Sol’s quota drain problem—and why the fix matters
OpenAI’s GPT-5.6 Sol quota limits refer to the usage caps and token allowances that apply to ChatGPT Work and Codex sessions, which were burning down much faster than expected in agentic coding workflows until OpenAI reset them and optimized Sol’s behavior to extend typical quota duration by around 18%. This is not a minor tuning tweak; it is a response to a design flaw in how Sol handled long-running, tool-heavy jobs. Power users quickly noticed that their ChatGPT Work usage limits were evaporating even when jobs spent most of their wall clock time waiting for tools to finish. When a coding agent spends tokens while it waits, the subscription model stops feeling predictable and starts feeling punitive. The fix matters because it restores trust in Sol as an everyday coding companion rather than a luxury tool reserved for rare, high-stakes runs.
What OpenAI changed: resets, kernels, and behavior while waiting
Under pressure from frustrated users, OpenAI reset usage limits for all ChatGPT Work and Codex accounts and updated the explanation of GPT-5.6 Sol quota limits, stressing that no subscription plan quotas were reduced. At the same time, Sol’s serving stack was tuned: inference kernels, API orchestration, and the agent harness behind Codex were optimized for token consumption efficiency. With Sol autonomously rewriting and improving production kernels, end-to-end serving costs fell by 20%, and speculative decoding lifted token-generation efficiency by more than 15%. Crucially, OpenAI also changed how Sol behaves in tool-calling workflows—reducing needless work while waiting for tools or running extensive web searches, and restoring the previously paused five‑hour limit once the changes were in place. In plain English: Sol now does less busywork, and more useful work, per unit of quota.
The real culprit: agentic coding workflows that burn tokens in the background
OpenAI’s own engineering lead admitted they underestimated how expensive real-world agentic execution would be once power users started pushing Sol. Sol is built to "work for longer, make additional tool calls, and coordinate complex workflows across tools and subagents," which makes it stronger on hard coding tasks but dramatically increases token usage. Programmatic tool calling—its so‑called code mode—lets Sol continue thinking and restructuring workflows while tools run in the background. That sounds ideal for complex CI pipelines or multi-repo refactors, but it also means every loop iteration, every cached input token, and every extra response eats into quota. Agents burn tokens while waiting, especially during heavy web search or slow external tools, and those long-tail scenarios were not properly accounted for in pre-release testing. The median user saw decent efficiency; the heavy users saw their limits disappear. That asymmetry is what turned a technical nuance into a public backlash.
An 18% longer runway—and a warning for anyone building agents
OpenAI says token consumption optimization should make typical Sol sessions last about 18% longer, and some users are already seeing even more noticeable improvements in practice. For ordinary ChatGPT Work users, the practical impact is simple: their reset quotas now stretch further, and agentic coding workflows feel more aligned with what they thought they were paying for. According to OpenAI, "GPT-5.6 Sol with maximum reasoning outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens," a reminder that efficiency can be a competitive feature, not only a cost-control measure. Yet the episode is also a warning for anyone building AI agents: subscription limits were designed around conversational chat, not autonomous code runners that happily loop, coordinate subagents, and wait on external tools. If you ignore the long tail of complex workloads, your users will find it—and they will force you to redesign your economics.
Where this leaves Sol—and what users should demand next
OpenAI’s response is, on balance, the right kind of intervention. By resetting usage limits, confirming that quotas were not cut, and optimizing Sol’s kernels and tool-wait behavior, the company has made Sol meaningfully more cost-efficient without neutering its agentic power. ChatGPT Work usage limits now feel less like a trap and more like a fair ceiling for intensive coding sessions. But users should treat this as a starting point, not a victory lap. Any serious agent platform needs transparent accounting of how tokens are spent in complex workflows, clearer controls over how aggressively agents call tools, and better defaults that account for heavy users, not just the median case. Sol’s quota fix shows that OpenAI can address those problems when pushed. The next step is to expect that kind of thinking upfront—before usage limits fail in production.
