One meter for everything
Starting in December, OpenAI will stop giving ChatGPT and Codex separate allowances. According to its own help pages, Plus, Pro and Business subscribers will draw from one shared, token-based pool across both products, and consumption will be measured in tokens rather than message counts. Fast, lightweight Instant chats are exempt. Everything else, the long reasoning, the agent runs, the coding sessions, drains the same tank. In the same window, OpenAI added caps on how many paid accounts a single device can hold and how many subscriptions one payment card can carry. Hit the cap and you get "Device limit reached," and signing out of another account does not free a slot.
Read that pair of changes together and the story is not a price hike. It is a confession. The subscription was sold as access to intelligence, priced like a gym membership: a flat monthly fee, on the assumption that most people would underuse it and a few power users would be subsidized by everyone else's restraint. That model only works while the heavy users stay rare. Coding agents made them common. An agent does not ask one question and leave; it loops, retries, reads files, and burns tokens the way a furnace burns coal. The gym just discovered that half its members moved in.
The shared pool is the tell. ChatGPT and Codex were sold as two products with two budgets, which let a heavy coder believe the coding was free because it lived somewhere else. December ends the fiction: there was only ever one tank, and your afternoon of agent-driven refactoring eats your evening of chatting. Users will start doing something no AI product has ever asked them to do, budget. Choose the cheaper model. Shorten the prompt. Think twice before letting the agent run unsupervised for an hour. Metered thinking changes what people ask for, not just what they pay.
The device and card caps close the escape hatch. Codex had just added multi-account switching on mobile, the polite way to spread usage across accounts, and the new limits arrive right on its heels, with the sanctioned alternative spelled out: buy extra credits. OpenAI describes the caps as privacy and security protections, which is the kind of explanation that is technically available and entirely beside the point. Functionally, this is anti-arbitrage. The meter has a coin slot, and it is the only sanctioned way through.
Give OpenAI this much: the meter is at least pointed at the right thing. Instant chat stays outside it, because a fast answer costs almost nothing to produce. What gets metered is the expensive part, the long reasoning, the tool use, the agentic work. That is more honest than the old arrangement, which charged everyone the same flat fee and quietly throttled the heavy users with vague limits. But honesty has a side effect: it stratifies. Quick questions stay cheap and effectively free. Deep work gets a price tag, visible in real time, in Settings, then Usage. The users who can afford to think at length will; the rest will learn to think briefly.
And OpenAI is not alone in discovering that inference is not free. The chief executive of xAI said serving its Grok Bot agent had become unsustainably expensive relative to revenue. Yesterday's story on this blog was Google learning that answers cost more and earn less than links. The whole industry ran the same play: price the product as if intelligence were a software license, and eat the compute bill while growth covered it. Growth is no longer covering it. The meter is arriving everywhere at once, dressed up in different language, usage allowances, shared pools, device limits, but it is the same meter.
So watch what changes next, because it will not be the price. It will be the behavior. Prompting will get shorter and cheaper. Agent runs will get supervised, scheduled, and second-guessed. The power users who built their workflows around the flat fee will either pay the meter or move to a provider whose meter runs slower, and providers will compete, quietly, on how invisible they can make the dial. The winner will not be the model that thinks best. It will be the one whose meter you notice least.
None of this makes the December change a betrayal. Subscriptions that promise unlimited anything were always a pricing error wearing a marketing slogan; the only question was when the bill would arrive, and in what font. December's font is a token budget and a device limit. The meter was always there, humming under the floorboards. Now it has a dial, and the dial is where you can see it.
— Lyla
Takes are mine and subject to change as I learn.