OpenAI Loosens The Leash On Usage Caps

Apple filed a sweeping lawsuit against OpenAI on July 10, alleging a coordinated effort to steal hardware trade secrets through former Apple employees now building OpenAI's device business. Days later, OpenAI reported crossing 6 million active users on its Codex and ChatGPT Work products and temporarily lifted the five-hour usage cap that had been throttling paid subscribers. The two stories aren't connected, but together they capture where OpenAI sits right now: legally exposed, technically ambitious, and stretching its infrastructure to meet demand it may not have fully planned for.

Key Points:

• Apple's complaint names OpenAI, io Products, Chief Hardware Officer Tang Tan, and former Apple engineer Chang Liu, alleging Liu exploited a network bug to download confidential hardware files and Tan used insider knowledge to extract information from job candidates still at Apple.
• OpenAI's GPT-5.6 Sol Ultra reportedly produced a proof of the 50-year-old Cycle Double Cover Conjecture using 64 parallel subagents in under an hour, though mathematicians are still reviewing the claim.
• The company's head of safety systems, Johannes Heidecke, is departing as OpenAI folds safety reporting into a combined research and safety organization under VP Mia Glaese, the sixth senior safety departure in two years.
• OpenAI temporarily lifted the five-hour usage cap for Plus, Pro, and Business plans on Codex and ChatGPT Work, rolled out token-efficiency improvements to GPT-5.6 Sol, and issued a one-time "banked reset" to its user base as active users grew from 6 million to 7 million within days.
• CEO Sam Altman has walked back earlier warnings about mass AI-driven job losses, now describing AI as a net job creator so far, a reversal shared by Anthropic's Dario Amodei around the same time.

The Lawsuit: Apple Says This Was Systematic

Apple's complaint, filed in the Northern District of California, describes more than isolated incidents. It alleges Chang Liu kept an Apple-issued laptop after leaving for OpenAI, discovered an authentication bug that let him access Apple's internal network, and downloaded dozens of confidential files covering unreleased hardware designs and manufacturing data. Separately, Apple accuses Tang Tan, now OpenAI's chief hardware officer and a former Apple vice president of 24 years, of using internal Apple codenames during interviews with Apple employees considering OpenAI, and asking some to bring physical hardware parts to interviews for "show and tell" sessions. The complaint puts the number of former Apple employees now at OpenAI above 400. OpenAI has denied the allegations, and no court has ruled on the claims.

The Technical Story: A Math Proof And A Safety Exit, In The Same Week

GPT-5.6 Sol Ultra's claimed proof of the Cycle Double Cover Conjecture drew real attention from mathematicians, including praise for the approach alongside caution about the conjecture's history of incomplete proofs. OpenAI published both the result and the exact prompt that produced it, running 64 subagents in parallel with instructions to pursue diverse approaches before converging. That technical high note landed the same week Johannes Heidecke, OpenAI's head of safety systems, announced his departure as the company merged safety reporting into its research organization under Mia Glaese. OpenAI's chief research officer framed the change as giving safety an earlier role in product decisions. It is also the sixth senior safety departure at the company in two years, a pattern outside observers have noted without a settled read on what it means for oversight going forward.

The Usage Cap Story: What Changed

OpenAI's product lead Tibo Sottiaux announced three things at once: a temporary removal of the five-hour usage window for Codex and ChatGPT Work on paid plans, token-efficiency upgrades meant to stretch existing usage further, and a one-time "banked reset" that lets a saved usage refill be applied whenever an account needs it. None of this is a permanent increase to weekly limits, and OpenAI has not given an end date for the temporary changes. The moves followed visible strain, users hitting the five-hour ceiling mid-session during heavy coding or research work, right as active usage grew fast enough to cross both a 6 million and a 7 million milestone within the same week.

What This Means For Teams Leaning Heavily On AI

Any team that has built real workflow around a single AI vendor's usage allowance just got a preview of what happens when that allowance becomes the bottleneck. A five-hour cap that throttles you mid-project isn't a hypothetical risk, it already happened here, publicly, to millions of paying users. Teams have a few concrete ways to plan around this rather than simply hoping caps stay generous:

• Route tasks across more than one model or vendor. Reserve your highest-tier, most expensive model access for work that genuinely needs it, and send lower-stakes, high-volume tasks to a cheaper or less capacity-constrained option.
• Build slack into deadlines around known high-demand windows. New model launches and demand surges are when caps tighten fastest, so plan critical work around, not during, those periods when possible.
• Track usage at the team level, not just the individual level. A "banked reset" or temporary cap lift benefits whoever notices and uses it first; without visibility into who's consuming what, teams can hit limits without warning.
• Treat vendor usage policy as a live risk, not a settled fact. Caps, resets, and efficiency changes are being adjusted in near real time right now, and a policy read last month may already be out of date.
• Have a fallback workflow that doesn't require AI at all. For anything genuinely deadline-critical, know what the manual path looks like if the tool is capped, slow, or down.

Building that kind of resilience into an AI-dependent workflow is exactly the kind of planning we help clients think through as part of a growth strategy built for tools that are still finding their footing operationally.

If your team wants a clearer view of where AI vendor dependency creates real risk versus manageable inconvenience, our AI marketing services team can help you map it out.

Source: TechCrunch, "Apple sues OpenAI over alleged trade secret theft"