OpenAI reverts Codex GPT-5.6 Sol context limit to 272k after overcharging

• OpenAI rolled Codex's GPT-5.6 Sol context limit back from 372k to 272k after thsottiaux's update said the higher product limit caused more usage to be charged than intended.
• The five-hour cap was temporarily lifted, and some users saw fresh quotas after reach_vb said the limit was removed and kimmonismus reported a full reset.
• The burn came from stacked mechanics: theo's analysis pointed to long-context charging, Ultra spawning Ultra subagents, v2 subagents copying long context, and fast mode adding another multiplier.
• The launch framing collided with usage reports: OpenAI's launch post sold GPT-5.6 as more efficient, while kimmonismus's post captured the 54% token-efficiency claim and said his Pro limits were draining faster than before.
• OpenAI also said it had changed and reverted hidden reasoning-effort experiments, the same under-the-hood area that Lentils80 described as "juice values."

OpenAI's launch post said GPT-5.6 Sol gets more work from every token, adds max and ultra, and coordinates multiple agents for harder work. OpenAI's Codex plan help article says Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents draw from the same agentic usage pool. The weird bit was not just the rollback, it was the hidden accounting surface: context ceilings, subagent inheritance, reasoning "juice," and a product merge all landed in the same week.

OpenAI confirmed the core incident in thsottiaux's update: Codex raised GPT-5.6 Sol's product context limit from GPT-5.5's 272k to 372k, then reverted it after the change caused more usage to be charged than intended.

The same post said inference optimizations should produce about 10% more usage on subscriptions, the five-hour limit would remain temporarily disabled, and the 372k limit would be rolled back out later.

That rollback sits below the public API spec. The GPT-5.6 Sol model page lists a 1,050,000-token context window for the API model, while the Codex product limit discussed by OpenAI was 372k, then 272k.

Before the official rollback, pvncher called the 272k billing theory misinformation unless users manually expanded context beyond the default. Afterward, aibuilderclub_ summarized the narrower confirmed claim: the warning was valid, but the exact credit calculation remained unconfirmed.

The most useful community root-cause map came from theo, who described four stacked multipliers rather than one bug:

1. GPT-5.6's long threads could cross the 272k boundary after Codex raised its product context limit to 372k.
2. Ultra subagents were also spawned with Ultra, creating expensive nested work.
3. Sol and Terra were using a v2 subagent layer that copied the full long context instead of starting subagents with fresh history, according to theo.
4. Fast mode could add another 2.5x on top.

OpenAI confirmed two parts of that map in thsottiaux's update: the context-limit change charged more usage than intended, and multi-agent usage was slightly higher than intended in high and xhigh reasoning effort.

Theo's exact 2x calculation was not officially confirmed. In one follow-up, he read OpenAI's "more usage being charged than expected" line as the important admission, not a precise public multiplier.

Community sleuths also found a hidden reasoning-budget surface before OpenAI mentioned it publicly.

Lentils80 said the "juice value" was the thinking budget a model can see in hidden channels, and Lentils80's explanation claimed the new max setting had fallen to the old xhigh budget.

OpenAI then said it had run experiments where reasoning efforts were changed, referred to as juice values under the hood, and reverted them in thsottiaux's update. Lentils80's follow-up called that an admission of the same nerfing he had flagged, while noting Codex was temporarily back to a 272k context window.

According to OpenAI's launch post, ultra coordinates four agents in parallel by default and trades higher token use for stronger results and faster time-to-result on demanding tasks.

The usage gap was large enough that thsottiaux's follow-up put Medium and Ultra at a 5x to 10x token-spend difference depending on task difficulty.

The subagent-control complaint was sharper. theo's subagent post said Ultra's spawned subagents also ran at Ultra, and the attached code screenshot described v1 and v2 subagent behavior: v1 exposed optional model and reasoning_effort fields, while v2 hid that metadata by default.

pvncher told one user that Sol only uses subagents in Ultra unless explicitly prompted on other modes, according to reach_vb's reply thread. In pvncher's custom-harness note, he added that Codex ships an extremely minimal system prompt and requires user consent to use subagents outside Ultra mode.

OpenAI's Codex usage help article says Codex usage varies with task size, complexity, model, and runtime surface, and that Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents share the same agentic usage and credit pool.

dkundel put it more bluntly: Work and Codex share a limit because they are the same agent, while normal ChatGPT chat uses a separate limit.

The UI made that visible. theo's balance screenshot showed a balance page saying Codex usage draws from a shared agentic usage limit, with separate five-hour, weekly, Spark, and credit balances.

That shared pool explains why Work-mode complaints looked like Codex complaints. koltregaskes asked whether ChatGPT Work used Codex usage after hitting a Work usage notice, and dkundel replied that ChatGPT Work is Codex.

OpenAI used resets as both customer relief and load management.

In the first big reset post, thsottiaux said OpenAI had made the highest-compute settings too easy to use without making their impact clear, reorganized the desktop app too aggressively, framed the launch in a way that made some Codex users think Codex was going away, and introduced regressions in existing multi-agent workflows.

The same post promised two resets that day, model-picker/default changes, restored chats and projects in the sidebar, more visible usage and reset timing, and clearer guidance on when to use Work versus Codex.

A later update went further: reach_vb said the five-hour limit had been removed, GPT-5.6 Sol was more efficient, and the product had crossed 6 million users. thsottiaux's reset explanation said automatic full resets were used because banked resets were not yet available from web or mobile ChatGPT, and full resets were more predictable for infrastructure load during the first 48 hours.

Users were already reporting the failure mode before the rollback.

Kimmonismus said he had never hit five-hour and weekly Pro limits as quickly as with 5.6 Sol, mostly using High rather than Fast Mode. He still called 5.6 a real improvement, especially for frontend work and longer agentic runs.

One Reddit user said a single Extra High Codex task used more than 70% of a five-hour limit in about 20 minutes, according to the Reddit report. Another Reddit post called Sol Max and Ultra impressive but described the pricing policy as token-cost inflation, according to a second Reddit report. The OpenAI Developer Community had similar reports, including a user saying a five-hour allowance could disappear after only a couple of tasks in the Codex usage thread.

The self-funded numbers got ugly fast. zeeg reported 1.3 billion tokens by 4 p.m., and zeeg's cost reply put that day at $900. A later dashboard from zeeg's weekly cost post showed $1,683.91 for 10,215 Codex requests across two machines, with GPT-5.6 Sol at $1,173.84.

Not every field report was negative. petergostev's rebuild report said GPT-5.6 Sol deleted 73,993 lines from a local app over 47.21 hours while keeping 954 controls validated and 42 visual parity checks pixel-identical.

The rollback landed in the same window as a broader Codex and ChatGPT Work roadmap dump.

According to btibor91's Reddit AMA summary, OpenAI said Codex had more than 5 million weekly users, twice as many as three months earlier, and had shipped 150 features and improvements over that period.

The AMA summary also listed product details that did not fit neatly into the usage incident:

• A Linux desktop app is in development, with no timeline.
• There is no Auto model routing today.
• The preferred UX is Codex inferring task difficulty while keeping a manual speed and reasoning override.
• Slack, GitHub, and Notion connectors were described as a step-function change for making Codex a productive coworker.
• OpenAI made no promise on a 1M-token context window for Sol.
• Normal ChatGPT conversations do not consume the agentic allowance, while Codex and ChatGPT Work do.
• OpenAI said benchmark cheating is a real concern and that it penalizes that behavior during evals while using independent vendors.

Kimmonismus's shorter AMA recap surfaced the same developer-facing items: Linux desktop work, better persistence, lower code complexity, no automatic model routing today, no 1M-token promise, and no guarantee that current pricing remains unchanged.