If you run an engineering team that has gotten comfortable with AI coding tools over the past year, today's GitHub announcement is the email you didn't want to read.
All Copilot plans move to usage-based billing on June 1, 2026. The headline subscription prices stay the same. Pro at $10 per month, Pro+ at $39, Business at $19 per seat. What's changing is everything underneath those numbers.
Under the new model, your subscription buys you a wallet of "GitHub AI Credits" priced at API token rates, equal in value to your monthly fee. Empty the wallet and you stop, you cap your spend, or you pay overage at published rates. The buffer between what you paid and what you could actually consume, the part that made these tools feel like a Netflix subscription, is gone.
GitHub's own announcement contains the most candid line in the AI coding industry this year:
It's now common for a handful of requests to incur costs that exceed the plan price.
Today isn't a surprise if you've been paying attention. It's the dam breaking on a story that has been building for three years.
For the past three years, software developers have been working with AI coding tools whose retail prices bore almost no relationship to the underlying cost of compute. Investors and strategic partners footed the difference. The difference was enormous.
In mid-2025, an investor analysis of Cursor, the most successful AI-native IDE on the market, found the company was paying roughly $650 million annually to Anthropic against approximately $500 million in revenue. That's a negative 30% gross margin, before salaries, offices, or infrastructure. One investor put it more bluntly to the journalist Ed Zitron:
Cursor is spending 100% of its revenue on Anthropic.
A user-level view is just as striking. An engineer who tracked her own Cursor consumption in 2025 paid $180 for the year. By her estimate, she had burned through somewhere between $838 and $2,015 worth of tokens at standard API rates. And she was an average user, not a power user. A user in the top 6% of Cursor consumption was running through 6.24 billion tokens against a $2,400 maximum-tier subscription. That isn't a margin problem. That's a customer paying roughly a tenth of what they're consuming.
The model providers were running their own subsidies on top of that. Anthropic spent $2.66 billion on AWS against $2.55 billion in revenue through September 2025. They paid their cloud provider more than they were bringing in from customers. OpenAI lost an estimated $5 billion in 2024 and was on pace for around $8 billion in 2025.
The cloud providers underwriting those losses were also on the cap tables. Google and Amazon as Anthropic investors. Microsoft as OpenAI's strategic partner. The whole arrangement is a circular bet on long-term market share.
Every actor in the stack was losing money on every transaction. The tools felt cheap because someone else was paying.
This pattern has a name in business strategy, and it predates AI by decades.
Amazon ran negative or thin margins through much of the 2000s while building a logistics moat competitors couldn't replicate. Uber lost money on every ride for over a decade while making car ownership feel optional. Airbnb subsidized hosts and hospitality experiences until hotels stopped looking like the obvious choice. Netflix borrowed billions to fund original content while training viewers out of the cable bundle.
In every case, investors funded losses to capture market share. Once switching costs were high enough, prices started to rise.
The AI coding playbook fits the template precisely. Subsidize the tools, change how engineers work, get teams to build their workflows and their staffing models around AI-assisted development, and let switching costs do the rest. Pricing power in the second phase is what pays for the losses in the first phase. Investors aren't being charitable. They're funding the customer acquisition that pricing leverage will later monetize.
What's distinctive about AI coding is the speed and the scale. Cursor went from launch to over $1 billion in annualized revenue in roughly 24 months and crossed $2 billion in early 2026. Anthropic closed a $30 billion funding round in February 2026 at a $380 billion valuation, the second-largest private financing round in tech history. The capital flowing in is enormous, and so is the loss rate. That combination forces the pricing transition on a faster timeline than it ran for Amazon or Uber.
The repricing didn't start with today's GitHub announcement. It started with Anthropic, six months ago, and worked its way down the stack.
In November 2025, Anthropic quietly began moving enterprise customers off seat-based pricing onto usage-based contracts at renewal. Where customers had been paying up to $200 per user per month with a generous bundled token allowance, they would now pay a smaller seat fee plus actual usage at standard API rates.
By February 2026, Anthropic had restructured the entire enterprise SKU. The legacy $40 and $200 seat tiers were retired. New customers paid $20 per technical user (Claude Code) or $10 per business user (Claude.ai) for platform access only. Usage was billed separately, with a mandatory monthly spending commitment based on Anthropic's estimate of consumption. The 10 to 15% API volume discounts that large enterprises had been negotiating disappeared. The IT procurement firm NPI Financial projected the new model would increase total cost of ownership for most organizations, sometimes by a factor of two or three for heavy users.
Anthropic also rolled out "Priority Service Tiers," pricing escalators for application providers that wanted reliable capacity. Cursor's AWS bill reportedly doubled in a single month after the change went live, from $6.2 million to $12.6 million.
The downstream effects accumulated through April. OpenAI moved its Codex agent from flat-message pricing to token metering in early April. On April 10, GitHub tightened Copilot usage limits. On April 20, GitHub paused new signups for Pro, Pro+, and Student plans, tightened weekly limits, and removed Opus models entirely from the Pro tier, keeping Opus 4.7 only on Pro+ while pulling Opus 4.5 and 4.6 from Pro+ as well. Today's announcement is the final piece. Usage-based billing across every Copilot plan, effective June 1.
It's worth pausing on what GitHub kept free. Code completions and Next Edit suggestions, the original Copilot features, still don't consume credits. What got squeezed was agent mode, Copilot CLI, and the long-running parallelized agentic workflows that have been driving most of the marketing energy. The pricing structure is finally tracking the underlying cost structure. Vibe coding and autonomous agents were the bait. Usage-based billing is the hook.
There is a second story underneath the first one. Anthropic's API uptime over the 90 days ending April 8, 2026 was 98.95%, well below the 99.99% benchmark that established cloud providers operate at. The pricing changes aren't only about extracting margin. They're also rationing access to compute that Anthropic literally cannot expand fast enough to meet demand at old prices. The dam isn't only breaking because the labs want it to break. It's breaking because the unit economics of frontier AI inference are finally catching up to the marketing.
One more relationship is worth flagging. Anthropic poached the lead engineer behind Claude Code from Cursor in 2025. Anthropic's largest API customer is now a direct competitor to Anthropic's own fastest-growing product. Pricing decisions made under that dynamic should not be expected to favor third-party tooling.
If you run a software engineering team that has integrated AI coding tools into your daily workflow, today's announcement is a budgeting event you should plan for now, not a story to read about.
Any business case that pencils only because of $20 flat-rate Cursor seats or unlimited Opus access on Copilot Pro+ is built on sand. The right exercise is to estimate the actual token consumption of your team's workflows and price it at standard API rates. If the resulting number is materially different from your current AI tooling spend, you are consuming a subsidy that has a known expiration date.
The new pricing models reward teams that match the right model to the right task. Opus-class models for genuinely complex reasoning. Sonnet-class for the bulk of day-to-day work. Haiku-class for high-volume simple tasks. Teams that defaulted everything to the most capable model when it was effectively free will need to develop more discriminating habits.
Long-running, parallelized agent workflows deliver real value. They also consume tokens at a rate an order of magnitude higher than chat-style usage. GitHub's announcement explicitly calls out this dynamic. Engineering leaders should understand which of their workflows are token-expensive before the bill arrives, not after.
Anthropic competes with its own customers. OpenAI competes with Microsoft's Copilot through Codex. Cursor is racing to substitute its own proprietary models for Anthropic's. Pricing decisions over the next twelve months will reflect those tensions, and the customers buying the products will absorb the consequences.
The deeper point is this. The past three years have been an unusual period in the history of developer tooling, a period when the tools available to engineers were dramatically better than anyone could currently afford to build sustainably. That window is closing.
The tools aren't going away. AI doesn't replace great engineering. It amplifies it, and the productivity gains are real. But the bill is finally arriving, and it's being addressed to the people doing the work.
The teams that will navigate this best are the ones who treat AI as one capable tool in the hands of highly-skilled engineers, not as a substitute for them. Subsidized or not, that has always been the right way to use these tools.
