GPT-6.2 stuck in compaction loop in Codex, burning 4.44M tokens in 9 minutes
Dear_Lock_5280 · reddit · 2026-11-18
A paying user documented a severe GPT-6.2 Sol Max bug in Codex on Reddit: after switching from 6.1 in an existing conversation, the model compacted its context after nearly every action.
Logs showed a single 9-minute turn processed 4.44M tokens, 2.06M from compaction alone, with six completed compactions and a seventh interrupted before any final response. The root cause appears to be the Codex client warning Unknown model gpt-6.2-sol, falling back to 258K context metadata instead of 475K and repeatedly triggering unnecessary compaction. The turn consumed half of the user's allowance; reverting to 6.1 fixed it.
The author also reports an unresolved October 31 case where another Codex issue burned 40% of their allowance, with support unresponsive for over two weeks.
More from coding & agent
- Claude Code says agents write 90% of its code — Yoav Goldman offers a skeptical take — PMinervini · 2026-11-18
- Claude delegates synthetic data generation to local Qwen3.8-27B, citing its own guardrails — maksym_andr · 2026-11-18
- GitHub Copilot app adds Azure DevOps repos and browser automation in public preview — DanWahlin · 2026-11-18
- Claude Code's context compaction silently turns offhand remarks into agent-constraining rules — nptacek · 2026-11-18
- Shipyard MCP server automates domains, DNS, DB and Stripe setup so AI-built apps launch in an hour — OnlyBullish1 · 2026-11-18
- Infinitty Previews Drag-and-Drop File Transfers and Auto Secure Tunnels for Remote Hosts — jasonkneen · 2026-11-18