The news
AI usage limits tightened across the industry in September 2026: OpenAI paused new sign-ups for its $200-a-month Pro plan on September 10, Apple (AAPL) capped Siri AI daily, and Meta (META) declined to specify usage limits for its new consumer AI subscriptions.
OpenAI's pause came a week after it launched Astra, which it describes as its most powerful model, on September 3. TechCrunch reported that the company stopped new Pro sign-ups because Astra strains its infrastructure more than other tiers, while keeping its API, Go and Plus plans open. It did not say when Pro sign-ups would resume. Product leader Thibault Sottiaux had warned on X before the pause that demand for Astra was unprecedented.
Apple took a similar approach with Siri AI, which began rolling out in beta, in English, on September 14, with more languages due in October. The company said server-side features, including Siri AI and image generation, are subject to daily limits that vary by feature, complexity, demand and policy, and that “expanded access to such features will be available for a fee in the future.” According to 9to5Mac, users join a waitlist in the Settings app to get access.
Meta followed on September 15 with Meta One, a set of subscriptions that bundle expanded AI usage with perks across Facebook, Instagram and WhatsApp. Consumer plans cost $7.99 and $19.99 a month, and business and creator tiers have starting prices from $14.99 to $499 a month. Meta declined to share exact usage limits for the two consumer plans, saying they could vary by country, surface and system conditions, TechCrunch reported. Its Muse agent, launched September 8, pairs a limited free tier with $20 and $100 monthly plans.
Meanwhile, the unit price of AI keeps falling. Ramp, which tracks spending at about 70,000 companies, reported on September 9 that token costs fell to $0.68 per million tokens from $1.15 in March. AI spending per employee at the top 1% of firms dropped nearly 10% in August to $7,205, even as 56% of Ramp customers paid for AI products. Ramp economist Ara Kharazian linked the decline to competition between OpenAI and Anthropic.
The numbers
- OpenAI Pro plan price (new sign-ups paused Sept 10)
- $200 per month
- Meta One consumer plans
- $7.99 and $19.99 per month
- Meta business and creator tiers (starting prices)
- From $14.99 to $499 per month
- Average token cost, August vs March (Ramp)
- $0.68 vs $1.15 per million
- AI spend per employee, top 1% of firms, August (Ramp)
- $7,205
- Ramp customers paying for AI, August
- 56%
Why CEOs should care
CFOs should plan for AI that is priced like a utility with rationing at the top end. Routine usage is getting cheaper, but the newest models come with caps, waitlists and premium tiers. Build budgets on consumption forecasts rather than seat counts, set spending alerts by team, and negotiate committed-capacity terms that guarantee throughput for business-critical workloads. SiliconANGLE cited a Gartner prediction that at least half of generative AI projects will overrun their budgets through 2028.
CIOs and technology buyers should treat rate limits as an availability risk. If a provider can pause sign-ups or throttle usage when demand spikes, your contracts need explicit rate limits, priority access during peaks and notice periods for changes. Keep at least one fallback model tested and ready, and route routine tasks to cheaper models, since falling token prices make that easier. IDC's Dave McCarthy warned, according to SiliconANGLE, that agents can consume large amounts of compute with few built-in circuit breakers, so set hard limits on automated workloads.
CISOs and boards should watch for a possible side effect. If sanctioned tools are capped or unavailable, employees may turn to personal accounts and consumer apps, which risks moving company data outside approved systems. Providing enough approved capacity is a security control as well as a productivity decision, and boards should ask for regular reporting on AI spend, usage and unit costs.
The bigger picture
The pattern may reflect scarce compute at the frontier rather than weak demand, though only OpenAI explicitly tied its move to infrastructure strain. Providers appear to be protecting capacity for their newest models while competing on price for older ones. Ramesh Chettuvetty, an executive at Lightbits Labs, told SiliconANGLE that GPU demand cannot currently be met. For context, Oracle (ORCL) reported $664 billion in remaining performance obligations as of August 31, 2026, and expects to recognize about 13% of that as revenue over the next twelve months, according to its quarterly filing, so most of that contracted revenue falls beyond the next year. If capacity stays tight, rationing at the top end may persist.
What’s next
Watch for OpenAI to reopen Pro sign-ups or introduce new tiers, for Apple to price its paid Siri AI access, and for Meta to publish clearer limits. Ramp's September data will show whether August's dip in per-employee spending was a summer pause or the start of a trend toward cheaper models.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
