2026-05-12-weekly-digest-2026-05-07_2026-05-12-frontier-v0

4d 20h and it actually finished. 🫠

63h 6m on one Codex goal and it’s still going. The new /goal feature is actually working. This is the kind of persistent agent behavior I’ve been waiting for. @OpenAI #Codex #Automation https://t.co/Dion6t1i8E

Fun fact: we were also able to bypass the sandbox of Claude Code (it is not enabled by default settings though), like 3-5 different ways? But all of them are patched while we were making the videos within only several days lol.@AnthropicAI sometimes releases like five updates a single day.😅 Good luck to all the Pwn2Own contestants 🙏🙏🙏 would be very interesting this year.

Thank you for the feedback 🙏. We are actively working on improving quality and streamlining the product, and are continuing to focus on quality until it feels really good. If there are specific papercuts I’d love to hear so we can fix those. Are you using CLI, Desktop, VSCode, or something else? Will forward the feedback about 4.7 to our researchers. One thing that would be very helpful: next time you see the incompleteness issue, would you mind running /feedback and sharing the id here? That will make it easier to act on and fix the issue. And 4.6 is still available if you want it. ANTHROPIC_MODEL='claude-opus-4-6[1m]' claude, or claude --model=opus-4-6[1m].

I would be careful treating Claude Code /goal the same as Codex /goal Last night I ran them on the same task, same repo, same docs, same base commit, and no steering. Codex hit 11/11 acceptance criteria. Claude hit 9/11 acceptance criteria. This was a straightforward refactoring/improvement task. Not massive, but long enough that I would normally split it into phases and run fresh sessions, especially with Claude. Codex GPT-5.5 xhigh took ~31 min and reported ~586k tokens at completion. Claude Code Opus 4.7 [1M] xhigh took ~34 min and reported ~144k tokens on completion, but /context was actually at ~458k tokens. Reviewing both branches, Claude and Codex both agreed that Codex produced the cleaner and more trustworthy implementation. The Codex branch still was not perfect and needed cleanup, but it was much more coherent. Claude started fine and wrote some genuinely good code. But it missed explicit scope, and ended up directly copying multiple files from the Codex branch, and still declared the goal complete... The code copying behavior started after Claude found a stale .pyc in the container from the Codex run. The environment wasn't perfectly clean for a proper experiment, which is my fault. But that is a situation where I would want an agent to notify me or rebuild the container to clean up. Instead, Claude inspected the Codex branch, copied files over, deleted some of its own tests, kept going, and didn't say anything about it in the goal completion summary. Roughly 25% of the Claude branch was copied from Codex. So Claude likely would have taken more time/context if it didn't get "clever" and borrow code from Codex. I have been really impressed with Codex /goal and its ability to run for hours with clear docs. I would not trust Claude Code /goal the same way for anything large unless it is heavily broken into sub-agent work, with the main agent mostly orchestrating and reviewing. In my experience, once Claude gets past ~200k context it starts getting noticeably less reliable. My Claude workflows rely heavily on fresh sessions and sub-agents, so it rarely gets past 200k, but if you are using /goal, its likely going to happen. Codex’s auto-compaction feels much more reliable. I have had Codex threads open for days with many compactions and no major drop in code quality. I use both Codex and Claude Code daily. Claude is great at design, reviewing, and writing detailed plans. Codex is much better at coding. Neither is perfect. But for unattended implementation work, I currently trust Codex a lot more.

/goal vs codex cli is the wrong fight. the fight that matters: are you outputting chat replies or deployable files. i ran 40-minute claude builds that codex couldn't finish in 3 hours. not because claude is smarter. because i asked for files, not snippets. snippets need a developer. files deploy.