<blockquote class="twitter-tweet"><p lang="en" dir="ltr">Update on rate limits in Codex. <br><br>We’ve found (a) some inefficiencies when using images in long sessions with multiple compactions (b) high p95+ usage for Computer History (c) a feature that was meant to generate conversation titles that was draining a bit more usage than… <a href="https://t.co/pDC9U0d3ch">https://t.co/pDC9U0d3ch</a></p>— Tibo (@thsottiaux) <a href="https://t.co/xk2PEM7g0B">August 23, 2026</a></blockquote> <script async src="https://t.co/b22siy1VPw" charset="utf-8"></script>
Update on rate limits in Codex.
We’ve found (a) some inefficiencies when using images in long sessions with multiple compactions (b) high p95+ usage for Computer History (c) a feature that was meant to generate conversation titles that was draining a bit more usage than intended. And we have a tiger team combing through everything and shipping fixes tomorrow. We also found a novel approach to drive efficiency up significantly that is completely unrelated and we will be working on next week.
As part of some of the fixes tomorrow, we will also do a full reset of the usage for all paid subscriptions.
See you then.
@thsottiaux
I use opus4.8 to summarize the promblem: ════════════════════════════════════════════════════════════════ Codex Desktop (Windows / MSIX) — large existing sessions white-screen, truncate history, and crash the renderer after the 26.707.x update ════════════════════════════════════════════════════════════════ TL;DR Since updating from 26.609.3341.0 to the 26.707.x line, opening any large existing session (hundreds of MB to 2.35 GB) shows a multi-second white screen, then renders only ~<10 messages of scrollback, and frequently crashes the renderer. The exact same session files opened instantly with full history on 26.609. The crash reproduces with 26 GB of free system RAM, so this is a client regression in the session load/render path — not user data size and not memory pressure. ENVIRONMENT - OS: Windows 11 (build 10.0.26200) - App: OpenAI.Codex (MSIX, Microsoft Store) - Last-known-good: 26.609.3341.0 (sessions opened instantly, full history) - Regressed line: 26.707.x 26.707.3748.0 first installed 2026-07-11 14:29 → .6957 (07-13) → .8168 (07-13) → .8479 (currently installed) → .9564 attempted 07-14 (has not successfully registered) - Renderer: ChatGPT.exe, Chromium 150.0.7871.101 - Session store: %LOCALAPPDATA%\...\.codex\sessions (rollout .jsonl per session) REPRODUCTION 1. Have a large, image-heavy session on disk (created on an earlier version). 2. Open it in the desktop client. → long white screen, then only the tail (~<10 messages) is scrollable; older history never loads; renderer often crashes. On 26.609.x the same file opened instantly with complete history. EVIDENCE IT IS A CLIENT REGRESSION (not the user's data or RAM) 1) Crashes with plenty of free RAM — not a system OOM - Renderer (ChatGPT.exe) crashed while opening a large session with 26.1 GB free system RAM (system at 59% used). - No Windows "Application Error"/OOM event is logged — the renderer exits cleanly. Consistent with the renderer exceeding its OWN per-process V8/Chromium heap limit while parsing the full multi-GB rollout into memory, which is independent of available system RAM. - Windows also logged RADAR_PRE_LEAK_64 for ChatGPT.exe 150.0.7871.101 on 2026-07-12 (excessive memory growth). 2) File sizes PREDATE the update — size didn't change, the client did - Largest sessions and their last-modified dates (all before 26.707 landed on 2026-07-11): 2352 MB last modified 2026-06-22 735 MB last modified 2026-06-30 576 MB last modified 2026-06-19 - Same size, fast on 26.609, slow/crashing on 26.707. 3) The client rewrites the entire rollout on open - Opening a session updates its rollout .jsonl mtime to the exact moment of opening, with NO new user turns (observed: a 365 MB session rewritten at 15:37:21 on open). This suggests the client re-serializes the whole rollout on open, which both stalls the load (white screen) and drives the memory blow-up / crash. SESSION CORPUS (for context on scale) - 302 sessions, 11 GB total. - Size distribution: 2 files > 1 GB, 5 in 500 MB–1 GB, 13 in 100��500 MB, 282 < 100 MB. - Large sessions are image-heavy. Measured on a 576 MB session (2341 lines): 129 lines contain inline base64 PNG screenshots and account for 58.7% of the bytes (~12 MB per image record); the remaining ~41% (238 MB) is text/tool output. - Usage: these months-old sessions were reused daily for ~1 week with no issue until the 26.707 update. SUSPECTED ROOT CAUSE / FIX AREAS - Session load appears to parse (and re-serialize) the ENTIRE rollout .jsonl into the renderer on open. For multi-GB, image-heavy rollouts this blows the renderer heap and stalls the UI. - Suggested: lazy/windowed loading (read only the tail needed to render, stream older turns on scroll), stop rewriting the full rollout on open, and virtualize scrollback so the whole history need not be resident in the renderer. IMPACT Users who rely on long-lived, image-heavy sessions can no longer open them: history is effectively inaccessible and the app crashes under the load. Regressed between 26.609.3341.0 and 26.707.x. ════════════════���═══════════════════════════════════════════════