AI providers move toward usage-based pricing — what is shifting right now
The major AI providers are softening their flat-rate plans and actively pushing toward metered billing. Anthropic quietly flipped several switches in April and May 2026, OpenAI has publicly said that “unlimited” cannot hold, and Cursor has already migrated its billing model. Individually these look like pricing tweaks; together they are an industry shift — and it lands hardest on the people who have wired AI deep into their workflows.
What has concretely changed
- Anthropic quietly removed Opus 4.6 from the model picker, tied Fast Mode to the separate Extra Usage budget and shipped a new Opus 4.7 tokenizer that emits up to 35 % more tokens for the same text. On May 1, 2026 staggered API volume discounts started (8 % above $10,000, 15 % above $50,000).
- OpenAI signals openly that “unlimited” plans are ending. Nick Turley (head of ChatGPT): “In a world where technology changes this fast, there is no world where the pricing system does not change dramatically.” Sam Altman now compares AI explicitly to an electricity meter.
- Cursor moved in spring 2026 from request-based billing to monthly token credits — same logic, different packaging: usage-based.
What used to apply
The first two years of mainstream AI ran on flat subscriptions: $20 for ChatGPT Plus, $200 for ChatGPT Pro, $20 for Claude Pro, $100 for Claude Max, $20 for Cursor Pro. Inside those plans sat generous quotas, often marketed as “unlimited”. The promise: pay a fixed sum and stop worrying about consumption.
This worked while people used AI as a chat tool — a few thousand tokens per session, a few hours of activity per day. With the rise of agentic workflows, long coding sessions in Claude Code, Cursor or Codex, and reasoning-heavy tasks that internally burn thousands of tokens per request, per-user consumption shifted by orders of magnitude. A serious session now easily runs at 500,000 to 1 million tokens — the old flat rates were calibrated two orders of magnitude lower.
What applies now
1. Flat rates get hollowed out without the list price moving. Anthropic leads here. Since April 16, 2026 Opus 4.6 is no longer in the default model picker; Fast Mode is gated behind the optional Extra Usage budget; the new Opus 4.7 tokenizer eats noticeably more tokens for the same text. The effect is a de facto price increase per task without any officially published token price moving. The existing news piece covers the Anthropic shift in detail: Anthropic changes Extra Usage settings — quietly.
2. Volume tiers become standard. Anthropic introduced staggered API discounts on May 1, 2026 — 8 % above $10,000 monthly spend, 15 % above $50,000, custom above $250,000 (detail news). The discounts are only attractive if you play at volume. For smaller teams, effective prices rise relative to the competitive landscape.
3. OpenAI is openly preparing the end of “unlimited”. On the Biztoopod podcast, Nick Turley said an unlimited flat rate is like an unlimited electricity plan — “that makes no sense”. He described the current subscription model as a “stopgap” for capacity constraints. Sam Altman has talked in multiple 2026 interviews about “AI on a meter, like electricity”. The signal is clear: OpenAI is preparing pay-as-you-go components or a move to tiered usage models. A concrete launch date is not officially confirmed as of May 16, 2026.
4. Cursor has already pivoted. The old request-based Pro plan turned into one with monthly token credits. Going over means buying extra credits or stepping up a tier. Practically, these are usage-based plans with prepaid budgets, not “unlimited” anymore.
Context
This movement is not coincidence. Per-token compute costs keep falling, but consumption per active user is rising faster. With agentic workflows, autonomous coding sessions and reasoning models that internally burn thousands of tokens per request, the margin on flat-rate plans has long been negative — at least for the top 10 % of power users, who also tend to be the most valuable customers.
What providers are doing now is a careful market correction: not a hard price hike, but a gradual shift of perception. A toggle today, a new tokenizer tomorrow, a volume tier the day after. People paying attention see the pattern — everyone else just wonders why their plan quota seems to drain faster every month.
For agencies, developers and anyone whose workflows truly depend on AI, the message is: redo the math. A 200 € flat budget will turn into anywhere from 50 € to 500 € under usage-based billing, depending almost entirely on how clean your model routing, context preparation and output discipline are.
What you can do right now
If you sit on flat-rate plans: Set up an API account and run a few typical tasks through it. That makes real token consumption visible — and it is the only solid basis for planning the coming months.
If you run “Opus for everything”: Break that habit now, before it becomes a cost problem. Sonnet is enough for most follow-up work, as long as the plan from the big model is clean.
If you use the Anthropic stack: Set a hard monthly Extra Usage limit in your account so unintended consumption cannot run away. The Opus 4.7 tokenizer effect quietly runs in the background — even without Fast Mode active.
If you use Cursor: Look carefully at the new credit logic. What used to be “Pro unlimited” now has a hard ceiling.
Deeper practitioner view
How I absorb this shift in my own work — model routing, KIDOKU/RAG, output discipline — is here: → Usage-based instead of flat-rate: The AI cost turn was predictable — and here is how to absorb it
Entdecke mehr
Usage-based instead of flat-rate: The AI cost turn was predictable — and here is how to absorb it
AI providers are pushing flat rates toward metered billing. Why it had to happen, what it costs and three levers that soften the shift.
Glossar$/MTok (Cost per Million Tokens)
Standard pricing unit for AI APIs — cost in US dollars per one million processed tokens. Listed separately for input, output and sometimes cache.
LexikonAI Pricing Made Simple
How AI models bill — tokens, input vs. output, hidden cost drivers and three levers to save. With price table and worked examples.