Codex Web GPT 2.0.0 is a major reliability release for long-running Codex tasks through ChatGPT Web.
Highlights:
- Improved streaming fidelity for reasoning, intermediate commentary, actions, and final Markdown. Partial labels and truncated final answers are no longer treated as complete output.
- Full context compaction support, including automatic compaction in the middle of an active tool loop. The old browser turn is retired, compaction runs in an isolated turn, and Codex continues in a fresh bound turn with the compacted history and tool results preserved.
- Large prompts are inserted through bounded browser edits and verified exactly instead of being rejected by an artificial character limit.
- Repeated Codex model-switch contracts and non-semantic one-pixel image placeholders are removed from the transported context.
- Current multi-environment Codex project metadata is supported, including one authoritative primary workspace.
- Browser sessions, retries, tab ownership, and concurrent turns are more stable.
- The launcher now checks GitHub once at startup and offers a user-triggered, checksum-verified in-place update on macOS, Windows, and Linux.
Context windows:
- ChatGPT Web Instant and Medium: 150,000 tokens.
- ChatGPT Web High: 185,000 tokens.
- ChatGPT Web Extra High: 256,000 tokens.
- ChatGPT Web Pro: 272,000 tokens.
- Automatic compaction keeps a 10% reserve before each hard limit.
Validation:
- Tested on macOS with real multi-turn chats and context compaction on both ChatGPT Plus and ChatGPT Pro accounts.
- The release matrix passed full verification, native packaging, and packaged-launch smoke tests on macOS ARM64, macOS x64, Windows x64, and Linux x64.
- Every published asset is covered by checksums.txt.
Release media will be added separately.