AI Coding Weekly

Theo calls GPT-6.2 Astra OpenAI's worst model yet

Theo and Ben independently asked GPT-6.2 Astra for a creative new game level and both got a bell-themed water level named "Bellwether". OpenAI's Thomas Sottiaux says the testers were running week-old builds.

The big disagreement

Theo calls GPT-6.2 Astra the worst model OpenAI has shipped

Theo says GPT-6.2 Astra is the worst model OpenAI has ever released, and that he and Ben, prompting separately for new levels in different games, both got bell-themed water levels named "Bellwether". @thsottiaux replied that early-access reviewers tested builds from before a lot of changes landed in the last three or four days.

Theo - t3.gg
@theo
X
It doesn't miss anything. It's SO thorough. But in it's thoroughness, I find it misses the most important thing: the actual point.
Nov 18, 2026 · View on X

Read the full story

Receipts

Malte Ubl: Astra 6.2 beats the previous best by 20 points on security

Malte Ubl says Astra 6.2 scores 20 points better than Sol 6 on cybersecurity tasks and is "entirely in its own class", and that it is worth rerunning security review with it because it will find vulns the previous pass missed. Guillermo Rauch adds that Vercel's DeepsecBench is a stealth benchmark no lab has had access to the codebases for.

Read the full story

Model and agent watch

Theo dates Anthropic's lead back to Fable 5, not Opus 5.5

Theo argues Anthropic has been ahead since Fable 5 and that most people missed the size of the gap because Fable 5 was expensive, pulled after three days, and had no ZDR policy for work use. He says Opus 5.5 brought the same wins to everyone, cheaper and with friendlier sub limits.

Theo - t3.gg
@theo
X
OpenAI has maintained pace with capability, at the cost of coherency and readability.
Nov 18, 2026 · View on X

Read the full story

Built from 606 posts by the engineers, language designers and agent builders we follow on X over 1 days.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.