Theo argues coding as we know it is dead, with Claude writing over 90% of his code. He finds OpenAI's Astra near-perfect at code review, and his logs show Gemini 4 Carbon about five times slower than any other model.
Four new videos and seven T3 Code updates. About 33 min to read, 2 h 02 min if you watch everything.
Theo finds OpenAI's Astra cheap and strong on paper ($8 per million input tokens, $40 output) but hates using it. Route it to code review instead, where it scored near-perfect on his evals and never lied.
Skim itSkip to the Pokémon ROM hack and the Outer Wilds planet to judge Astra's game work for yourself.
Theo audited his usage logs and found Gemini 4 Carbon running about five times slower than any other model on comparable work, worse than his first review let on.
Skim itSkim Theo’s timing breakdown to see how much of Carbon’s wait happens before the request gets going.
Theo argues that coding as we know it is dead. Claude already writes more than 90% of his own code, and he expects hand-written code to become a niche artisan craft, like Italian shoemakers.
Briefing is enoughYou can skip the video. The briefing already covers Theo's main points, including his tip to change the prompt and rerun instead of fixing an agent's code by hand.
The desktop app can forward servers running in a remote environment to your machine. You can then open them in your own browser or local tools.
You can add a side chat to the main thread's composer or reference it with @, even while the main agent is busy.
The Issues page can show any mix of projects you choose, and keeps your choice after a restart. A fork now lists both its own issues and its upstream repository's issues.
Antigravity gained several abilities. Auto mode approves edits and safe commands but asks before irreversible actions. You can steer it without stopping its subagents, and use /compact where the session supports it.
Agents can read, open, update, respond to and merge pull requests, and can merge a whole stack of PRs at once. That avoids repeated CI (automated test) runs on each layer of the stack.
A new Issues page shows issues from Linear, GitHub, GitLab, Azure DevOps and Forgejo beside your threads. You can filter by label, assignee and project, save named views that carry across your devices, and hand an issue to an agent, which can carry it through to a merged PR. Agents can now also read, open, update, respond to and merge pull requests on their own, and read and re-run failing CI checks.
You can now reply directly to a specific assistant response, with the quoted part shown as an inset card. Stopped Claude, Codex, OpenCode and Cursor threads can also be resumed without typing a new message.