The flat-rate era of AI coding is ending. Here's what my stack costs now.
A while back I wrote about how my AI-coding setup settled into a ladder: Claude Code on a business subscription as the daily driver, OpenCode as the overflow lane, and GitHub Copilot demoted to a backup provider after Copilot Business moved to token-based billing and my costs went vertical. I framed that demotion as a one-off annoyance. It wasn't. It was the first raindrop. Over 2026 the whole category walked away from the flat monthly plan, and I've stopped pretending my bill is predictable.
What actually changed
This isn't vibes. Three concrete moves, all in the open:
- GitHub Copilot switched to usage-based "AI Credits" on 1 June 2026. Code completions stay included on every plan, but agent-style work now draws down a credit balance, and the announcement confirms the old safety net is gone: no more automatic fallback to a cheaper model when you run dry. Business and Enterprise seats pull from an org-pooled balance rather than a fixed per-seat allowance. That "fallback removed" line is exactly the mechanism that quietly makes a former flat seat cost more.
- Cursor restructured Teams in June 2026 into two separate included-usage pools per seat: one for its own models (Auto and Composer), one for third-party APIs (Claude, GPT, Gemini). Standard seats run $32/seat/month annual ($40 monthly), Premium $96 ($120 monthly). Read that structure carefully — the model you reach for determines which bucket drains.
- Anthropic kept its subscriptions flat (Pro at $17–20, Max from $100 up to $200), which is the whole point of them. But on the API side, Sonnet 5's introductory $2/$10 per million tokens ends 31 August 2026 and reverts to $3/$15. And there's a detail people miss: the newer models use a tokenizer that produces roughly 30% more tokens for the same text. A headline per-token price cut isn't a price cut if each request now counts more tokens.
Why the flat plan broke
The flat monthly subscription was priced for a human typing. You can only ask so many questions an hour. An agent has no such ceiling — it plans, writes, runs the tests, reads the failures, and tries again for hours, burning compute the entire time. Developer Tech put it plainly in June: Copilot "is not the same product it was a year ago." The canonical data point is older: back in January 2025, Sam Altman said OpenAI was losing money on the $200/month ChatGPT Pro plan because people used it more than expected. When the vendor loses money on their own premium tier, the meter is coming. It just took another year to arrive everywhere.
The honest math on my own stack
I'm not going to invent a precise euro figure for you, because my usage swings wildly week to week and anyone who quotes you a tidy monthly number is guessing. What I can tell you is the shape of it, and the shape is the whole lesson.
A flat subscription is the only thing left that caps a cost with no natural ceiling. That's not a discount. It's insurance.
That reframing changed how I buy. A heavy Claude Max user on the $200 flat plan can consume the equivalent of many hundreds of dollars of raw API tokens in a month — the gap is entirely a function of which billing path you're on. So the flat sub isn't the cheap option; it's the bounded option. Which also means it's the option most under pressure, because vendors can read that arbitrage as well as I can. I fully expect today's flat plans to sprout weekly caps and tiers (Max already has weekly limits), and I'm budgeting as if they will.
For scale, one developer estimated their Copilot month would run about $700 metered versus $28 flat — a ~25× jump. That's one person's estimate of their own heavy usage, not a typical bill. But nobody's estimate came out lower, and that tells you which way the wind blows.
What I'm doing about it
- Keep exactly one flat sub as the anchor. For me that's Claude Code — the daily driver lives where the cost is bounded.
- Meter everything else, deliberately. Overflow tools run on pay-per-token, and I only reach for them when I've actually hit a wall, not out of habit.
- Watch the cliffs. Intro pricing expires (Sonnet 5 on 31 August), tokenizers change, "included" quietly shrinks. I re-read the pricing pages now, because they change monthly.
- Take local models seriously for the boring work. A lot of what I hand an agent is mechanical. The economics of running a self-hosted coder for that are starting to pencil out — which is its own post.
The era where you paid twenty bucks and stopped thinking about it is over for anyone using these tools like tools. That's not a tragedy. But "unlimited" was always a marketing word, and 2026 is the year it got quietly returned to the shelf.