The flat AI plan wasn't built for agents that never log off
4 Nov 2026 · News
A monthly AI plan is priced for a person who logs off. The new agents don't. The big labs are keeping their prices and starting to count the work underneath.
The flat AI subscription was built for chat, and the always-on agent is the first thing it can't pay for. OpenAI sells an agent that works around the clock on "extended limits" with no end date and no published price. From December it will measure ChatGPT usage in tokens. Anthropic's IPO filing shows its computing bill growing faster than its revenue, and says changing usage limits has already brought legal claims. The $20 price tag will survive. What breaks is the open-ended allowance behind it, and agent work will be sold by the month.
An agent with no meter yet
OpenAI launched dots on 29 September. They are always-on agents with "their own cloud computer" that "can work towards your goals 24/7." The launch post says the first dot comes with Pro "at no extra cost," with extended limits "for the first month after launch." It also says that "in the future" users will scale each dot by "the total amount of work it can take on per month."
The first month ended on 29 October. That day OpenAI went the other way. Its release notes say Pro 100 and Pro 500 now get twice their previous Work and Codex allowance, and that it "canceled the Pro 200 reduction announced in our September 29 email." OpenAI's Thibault Sottiaux said on X that dots keep extended limits "for now." Nobody has said for how long.
The unit is already chosen. OpenAI's usage limits page says that starting in December, Plus, Pro and Business plans will share one allowance across ChatGPT and Codex, and "your usage for ChatGPT and Codex will be measured in tokens." Instant chat won't count. Agent work will. Run out, and you can wait, upgrade, or buy credits.
The seller's own numbers
Anthropic's amended S-1 shows why the open-ended allowance is the weak spot. From the second quarter of 2026 to the third, its compute and infrastructure cost rose 77%, from $5.6 billion to $9.9 billion. Revenue rose 47%. Compute went from 48% of revenue to 58%.
The filing keeps flat plans out of its main growth measure. Its net revenue retention figure leaves out Claude Pro, Max and Ultra subscribers, "whose fixed-price subscriptions do not reflect the expansion and contraction behavior this metric is intended to capture." Subscriptions were $6.9 billion of $33.2 billion in revenue for the first nine months of 2026. The rest is billed by use.
It also says how the flat plan is meant to end for heavy users. When subscribers "approach or exhaust the usage capacity included in their tier," they can move up a tier or "purchase more intelligence directly," which the filing calls "a seamless bridge from subscription to incremental pay-for-intelligence revenue." In the risk section, Anthropic says it has "changed, and may in the future change, usage limits," and that some of those changes "have resulted, and could in the future result, in claims" that its disclosures were "inadequate or misleading." It reports several proposed class actions over how it markets Claude subscriptions.
Same prices, new labels
The plans themselves show the shift. On 30 October Anthropic renamed its $200 Max 20x plan to Ultra. Its help page says existing subscribers move over "with no price change," and that Ultra includes "10 times the Pro plan's usage every week." Anthropic and The Next Web say weekly limits didn't change. The 20 just dropped off the label. Claude's pricing page now lists the Fable model on the $20 Pro plan as "paid with usage credits," and offers usage credits "at standard API rates" once a limit runs out.
Google's AI plans page still describes most Gemini limits as "Expanded," "Higher" or "Highest." Its Flow video tool already has a number: 1,000 credits a month on the $19.99 Pro plan and 25,000 on the $199.99 Ultra 20x plan. Pro and Ultra members who hit a limit can buy AI credits for Flow and for Antigravity, Google's coding agent tool.
On 1 November a user asked Elon Musk on X for a usage reset, saying "usage ran out really fast this week." Musk replied: "The cost of serving Grok Bot was unsustainably high relative to revenue." It is one company's statement, and it matches what Anthropic's filing shows in numbers.
Then and now
On 12 October we wrote that your AI subscription is a meter you can't see, and led with OpenAI halving its $200 plan. OpenAI reversed that cut on 29 October, the day it was due to reach existing subscribers, and doubled the allowance for those already moved to the lower limit. Competition for heavy users is still pushing allowances up.
Our call from that post was that by the end of April 2027 OpenAI, Anthropic and Google would each show an actual unit for their top consumer plan. OpenAI now names the unit, tokens, starting in December, but hasn't published the number. Yesterday we wrote that the price war moved to the bond market. The overflow compute that agents need is the most expensive kind to buy.
Our call
By 31 March 2027, OpenAI will end the open-ended "extended limits" for dots and publish a monthly allowance or price for agent work. On the same date, ChatGPT Plus, Claude Pro and Google AI Pro will still list at $20, $20 and $19.99 a month. We're wrong if dots are still on extended limits with no published allowance on 31 March, or if any of those three plans raises its list price.