened, but one that got stopped before it could.
MPs are relieved. They're also still asking how someone ended up holding a government research post and an Anthropic role at the same time in the first place — that's The Register.
Nobody has alleged wrongdoing. That's almost the point.
The people setting national AI strategy and the people building the frontier models are increasingly the same small pool of people, moving between jobs faster than the ethics paperwork can catch up.
If you sell into government or work anywhere near AI policy, this is worth watching. The revolving door only has to spin once for the story to become "why didn't anyone notice."
What shipped
GPT-6 Astra is now generally available inside GitHub Copilot, across Copilot Chat, inline edits, and agent mode. GitHub's pitch is that it's built for "long-horizon, autonomous coding and agentic tasks" — the kind of multi-step work you'd previously have had to babysit. It's also live on Vercel's AI Gateway as of the same week, and OpenAI is publishing its own case studies for it: Legora used it to review 41 documents in minutes and catch all four planted errors, and Playco says it cut manual fixes in game prototyping by 50%. Worth remembering these are OpenAI's own selected examples, not independent benchmarks — read them as "here's what good looks like," not "here's what you'll get." Details at GitHub's changelog.
A new benchmark puts numbers on the coding-agent gap that's usually just vibes. A paper on arXiv found PR merge rates of 84% for Claude, 74% for Codex, and 43% for Devin — against 85% for humans. Claude is now essentially matching human merge rates on real pull requests. That's a genuinely different claim than "the model writes plausible code," and it's the kind of number worth pulling into any build-vs-buy conversation about agentic coding tools this quarter — see the paper.
GitHub is also deprecating a batch of Copilot models on October 2nd, so if your team has pinned a specific model in a workflow or eval harness, this is your prompt to check which ones — full list here. Not exciting. Still the kind of changelog line that breaks someone's Tuesday if nobody reads it in advance.
What I'd actually do this week
Pull up the GitHub Copilot deprecation list and check it against anything your team has hardcoded — CI configs, eval scripts, internal tooling — before October 2nd does it for you.
If you're evaluating coding agents for your team, use the 84/74/43 merge-rate numbers as a starting filter, not a final answer — then run your own PRs through whichever one clears the bar, because your codebase's edge cases won't match a benchmark's.
If anyone at your company sits on both a government advisory panel and a vendor's payroll, ask now how that's disclosed — Clifford's exit shows the cost of waiting until someone else asks first.
Reply and tell me what you're seeing — especially if you've actually run GPT-6 Astra on something real yet.