OpenAI adds banked reset card for Codex paid users

• Codex and ChatGPT Work hit 7M active users, and OpenAI added one banked reset to accounts after thsottiaux's milestone post announced the threshold.
• The reset moved from a forced-reset ritual toward a user-controlled card, with testingcatalog's usage screenshot showing web and mobile reset UI and testingcatalog's follow-up showing paid users with an extra banked reset.
• The burn-rate patch had four parts: inference savings, a 372K to 272K product context rollback, reverted reasoning-effort experiments, and multi-agent plus auto-review fixes, all listed in thsottiaux's efficiency update.
• ChatGPT Work uses the same agentic allowance as Codex, while normal Chat does not, according to dkundel's limit clarification.
• Ultra was the sharp edge: theo's breakdown tied the burn to long-context billing, Ultra subagents, and v2 subagent context copying, while reach_vb's reply said Sol uses subagents in Ultra unless explicitly prompted on other modes.

OpenAI's release notes define ChatGPT Work as an agent for longer tasks across connected apps and files. The ChatGPT Work and Codex help page splits the product into Chat, Work, and Codex. A June OpenAI Developer Community post says banked resets began as savable rate-limit resets for Go, Plus, Pro, and Business users. OpenAI's GPT-5.6 Sol model page lists a 1,050,000-token context window, while the live product temporarily rolled Codex back to 272K.

OpenAI framed the reset as a celebration: 7M active users now using Codex and ChatGPT Work, plus one banked reset that replenishes weekly usage when applied.

latentspacepod put the spike beside an older Claude Code number and claimed Codex may have added roughly 1M active users in about a day, while thsottiaux's reply added one caveat: token growth was steeper than user growth.

Simon Willison's Fable bump note tied the adoption curve to Anthropic extending Claude Fable 5 access through July 19. Willison wrote that OpenAI appeared confident it could keep GPT-5.6 available without the same subscription uncertainty.

The immediate operator-facing change was simpler than the pricing discourse: the 5-hour Codex limit stopped applying temporarily, and users reported full rate resets.

OpenAI also said GPT-5.6 Sol was available on all paid plans with 100% rate limits while the model remained available, per reach_vb's availability post.

The reset cadence changed because banked reset UX was not fully rolled out. thsottiaux's explanation said OpenAI applied resets automatically because web and mobile could not yet use banked resets, and because full resets produced more predictable infrastructure load during the first 48 hours.

OpenAI's patch list reads like an incident note hidden inside a celebratory reset thread:

• Inference optimizations, expected to produce around 10% more usage.
• GPT-5.6 Sol context limit in Codex reverted from 372K to 272K after extra usage was charged unintentionally.
• Reasoning-effort experiments, called “juice values” under the hood, reverted.
• Multi-agent usage in high and xhigh reduced going forward.
• Auto-review efficiency fixes.
• The 5-hour limit kept temporarily suspended.

The context issue had an unusually public correction cycle. pvncher's earlier post called the 272K higher-rate claim misinformation for Codex's expanded context window, while aibuilderclub_'s follow-up said thsottiaux later confirmed the 372K product change caused more usage to be charged than intended and was reverted.

A separate ablation thread tested the harness side of the same problem. mattlam_'s OpenBench table found trimmed Codex variants with similar or better correctness, shorter wall-clock time, and lower token use.

The reported failure mode had three stacked multipliers:

• Long threads crossed the product's intended context behavior.
• Ultra spawned subagents with Ultra behavior.
• The v2 subagent layer copied the full long context, unlike v1 fresh-history subagents, according to theo.
• Fast mode added another multiplier in theo's account.

OpenAI-side replies narrowed the intended design. pvncher's Ultra reply said Ultra uses max reasoning, and pvncher's Max reply said Max was disabled by default because it is slow and token-consuming without subagents.

The disputed “juice value” claims were not just vibes. Lentils80's post alleged Sol thinking budgets had been degraded versus release day; thsottiaux's update later said reasoning-effort experiments were reverted, without publishing the old and new values.

ChatGPT Work is the same quota story wearing a more general-purpose UI. dkundel said Work and Codex share a limit because they are the same agent; ChatGPT in Excel or PowerPoint shares it too, while normal Chat has separate limits.

OpenAI's release notes say Work can research, analyze, work across connected apps and files, create documents, spreadsheets, presentations, reports, and Sites, and run Scheduled Tasks. On web and mobile, the rollout covers paid plans except Free and Go, with Pro, Pro Lite, Enterprise, and Edu first.

reach_vb gave the non-coding demo: his Gmail example had Work review the 500 most recent emails and produce an unsubscribe list in 46 seconds.

The quality reports explain why the reset card became a story at all. Users were burning limits because they kept finding tasks worth handing to Sol.

altryne used Codex and Sol for Chrome shortcut remapping, 1Password diagnosis, a custom macOS window mover, cross-machine session work, Raycast scripting, disk cleanup, and Home Assistant maintenance. One thread kept going after the visible 5-hour credits were exhausted, according to altryne's limit note.

kimmonismus called GPT-5.6 better overall, especially for frontend work and longer agentic runs, but said the token burn made Pro usage difficult. zeeg reached a harsher economic split: zeeg's comparison preferred Fable's outputs in limited runs but still favored Sol's cost and speed tradeoff over paying out of pocket.

External eval chatter pointed the same direction. danielhanchen's Margin Lab post cited GPT-5.6-Sol xhigh at roughly 82% on a 50-question SWE Bench Pro sample, up from roughly 54% for 5.5 xhigh.

The reset system still had edges after the milestone. OSS program users reported banked resets not working, and reach_vb said OSS program users did not get resets yet while OpenAI worked on support.

Gifted or granted subscriptions were another exception. reach_vb told one user that paid subscriptions were eligible, but a small percentage of gifted or granted subs were not.

The GitHub tracker shows separate limit-state bugs. One open issue says a reset-available UI turned into 100% 5-hour usage within about five minutes, with the first local rate_limits reading already at 79%. Another open issue says a Plus account showed 83% weekly limit remaining while every prompt was blocked as if the account were Free.