Codex Ultra struggles with long docs: burns 80% context with little code
lucasmeijer · x · 2026-08-30
A user experimented with the 'Ultra' model in Codex Mac App for a well-specified implementation task that didn't fit into the 'medium' context window. The result was underwhelming: the agent ran through 80% of its context window without seemingly having written much code.
Related event: Codex Mac App Subagents Fall Short in Real-World Tests(2 posts)→
More from coding & agent
- Selecting local coding models for high RAM, limited VRAM setups — Frail_Waif · 2026-08-30
- User Reports Codex Mac App Subagents Not Frugal on Main Thread — lucasmeijer · 2026-08-30
- Maersk Engineering: Translating SOPs and 100k Corrections for Agents — AI Engineer · 2026-08-30
- AI tools eroding code review: team reads only 5% of PRs — Upset-Day9099 · 2026-08-30
- Tesana Launches Loop Mode for Self-Testing 3D Game Generation — Aiden_Tech_Ai · 2026-08-30
- Open Source quackd: Control Bipedal Robot via Natural Language — DynamicWebPaige · 2026-08-30