The AI coding subsidy era ended today

Every developer on Reddit spent the weekend posting screenshots of their new Copilot bill. The numbers looked terrifying: $29 a month ballooning to $750. One user shared a screenshot showing $3,000. "What a joke," wrote one. "I'm cancelling," wrote another. The consensus formed fast: Microsoft pulled a bait-and-switch.

I think the consensus is wrong about one thing. The old price wasn't real. It was a subsidy. And the people posting those screenshots are finding out what their workflow actually costs when someone stops covering half the bill.

The buffet defense

The anger isn't irrational. Microsoft spent two years building Copilot features that burned through tokens at industrial scale. Sub-agents that chain for hours. Autonomous coding sessions that span entire repositories. Agentic workflows that spawn dozens of parallel requests. GitHub's own marketing celebrated this usage. Then on June 1, they switched from flat-rate to token-based billing, and the people who used the product exactly as designed got a bill they didn't recognize.

"Microsoft provided this billing method and they kept making it easier and easier to burn through massive numbers of tokens," one Reddit user wrote. "The only one at fault here is Microsoft." That's a fair argument. If you tell people the buffet is all-you-can-eat for $29, then hand them an itemized receipt afterward, you can't blame them for eating.

Anthropic is two weeks behind. On June 15, Claude Code splits into two billing pools: interactive use stays on your subscription, but every claude -p call, every Agent SDK invocation, every GitHub Actions integration moves to a separate credit at full API rates. A Max 20x subscriber gets $200 in monthly programmatic credit. At Sonnet 4.6 pricing, that's $3 per million input tokens and $15 per million output. A team running moderate CI volume can exhaust that credit in days. One community analysis pegged the effective price shift between 12x and 175x depending on workload. That's not a rounding error. It's a different business model entirely.

Why Copilot and Claude Code pricing was never real

Here's where I disagree with the outrage. The subsidies were never hidden. GitHub wrote it down in public. In their April 27 blog post, Mario Rodriguez said explicitly: "A quick chat question and a multi-hour autonomous coding session can cost the user the same amount. GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable." A week earlier, Joe Binder posted that "it's now common for a handful of requests to incur costs that exceed the plan price." They told us they were losing money on every power user. We chose not to listen.

One analysis found that a $20 Claude Pro subscription could produce roughly $236 in API-equivalent value per month through tools like OpenClaw. That's not a pricing model. That's venture capital buying market share. Boris Cherny, head of Claude Code, said that third-party tools operating outside the prompt cache system are "really hard to do sustainably." When Claude Code's source leaked earlier this year, the internals revealed a product generating $2.5 billion in annual revenue. The subsidy was never charity. It was a growth tactic with an expiration date. Both vendors spent months signaling. The question was never if it would end. Only when.

The $750-a-month screenshots tell a more specific story than "prices went up." GitClear's January 2026 report found that regular AI tool users averaged 9.4x higher code churn than their non-AI counterparts. More than double any productivity gains the tools delivered. Faros AI's March data showed an 861% increase in code churn under high AI adoption. Jellyfish analyzed 7,548 engineers in Q1 and found that the ones with the largest token budgets achieved 2x the throughput at 10x the cost. Volume, not value. The person posting a $750 bill and the person posting a $29 bill were using the same tool. The price difference is the efficiency difference. One was coding. The other was vibe coding.

What tokenmaxxing cost Uber, Salesforce, and Microsoft

The old flat-rate model didn't just obscure individual expenses. It hid costs from entire organizations and let broken incentives compound unchecked. Uber burned through its full 2026 AI token budget by April. Their COO, Andrew Macdonald, said on the Rapid Response podcast that "if you're not actually able to draw a direct line to how much useful features and functionality you're shipping to your users, the token costs are harder to justify." Salesforce's projected Anthropic bill hit $300 million a year. Marc Benioff started asking for a "smart router" to stop sending trivial queries to frontier models. Microsoft cancelled its own employees' Claude Code licenses and redirected them to GitHub Copilot CLI.

Meta quietly took down its internal tokenmaxxing leaderboard after reports that employees at other companies were caught spinning up meaningless agent tasks just to game consumption stats. Jensen Huang told the GTC audience in March that every Nvidia engineer would need an annual token budget that could reach half their base salary. He wasn't warning anyone. He was planning for it. Flat-rate pricing let all of this happen without friction. Nobody asked whether a prompt was worth $0.47 when the answer was always "it's included." Usage-based billing makes the cost visible. It forces the question engineering teams should have been asking for a year: what is this workflow actually worth?

But predictability matters

The consensus gets one thing right, and it's worth conceding cleanly. Token-based billing is harder to forecast than flat-rate. For teams that built CI pipelines on Claude Code at subscription rates, the mid-stream change is operationally disruptive in ways the pricing math doesn't capture. The enterprise CFO who says "I can't budget for a cost that varies 10x depending on the workday" has a legitimate complaint. Knowing the subsidy was unsustainable doesn't make the transition painless.

The timing deserves criticism too. GitHub has 4.7 million paid Copilot subscribers, up 75% year over year. Waiting until adoption is locked in to flip the pricing model is a textbook platform play: subsidize to capture, then meter to extract. The economic logic can be sound and the sequence can still erode trust. Both things are true at the same time.

Who this changes things for

If you treat AI coding tools the way you treat any other infrastructure cost, with awareness, targeted prompts, and an architecture that doesn't multiply inference overhead across service boundaries, the pricing changes are manageable. I run four agents on a single VPS sharing one PostgreSQL, one Redis, one Node.js process for $7 a month. My Claude Code subscription is $200 a month. June 15 changes my economics too. My publishing pipeline runs claude -p three times a week. I'll audit my token volume, adapt, and keep shipping. The cost is visible now. That's an improvement over pretending it wasn't there.

If you built critical infrastructure on the assumption that flat-rate meant flat-rate forever, the trust argument applies even when the economic argument doesn't. If your enterprise needs pricing predictability for annual budgets, skip the subscription and go straight to the API. And if the subsidy was the entire value proposition, if the tool stops being worth it at real prices, that's an honest finding. It just isn't a scandal. The coding tool platforms that locked us in are now metered platforms. The real question is whether what's inside them justifies the meter.

The vendors stopped undercharging. That's not a crisis. It's a price correction. The question was never what AI coding tools cost to buy. It was always what they cost to run.