OpenAI has rolled out changes to reduce consumption when using GPT-5.6 Sol in Codex and ChatGPT Work. It has also temporarily removed the 5-hour usage limits that applied to the Plus, Business, and Pro plans. The company expects that reasoning optimization alone will allow for roughly 10% more usage. The same week, Anthropic made its own move: it re-extended the period during which Claude Fable 5 can be used within paid plans to July 19, and continues to offer a 50% increase to Claude Code's weekly usage cap. Competition among top-tier models has shifted from benchmark scores to how much work can actually get done within the bounds of a subscription fee.
Removing the 5-hour cap and two consumption fixes
OpenAI's product lead Thibault Sottiaux announced on July 13 JST that the 5-hour limits for Codex and ChatGPT Work would be temporarily suspended. This applies to the Plus, Business, and Pro plans. Since the weekly cap remains in place, usage is not truly unlimited. Users who work intensively over short bursts can avoid interruptions every 5 hours, but the amount they use is still deducted from their weekly allowance.
In a detailed post the same day, the company described two separate changes aimed at lowering consumption. The first is an optimization to the inference infrastructure, expected to let users get roughly 10% more use out of GPT-5.6 Sol. The second is a temporary reduction to the product-side context limit. In addition to these, OpenAI says it is also fixing an issue where multi-agent usage under the high and xhigh reasoning settings consumed more than intended, as well as improving the efficiency of auto-review.
It's worth distinguishing here between subscription usage allowances and API token pricing. The roughly 10% figure OpenAI cited refers to the effect of the reasoning optimization on how much Sol usage a paid plan allows—it is not an announcement of a 10% cut to API pricing. The combined effect of all the fixes, including the other changes, has not yet been quantified.
372,000 tokens drove up consumption
What caused the unexpected spike in consumption was the in-product context limit for GPT-5.6 Sol. OpenAI had expanded this limit from 272,000 tokens under GPT-5.5 to 372,000 tokens for Sol. However, this change reportedly caused more usage to be counted than intended. The company has now rolled the limit back to 272,000 tokens and expects this to significantly slow the rate at which usage is consumed.
Context refers to the range of conversation, code, and documents a model can reference in a single pass. A higher limit makes it easier to retain large repositories or long work histories, but it can also mean more information gets referenced each time. According to this explanation, both the model's own inference efficiency and how much history the product passes along affect how quickly usage allowances get consumed.
That said, the 372,000-token limit isn't gone for good. OpenAI plans to expand the limit again at a later date, though it hasn't announced a timeline. Nor has it clarified how it might change the way usage is counted at that point, or what percentage of improvement the rollback to 272,000 tokens will actually yield. There is likewise no end date set for the suspension of the 5-hour limit.
Fable 5 extended to July 19, terms unchanged
Anthropic announced on July 13 JST that it was extending the Claude Fable 5 promotion until 11:59:59 PM Pacific Time on July 19—3:59:59 PM JST on July 20. The original deadline was July 7 U.S. time, and this marks the second extension following the earlier push to July 12. The 50% increase to Claude Code's weekly usage cap will also continue through the same deadline.
This applies to Pro, Max, Team, and premium seats under seat-based Enterprise plans. Users can allocate up to 50% of their normal weekly cap to Fable 5 at no extra charge. It's available on Web, desktop, and mobile, as well as through Claude Cowork and Claude Code version 2.1.170 or later.
The "up to 50%" figure doesn't mean the weekly allowance itself is being separately increased. Fable 5 and other Claude models share the same weekly cap, and Fable 5 simply consumes that cap faster than other models. Once this limit is reached, users must either switch to paid usage credits or fall back to another model to use up the remaining weekly allowance. After July 19, Anthropic plans to remove Fable 5 from the weekly cap system and offer it solely through usage credits.
What the usage-cap race reveals about the cost of inference
The moves by both companies reflect the difficulty of fitting top-tier models into subscription plans. Via the API, GPT-5.6 Sol costs $5 per million input tokens and $30 per million output tokens, while Fable 5 costs $10 and $50 respectively. With pay-as-you-go pricing, companies can simply bill for what's used. But under a flat-rate plan, a model's inference volume, context length, and the number of concurrently running agents all have to be translated into a usage cap.
Anthropic says demand for Fable 5 is hard to predict, and it has indicated that once sufficient compute capacity is secured, it plans to return the model to standard subscription functionality. Extending the deadline week by week has effectively bought time to gauge that supply and demand. OpenAI, meanwhile, is removing the 5-hour cap while trying to improve how far the weekly allowance stretches, working on both the inference infrastructure and the product-side settings.
For users, the real difference isn't which model they can choose, but how much of a long task they can complete without interruption. And there are clear numbers to watch for: whether OpenAI reveals improvements beyond that roughly 10% figure and sets an end date for the suspended 5-hour cap; whether reintroducing the 372,000-token limit causes consumption to spike again; and whether Anthropic manages to move Fable 5 into the standard usage cap after July 19. These three points will show whether performance gains actually translate into more usable work in daily practice.