The AI industry's first few years ran on flat, predictable prices that didn't reflect what the models actually cost to run — a teaser rate, the same way an adjustable-rate mortgage starts low before it resets. In 2026, that period ended for the platforms most people actually build workflows on: Anthropic's Claude, OpenAI's Codex and ChatGPT Business, and GitHub Copilot all moved from flat or bundled pricing to usage-based billing within a few months of each other.
Short answer: The "teaser period" was 2023–2025's flat, often subsidized AI pricing that made adoption easy to budget for. Through 2026, Anthropic, OpenAI, and GitHub Copilot all shifted their business plans from flat or bundled pricing to usage-based billing at closer-to-real API rates, so heavy users now pay for what they actually consume instead of a predictable flat number.
I track pricing pages for a living — every tool I recommend on this site gets checked against its own vendor page before I write a word about what it costs. Going back through my notes from this year, the same story shows up three separate times. A plan that used to be one flat number turned into a seat fee with a meter running underneath it. In my testing of Claude Enterprise, ChatGPT Business, and GitHub Copilot's own admin dashboards, none of the underlying model quality got worse — if anything, the models kept improving. What changed is who absorbs the cost of a long context window, a multi-step agent run, or a team that leans on autocomplete for eight hours a day. That cost used to be smoothed into a flat number built for easy budgeting. Now it's itemized, and itemized bills move when usage moves.
What you'll need
You don't need a finance background to see this coming, just your own invoice and about ten minutes. Pull up whatever AI tools your team already pays for — Claude, ChatGPT, Copilot, or a raw API key billed directly — and find last month's actual bill, not just the plan name you signed up under. Check whether it shows one flat total or a line-by-line usage breakdown; that alone tells you which side of the reset you're on. If you manage a team, grab the seat count and renewal date too, since managing AI coding costs at scale gets a lot harder once usage is metered than it was under a flat fee. Nothing below required a sales call — every figure came straight off the vendors' own pricing and billing pages.
Step-by-step: check if your own AI stack is about to hit its reset wall
1. Find out if your plan already switched to usage-based billing
Anthropic restructured Claude Enterprise so a $20-per-seat monthly fee no longer includes any bundled tokens — usage is billed separately at standard API rates, with no seat-level allowance, a 20-seat minimum, and an annual commitment, according to Anthropic’s own Enterprise billing help article. That replaced older Standard and Premium seats, which charged one flat monthly fee that included a usage allowance. If you're still on a legacy seat, the switch typically lands at your next contract renewal, not the moment it's announced.
2. Compare your seat price to what you actually consume
OpenAI's Business and Enterprise plans still list Codex access at $20 per user per month billed annually, or $25 billed monthly, according to OpenAI’s own Codex pricing page. What changed is what that seat price covers: Codex usage moved off a flat per-message allowance onto token-based billing at API rates starting April 2, 2026 for Business and new Enterprise plans, then extended to all existing Enterprise, Edu, and Gov plans on April 23, 2026. The seat price alone no longer caps what a heavy Codex user costs your account.
3. Check for seat minimums and lock-in terms
Usage-based Enterprise plans often carry stricter commitments than the flat plans they replaced. Anthropic's current Enterprise terms require at least 20 seats and an annual contract before you can access the usage-based structure at all, which is a different commitment than signing up for a monthly seat and canceling anytime.
4. Model your worst month, not your average month
GitHub replaced Premium Request Units with metered AI Credits starting June 1, 2026, tracked by input, output, and cached token consumption at each model's published rate, per GitHub’s own announcement. Copilot Pro stayed at $10 a month with $10 of credits included; Pro+ stayed at $39 a month with $39 included. A quiet week barely touches that allowance. A week spent running an agent through a large refactor can burn through it fast, and overage is billed on top.
5. Watch your renewal date, not the announcement date
Because these shifts tend to roll out at contract renewal rather than the day a company announces them, the plan you're quoted today may not be the plan you're billed under in six months. Mark your renewal date and re-check the vendor's current terms a few weeks before it comes up.
6. Decide: renegotiate, downgrade, or diversify
Once you know your worst-case usage cost, you have three real options — push back on your account rep for a usage cap or discount, downgrade seats for people who barely touch the tool, or split workloads across more than one vendor so no single meter decides your whole bill.
Example prompts you can copy
I used versions of these to sanity-check my own AI spend before writing this piece:
- "Here's my last three months of [tool] usage data: [paste]. Estimate my cost under token-based billing at [rate] versus what I paid under a flat plan."
- "List the specific actions in [tool] that consume the most tokens per use — long context, image generation, agent mode, multi-turn chains — so I know what to watch."
- "I'm about to renew a [tool] contract. What questions should I ask about usage caps, overage rates, and seat minimums before I sign?"
- "Compare my current AI tool spend against running the same workload through the raw API directly. Where would usage-based billing actually save money?"
Paste real numbers from your own invoice rather than a hypothetical — vague usage estimates produce vague, unhelpful answers back.
Common mistakes to avoid
The mistake I made first: assuming a vendor's headline seat price was still the whole story, when for Codex and Claude Enterprise it's now just the entry fee. Second, budgeting off an average month instead of the busiest week of the quarter — a metered plan bills the spike, not the average. Third, missing the actual switch date because it happens quietly at contract renewal instead of on the day it's announced in a blog post; I've since started calendaring renewal dates for every AI tool the way I would a lease. Fourth, treating "still $20 a month" headlines as good news without checking what got removed from that $20 — in Codex's case, the flat message allowance that used to cap your cost. Fifth, canceling everything in a panic instead of checking which specific seats in your account are actually heavy users; in most teams I've looked at, a handful of accounts drive most of the usage-based cost.
Where the flat-fee "teaser" era ended in 2026
| Platform | Old model | New model | Effective date |
|---|---|---|---|
| Claude Enterprise (Anthropic) | Flat per-seat fee with a bundled, discounted token allowance | $20/seat/month + usage billed separately at API rates, no included tokens, 20-seat minimum | Renewals from Nov 2025; default for new contracts by Feb 2026 |
| Codex / ChatGPT Business (OpenAI) | Flat per-message allowance bundled into the seat | Seat stays $20/user/month annual ($25 monthly), but usage billed by token consumption at API rates | April 2, 2026 (Business); April 23, 2026 (all Enterprise) |
| GitHub Copilot | Premium Request Units — a fixed monthly allowance of flat-rate requests | AI Credits metered by token usage (input, output, cached) at each model's published rate | June 1, 2026 |
Tools that make this easier
None of this means you need to abandon these tools, just budget for what they actually cost now instead of what they used to cost. My deeper breakdown of why AI is getting way too expensive covers the consumer side of this same shift across ChatGPT, Claude, and Gemini plans. If you run a development team specifically, managing AI coding costs at scale walks through budgeting and cutting waste once seats stop being the only variable. For the Codex side of this change specifically, how to use ChatGPT Codex covers the workflow itself, and if your team lives in the editor, how to use GitHub Copilot in VS Code is worth a read before your next renewal. If you want to track Claude Code usage before it shows up as a surprise on an invoice, Claude Code usage tracking by LangWatch covers a tool built for exactly that. And if you're curious how this pricing reset connects to the bigger financial picture — the capital being spent to build the compute behind these models — I've covered that separately in is there an AI bubble and the AI trade now runs on borrowed money.
My take
The teaser-rate comparison holds up better than most metaphors borrowed from finance. Flat pricing wasn't a permanent feature of these tools, it was a launch strategy — a predictable number that made AI easy to slot into a budget while the vendors figured out what usage actually looks like at scale. Now that they know, three of the biggest platforms converted within months of each other, and that's not a coincidence, it's what happens when subsidized adoption pricing runs into the real cost of running frontier models at scale. None of that makes Claude, Codex, or Copilot worse tools. It does mean the number on the pricing page is no longer the number on your invoice, and the gap between the two is exactly what I'd check before signing your next renewal.
Frequently Asked Questions
What does "the teaser period" mean for AI pricing?
It refers to 2023–2025, when major AI vendors sold flat, often subsidized plans that made costs predictable and adoption easy, closer to a low introductory rate than a price that reflected real compute costs.
How do I know if my AI plan already switched to usage-based billing?
Check your last invoice for a line-by-line usage breakdown instead of one flat total, and check the vendor's current pricing page against what you signed up under — Anthropic, OpenAI, and GitHub Copilot all made this switch during 2026.
What is the easiest way to avoid a surprise AI bill?
Pull your busiest week of usage, not your average week, and estimate cost from that instead of the headline seat price. A metered plan bills your spikes, not your typical month.
Are Claude, ChatGPT, and GitHub Copilot still worth paying for after the reset?
Yes, for most regular users — the models didn't get worse, and light users often see little change. The reset mainly hits heavy, high-volume accounts that used to get flat pricing regardless of how much they consumed.
Will AI subscription prices keep going up after 2026?
Probably, at least for usage-based tiers, as long as compute costs stay high and vendors keep converting flat plans to metered ones. The safest habit is checking your renewal terms every cycle instead of assuming last year's price still applies.