OpenAI reset usage limits after GPT-5.6 Sol started draining some ChatGPT Work and Codex allowances faster than users expected, but the fix exposes a bigger pricing problem for agentic tools.
OpenAI's Sol launch ran into the kind of problem you only see when a powerful model meets real users at scale. The model was made broadly available in ChatGPT on July 9, according to OpenAI's own release notes, and within days the company was resetting usage counters and lifting the five-hour restriction for paid plans.
That is not small.
According to reporting from BleepingComputer and PC Watch, OpenAI product lead Thibault Sottiaux posted on X on July 12 that the last 48 hours for Codex and ChatGPT Work had been intense. He said OpenAI was temporarily removing the five-hour usage limit for Plus, Business and Pro plans, rolling out efficiency changes for GPT-5.6 Sol, and resetting usage after the products reached 6 million active users.
The important part is not that OpenAI gave users more room. The important part is why it had to. Sol is a more ambitious model than GPT-5.5. It is built for coding, research, science, cybersecurity, computer use and design, according to OpenAI's release notes. It can run longer jobs, use tools, and push through work that older assistants would more often stop and hand back to you.
That is exactly what developers want. It is also exactly what burns allowance.
The complaints were not just ordinary launch grumbling. Public GitHub issues in the openai/codex repository described high-frequency waiting behavior, serial tool execution and very large resumed contexts driving heavy token processing during Sol sessions. One issue described a 43-minute interval with 292 model responses, 96 exec calls and 192 wait calls, with the five-hour meter moving from 58% to 100%. Another said a long GPT-5.6 trace used parallel execution in only 5 of 739 exec cells, even though the work appeared to include independent operations.
You do not need to know every implementation detail to understand the user problem. If a coding agent keeps checking on tools, rereads a large context, and spawns more model turns than expected, the allowance can disappear while it feels as if the task is merely waiting.
The reset fixed the symptom #
OpenAI's first move was blunt and useful. It reset usage. It also d the five-hour cap for Plus, Business and Pro subscribers, and later reporting from Neowin said OpenAI had landed inference optimizations that should give users around 10% more usage for GPT-5.6 Sol. PC Watch also reported that OpenAI expanded banked resets to web and mobile after they had initially been available only in the desktop app.
The reset was necessary.
But it was not enough.
OpenAI's own launch framing had already made efficiency part of the Sol pitch. In a CNBC interview on July 9, Sam Altman said GPT-5.6 Sol was 54% more token efficient on agentic coding tasks than other leading models, while also saying cost and speed were central to the release. CNBC's transcript is careful on one point the marketing chatter often blurs: Altman did not give a precise baseline model in that exchange.
That distinction matters when you are paying for usage. A model can be more efficient on a benchmark and still feel expensive in a real Codex session if it runs more steps, holds more context, calls more tools and keeps going long after a chat model would have stopped. Efficiency per unit of intelligence is not the same thing as a predictable allowance for a working developer.
Agentic tools need different limits #
Here is the thing. Subscription limits were designed around a simpler pattern: you ask, the model answers, and the exchange ends. Sol is closer to a worker sitting inside your project. It can continue through tests, code edits and the searches in between. It can also spend tokens while coordinating all of that.
That frame is breaking.
OpenAI has some room to improve the mechanics. Waiting states should not chew through quota unnecessarily. Independent tool calls should run in parallel where they safely can, not march through a long serial queue. Cached input should help users, not leave them wondering why a mostly cached session still feels punishing against their plan.
But the harder question is commercial, not technical. If AI companies want you to use agents for serious work, they cannot price and meter them as if they were ordinary chatbots with a stronger coding badge. A software task that runs for 45 minutes is not the same product as a quick answer about an API. Users understand that more capability costs more. What they hate is discovering the cost only after the meter has jumped.
OpenAI knows this now. Sottiaux's July posts, the usage resets, the temporary cap removal and the banked reset rollout all point in the same direction. The company is learning in public because the heaviest Codex and ChatGPT Work users are doing the same thing.
The next launch has to account for those users first. The median user may be easier to model, but power users are the ones who find the real limits of an agentic coding system. If Sol is going to be sold as the model that keeps working, OpenAI has to make the meter as clear as the capability. That is the real test.
Also read: Epsagon founders raise $34M seed round for Harmony, an AI agent platform built inside Slack and Teams • Washington bans Chinese humanoid robots and power inverters in its boldest tech decoupling move yet • Act Security raises $60M and exits stealth with a platform built to stop AI agents from inheriting dangerous cloud permissions