# Sol limits reset after usage drain > OpenAI reset ChatGPT Work and Codex limits, says typical GPT-5.6 Sol use should last about 18% longer, and plans to restore the paused five-hour limit. _Source: Tibo (@thsottiaux), Codex & ChatGPT at OpenAI · 2026-07-29 · 7 min read · Verified against primary sources_ Canonical: https://iyu.app/e/gpt-5-6-sol-usage-limits ## The 60-second version OpenAI reset ChatGPT Work and Codex limits after GPT-5.6 Sol consumed some users' allowances faster than intended, and says typical use should now last about 18% longer. **Key points** - OpenAI says subscription-plan allowances were not reduced. - The temporarily paused five-hour limit is scheduled to return after the investigation. - Sol tends to work longer, make more tool calls and coordinate more subagents on difficult tasks. - Code mode sometimes created extra responses and cached input while tools were pending or web searches multiplied. - Median efficiency looked acceptable, but demanding long-tail workflows were much more expensive. **Verdict.** The reset provides immediate relief and the fixes should help typical use, but the 18% figure is an OpenAI estimate. Power users should benchmark their own tool-heavy workflows again. ## Full explainer > **i** This is a first-party operational update from Tibo, whose profile identifies him with Codex and ChatGPT at OpenAI. The reset and policy announcement are verified as statements OpenAI made; the 18% improvement is OpenAI's expectation, not an independent measurement. ### What changed — Limits were reset and the five-hour window is returning OpenAI has reset usage limits for ChatGPT Work and Codex users after complaints that GPT-5.6 Sol was draining allowances faster than expected. The company says it did **not reduce subscription-plan allowances**. It also says the five-hour limit that was temporarily paused during the investigation will be restored the following day. - **18%** — Expected longer typical Sol usage, according to OpenAI - **5 hours** — Temporarily paused limit window scheduled to return - **All** — ChatGPT Work and Codex users included in the reset > **⚑ Caveat:** The approximately 18% figure is a projected typical-use improvement from OpenAI. Individual results can differ substantially by task complexity, reasoning setting, tool latency, search volume and subagent use. ### Root cause — Sol often keeps working longer than earlier models OpenAI's explanation is not that the quota became smaller. Sol is more willing to continue a difficult task, call additional tools and coordinate workflows across tools and subagents. That persistence can improve hard-task performance, but it can also consume more of a user's allowance per turn. The same named reasoning effort is not a fixed token budget across model generations. OpenAI says `High` on Sol can use more tokens than `High` on GPT-5.5 because Sol does more work at that setting. - **Longer autonomous runs:** Sol may continue planning, checking and calling tools instead of stopping early. - **More tool coordination:** Complex workflows can involve additional tools and subagents, increasing work per turn. - **Reasoning setting:** The same label, such as High, can represent more token use on Sol than on GPT-5.5. - **Uneven impact:** Median users were relatively efficient, while some power users in the long tail consumed much more. ### Code mode — Parallel tools created hidden usage overhead Programmatic tool calling, also called code mode, lets Sol launch tools in parallel and continue while other work is pending. OpenAI says this flexibility sometimes produced more responses per turn, more cached input tokens and higher usage than intended. Waiting for slow tool calls and running many web searches were especially costly. OpenAI says it has improved both cases and is continuing to optimize code mode. The key distinction is that parallel execution can reduce elapsed time while still increasing the amount of model work counted against a limit. > Faster wall-clock completion and lower quota consumption are different optimization goals. ### Why it was missed — The average concealed the power-user tail OpenAI says the median user found Sol reasonably token-efficient. The problem concentrated among people running harder, longer and more tool-heavy jobs. Pre-launch analysis focused too much on average and median behavior, so expensive long-tail workflows were underrepresented. That explains why reports appeared contradictory: both statements can be true. A typical user can see good efficiency while a smaller group of intensive users exhausts the same allowance unusually quickly. ### What users should do — Retest the same workflows after the reset - **1.** Retry the same representative Sol tasks after the reset; compare completed work, not just turn count. - **2.** Track tool-heavy jobs separately from plain chat, especially web-search loops and slow external tools. - **3.** Do not assume identical reasoning labels consume identical amounts across GPT-5.5 and Sol. - **4.** Treat the 18% figure as a typical-use estimate and measure your own long-running workload. The practical takeaway is straightforward: allowances were reset, the five-hour window is returning, and several efficiency fixes are live. Users should expect improvement, but demanding agent workflows may still sit far above the median and need their own measurement. ## Primary sources - [Tibo (@thsottiaux) — GPT-5.6 Sol usage-limit update](https://x.com/thsottiaux/status/2082317452755751098) --- _Published by iyu (https://iyu.app) — the day's AI news, checked against primary sources and rewritten in plain language. Free to quote with attribution and a link to the canonical URL._