If you run Codex or ChatGPT Work for multi-hour agent tasks, the pain point is rarely model IQ — it is reopening the thread. Internal OpenAI benchmarks circulated in mid-August 2026 show a fix: a 741-turn, 231 MB conversation that took 27.62 seconds to load now opens in 1.66 seconds — about 94% faster (roughly 16.6× speedup).
Codex desktop app lead Andrew Ambrosino published the numbers on X; aggregators and TestingCatalog framed it as a rollout aimed at the following week. OpenAI's release notes also document faster long conversations on the web via sectional loading — related problem, separate client.
TL;DR
| Metric | Before | After (benchmark) |
|---|---|---|
| Load time (741 turns, 231 MB) | 27.62 s | 1.66 s |
| Conversation entries fetched | 15,529 | 64 |
| Network requests | 894 | 16 |
| Renderer JS heap growth | (baseline) | ~87.8% less |
| App memory growth | 1030.7 MiB → | 606 MiB (~41% less) |
Scope: Client history load + render — not inference tokens/sec.
Why this matters for agent builders
OpenAI's own usage data (June 2026 research post cited in coverage) says 70%+ of Codex users in May delegated tasks exceeding one hour of human work. Those sessions accumulate:
- Tool stdout and file diffs
- Subagent transcripts (Work orchestration)
- Browser/computer-use screenshots metadata
- Retries and planner notes
Without pagination, every reopen becomes an O(history) tax — the opposite of how persistent Codex mode is supposed to feel.
The benchmark's two telling drops — 15,529 → 64 entries and 894 → 16 requests — imply lazy / windowed hydration of conversation records, not loading the entire DAG on mount.
Desktop vs web vs Linux preview
| Surface | Status (Aug 31, 2026) |
|---|---|
| ChatGPT desktop (macOS/Windows) | Unified app with Chat, Work, Codex — target of 94% benchmark |
| ChatGPT web | Release notes: sectional long-chat loading |
| Linux preview | See ChatGPT desktop Linux preview — verify perf separately |
Compare harness choice in Codex vs Claude Code: Claude keeps more session state local; OpenAI is now optimizing the server-backed history path that made giant threads painful.
What people are asking
Is 94% the same as "90% faster load" in headlines?
Close enough — 94% reduction in elapsed time ≈ 16× faster, sometimes rounded to "90%" in social posts. Use Ambrosino's 27.62s → 1.66s pair when citing precisely.
Will this fix Codex capacity errors?
No. Codex capacity boosts address queueing and rate limits; this fixes client reopen latency only.
Should I still split threads?
Yes. Faster load ≠ unlimited history. For auditability, export artifacts (Work deliverables guide) instead of one immortal chat.
Related on explainx.ai
- ChatGPT Work vs Codex — complete guide
- OpenAI Codex persistent mode
- ChatGPT desktop Linux preview
- Codex vs Claude Code comparison
- ChatGPT Work thread orchestration on mobile
- OpenAI Codex capacity boost
Sources: OpenAI Help Center release notes · RuntimeWire benchmark summary · TestingCatalog Aug 2026 report
Benchmark figures reflect OpenAI internal tests circulated August 2026; your thread size, OS, and client version will differ.