In June, GitHub changed how it bills for Copilot. The sticker price didn't move. Copilot Pro is still $10 a month, and Pro+ is still $39. Everything underneath it moved.
The old deal counted requests. The new one counts tokens. It counts every input, output and cached token an agent chews through, drawn down against a monthly pool of credits where one credit is worth a cent. GitHub's own announcement was open about why, saying Copilot "is not the same product it was a year ago" and that aligning price to actual usage was the route to a sustainable one.
When the first full billing cycle closed on 30 June, the bills landed. The Register covered developers threatening to walk, including one who burned 16 per cent of a monthly Pro+ allowance in a single morning for what they judged to be very little. GitHub added a hard spending cap on 2 July, which tells you roughly how the first month went.
Copilot has been one of those genuinely useful, genuinely daft-value tools for anyone working in and around web development. Natural language in, working code out, for less than a round of drinks. The question I keep coming back to is whether June was a one-off or the shape of things.
I reckon it's the shape of things, though not quite in the way I first assumed. Here's what's actually going on, and what I'd do about it if I were running a small business on top of these tools. Which I am.
Why your $20 a month plan isn't really $20
Whatever plan you're on, the chances are your AI of choice is subsidising you. We now have a decent measure of by how much.
The research firm SemiAnalysis bought every paid tier from OpenAI and Anthropic and deliberately ran them into the ground, pushing each with long agent-style coding tasks until it hit the weekly cap, then working out what those tokens would have cost at standard pay-as-you-go rates. Tom's Guide covered the findings. A maxed-out $200 ChatGPT Pro plan pulls around $14,000 worth. Claude's equivalent top tier lands near $8,000. Even the $20 plans return many times what they cost, roughly $700 on ChatGPT Plus and $400 on Claude Pro.
Worth being precise about what that means, because a lot of the coverage wasn't. Those are API-equivalent figures, not the labs' actual compute bill, and API prices carry margin. The real cost of serving you is a good deal lower. But the gap is still enormous, and it's still the wrong way round.
The reason is compute. Every prompt you type has to be processed and answered, and that costs real money in real data centres. Complex prompts cost more. Simple prompts still cost something. Agent workflows, where the model goes away and fires off dozens of calls on your behalf, cost a lot more than either.
None of this is a scandal. It's a strategy, and an old one. Buy the market, build the habit, work out the monetisation later. Uber did it. Amazon Prime did it. The difference is scale. As The Register put it in May, the data centres behind these tools were built with borrowed money and were never meant to run at a loss indefinitely.
Anthropic's annualised run rate crossed $47bn in May 2026, on the back of a $65bn funding round at a $965bn post-money valuation. Weeks later the company confirmed it had confidentially filed for a public listing, and CNBC reported OpenAI was readying its own.
Nobody raises money like that without a plan to get it back.
What actually changed, and what didn't
The industry is struggling to stick the landing. We've seen three distinct attempts in as many months, each with a different outcome.
GitHub Copilot, on 1 June, replaced flat requests with metered tokens. The monthly fee remained, but it now describes an allowance rather than a ceiling. There was a loud backlash, and GitHub didn't reverse the change.
Anthropic took a different path. It announced a split in its subscriptions for 15 June, creating two pools, one for interactive chat and one for agent workloads running through the SDK, headless commands and third-party tools. Then, as The New Stack reported, on the very day it was due to go live Anthropic pulled it, saying it was reworking the plan and would give advance notice before anything else took effect. This followed an earlier, failed attempt in April to remove Claude Code from the $20 Pro plan. In the same week as the reversal, a proposed class action was filed in California alleging the Max tiers fall short of their advertised usage multipliers.
Cursor also moved in June, splitting Teams seats into two usage pools and adding a premium tier at five times the usage for three times the price. Cursor's own announcement reckoned the change would lower costs for 90 per cent of teams, and the market largely accepted it.
So the direction of travel is clear enough, but it isn't a done deal and customers aren't powerless. Two of those three got walked back, at least partially, because enough people made enough noise. That's worth knowing.
If you want the one-sentence version: my read is that the sticker price is becoming the floor rather than the cap, but not without a fight.
The bit that surprised me
Here's what properly changed my mind.
Tokens are not getting more expensive. They're collapsing in price, and have been for years. Epoch AI found that the price to reach a given performance milestone has been falling somewhere between 9x and 900x per year depending on the task, though it's careful to note the steepest falls came most recently and may not hold.
And yet the bills are going up.
Researchers at MIT found both halves of this in the same study. The price for a given level of benchmark performance is falling around five to tenfold a year. Meanwhile the price of running frontier models is rising between three and eighteenfold a year, because the models got bigger and reasoning demands more.
You can see the same thing in your own workflow. A chat message is one call. An agent asked to fix a bug across a repository might be 50, with the full context sent each time. Cheaper tokens made agents possible, agents ate the savings, and then some.
This matters for how you plan, because it kills the comforting assumption that you can wait for prices to come down. They already have, dramatically. It didn't help.
The grandfathered plan, and why it isn't a plan
Here's the other pattern worth watching, and it's the one that affects businesses nowhere near a code editor.
Canva got there first. In September 2024 it moved its Teams plan from a flat rate to per-seat billing, and TechCrunch, who broke the story, reported a Canva spokesperson pointing to the company's growing suite of generative AI tools as the reason. The original pricing, they said, reflected an early-stage product and had been unchanged for four years. For some customers the annual cost roughly tripled. The reaction was loud enough that by October Canva had walked part of it back, restoring original pricing for early adopters and publishing a 'Pricing Promise' committing to at least 60 days' notice of future changes. New customers pay the new rates. Long-standing ones largely kept what they had.
Which is where we come in. You open the invoice each month, see the same modest figure that's been there for years, and quietly get on with your day. We're on a legacy Canva plan that no longer exists for anyone signing up today. It's a lovely thing to have. It isn't a strategy, and I try to remember that.
Because a grandfathered rate is a retention mechanism, not a promise. It exists for as long as it's cheaper to keep you than to lose you.
Microsoft ran the same play at a much larger scale this summer, and it was a masterclass in value-add framing. On 1 July 2026 it raised commercial Microsoft 365 prices by up to 43 per cent depending on the plan, with Business Standard going from $12.50 to $14 per user. Office Watch has the full breakdown. Microsoft simultaneously folded Copilot Chat capability into every base tier, so the story is added value rather than a price rise, for a set of AI features many of those customers never asked for. Anyone who renewed before 1 July keeps their old price until the next renewal, and gets the new features anyway. Grandfathered, right up until the day you aren't.
There's a third route too, which is to stop charging altogether and sell attention instead. OpenAI announced in January that it would start testing ads in ChatGPT for free and Go tier users in the US, reversing Sam Altman's earlier position that advertising was a last resort. By late March the company said early results were encouraging and began expanding beyond the US. If you're a small business using a free tier as part of your process, you're not outside all this. You've been moved to the advertising-funded side of it.
Where I think this goes
I don't have a crystal ball, and this is tech, so it moves fast. But there are only really three ways a provider closes the gap, and they aren't mutually exclusive.
First, meter it. Predictable for the provider, unpredictable for you. That's precisely the wrong way round if you're a small business.
Second, advertise into it. Fine for casual use, awkward the moment you're relying on the output for client work.
Third, tier it harder. Move the useful capability up a level. You don't get a price rise email, you just find the thing you were using now needs the plan above. This is the one SemiAnalysis reckons most likely, since openly cutting subscription limits invites a public backlash. Keep the plans looking generous, but hold the newest and most expensive models back for API and enterprise channels. The headline price stays flat while the value quietly erodes. No lab has confirmed this, to be fair. It's an inference.
There's a fourth possibility worth naming, which is that competition keeps everything cheap indefinitely. Open-weight models are genuinely good now and getting better, and they put a floor under how much anyone can charge for commodity work. There's a sign of it already, with OpenAI reported in June to be weighing significant price cuts in a fight with Anthropic over users. Price wars are lovely for buyers, right up until one side wins.
I wouldn't bet the business on it. I wouldn't dismiss it either. What I would say is that public companies answer to shareholders, and shareholders tend to prefer margins to your all-you-can-eat plan.
What I'd actually do about it
None of this is a reason to stop using AI. We use it every day and it's made real, measurable differences to how we work. It's a reason to stop treating it as a fixed cost.
Know your number. Add up what you spend across every AI tool each month, including the ones bundled into things you already pay for. Most people can't tell you this figure, which is rather the problem. You can't manage exposure you haven't measured.
Set hard caps wherever they're offered. GitHub, Anthropic and Cursor all now expose spending limits and usage dashboards. Turn them on before you need them, not after the bill.
Stress test at 3x and 10x. If your AI spend tripled next month, what would you cut? If it went up tenfold, would the business still work? You don't need a spreadsheet. You need an honest five minutes.
Know what you'd pay at today's rates. If you're sitting on a legacy plan, work out the current cost of the same thing. Not to panic about it, just so the number isn't a surprise if the letter ever arrives.
Check annual versus monthly. When GitHub switched, annual subscribers stayed on their old terms until their plan expired. Microsoft did the same at renewal. Annual billing locks you in, but it also locks the terms in. That cuts both ways and it's worth deciding deliberately rather than by default.
Keep your work portable. Prompts, context documents, house style guides and process notes should live in your own files, not solely inside one vendor's projects or workspace. Not because a provider is about to vanish, but because the ability to move at short notice is the only real leverage a small customer has.
Say something when it changes. Two of the three billing changes above got reversed or softened after customers complained. That is not nothing.
Be honest about your pricing. This is the uncomfortable one. If AI has halved the time a job takes and you've passed all of that saving on, you've quietly built somebody else's subsidy into your rates. When the subsidy moves, so does your margin, and you'll be having a difficult conversation about a cost your client can't see.
What's changed since this was published
I'll keep this section updated as things move, because they will. Newest first.
- 10 August 2026. First published.
