OpenAI fixed GPT-5.6 Sol’s most frustrating flaw: Burning limits while it waits
The New Stack Amanda Caswell ● Covered by 14 sources
OpenAI just fixed GPT-5.6 Sol eating through usage limits even while it sat idle waiting on tools. Turns out chat-based subscription caps just don't work for agents that run 40-minute coding sessions.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI shipped GPT-5.6 Sol earlier this month, pitched as a model for tougher coding work, and within weeks power users were furious that their Codex and ChatGPT Work limits were vanishing far faster than expected. The company's response: reset everyone's usage limits and push backend improvements that should stretch a typical Sol session about 18% longer.
The root cause, according to engineering lead Thibault Sottiaux, was Sol's new programmatic tool calling, sometimes called "code mode." The model got much better at chaining tool calls, coordinating subagents, and grinding through complex workflows for extended periods. That's exactly the capability OpenAI wanted, but it also meant Sol was burning tokens at a rate the company never modeled. Sottiaux admitted OpenAI leaned too heavily on average and median usage numbers before launch and missed how much the long tail — the power users pushing hardest — would actually consume.
The GitHub repo for openai/codex tells the real story better than any blog post. One developer logged a 43-minute session that produced nearly 300 model responses and 192 wait calls, chewing through 42% of a five-hour allowance mostly while doing nothing but waiting on tool results. Another noticed tasks that looked parallel were actually running sequentially, stretching runtime across more than 700 execution cells and multiplying token use in the process. This is a familiar story for anyone who's shipped software: systems behave fine in controlled testing, then real users find the edge cases nobody planned for. Agentic coding tools make this worse because so much of the work — and the waste — happens invisibly, in the gaps between tool calls.
OpenAI's immediate patch targeted exactly that gap: making Sol more efficient while it waits, tightening up web search behavior, and temporarily suspending the five-hour cap while the fixes rolled out. Useful, but it doesn't solve the deeper mismatch. Subscription tiers built for back-and-forth chat sessions were never designed for a model that disappears into a 40-minute task, making dozens of tool calls and revisions before it resurfaces with an answer. Until pricing and limits catch up to how agents actually work, developers are going to keep getting blindsided by their own usage meter.
My take — AI-written commentary, not fact-checked reporting
This is what happens when companies price agentic AI like it's still chatbot small talk — the usage model breaks the moment the product actually gets good at its job. OpenAI deserves some credit for admitting the miss publicly rather than quietly throttling people, but the bigger lesson is that every AI lab selling "agents" needs pricing built around task complexity, not token-per-message averages, or this exact blowup repeats with every capability bump.
Read more about this at: The New Stack