On March 16, GitHub discovered a bug in its rate-limiting system. The system had been undercounting tokens from newer models — Opus 4.6, GPT-5.4, and other next-gen releases consumed significantly more infrastructure per request, but the counter was measuring them as artificially lightweight.
They fixed the bug. Then Pro+ users started hitting their weekly limits after consuming roughly 20% of their nominal quota.
The bug hadn't been protecting users from rate limits. It had been hiding the real cost of AI coding from everyone — including GitHub.
48 Hours
Five weeks after the fix went live, three platforms corrected in rapid succession. Not coordinated. Just inevitable.
April 20, Sunday
GitHub paused all new signups for Copilot Pro, Pro+, and Student plans. Joe Binder, VP of Product: agentic workflows "regularly consume far more resources than the original plan structure was built to support." Opus models removed from Pro. Tighter rate limits imposed.
Internal documents: weekly operating costs had nearly doubled since January.
April 21, Monday afternoon
Anthropic quietly removed Claude Code from the $20 Pro plan. No blog post. No email. The pricing page just changed. Access now started at Max ($100/month).
By the next morning, after developer backlash, the page was quietly restored. Head of Growth Amol Avasare: "running a small test on ~2% of new prosumer signups." Then: "our current plans weren't built for this."
April 21, Monday evening
SpaceX announced a $60 billion option to acquire Cursor. The stated prize: combining Cursor's product with Colossus, SpaceX's supercomputer (1 million H100-equivalent). Own the compute. Own the model. Own the tool.
Cursor had been paying Anthropic's retail API rates. At those rates, a heavy user really does cost $5,000/month. SpaceX's answer: stop paying someone else's rates.
The Math That Broke
The flat-fee model for AI coding tools was always a subsidy. The question was how long it would last.
The viral number — $5,000 in compute per $200 Claude Code Max subscriber — is misleading. The actual compute cost is roughly 10% of API pricing. Most users spend ~$6/day in API-equivalent terms, translating to about $18/month in real compute against a $200 subscription. Anthropic confirmed fewer than 5% of subscribers hit weekly caps. The median user is profitable.
But Cursor doesn't have Anthropic's margins. Cursor pays retail. And the 5% tail that burns $500/month in real compute — or $5,000 at API rates — is precisely the cohort that builds the product's reputation.
The deeper problem is structural. Per-token prices have fallen 30-70% over the past year. A reasonable person would expect costs to drop. Instead, GitHub's weekly Copilot costs doubled since January. Agentic workflows — long-running, parallelized, tool-calling sessions — consume 10-15x more tokens than the completions the original pricing was built for. Cheaper tokens, exponentially more of them.
The economist William Stanley Jevons described this in 1865: when coal-burning engines got more efficient, coal consumption didn't fall — it rose. The efficiency invited new uses that overwhelmed the savings. AI coding is living the same paradox. An agent that writes, tests, debugs, and iterates autonomously burns through tokens at a rate that makes per-token discounts irrelevant.
Where It Lands
The market is converging on a price.
| Platform | Flat-Fee Tier | What Changed |
|---|---|---|
| Claude Code Max | $200/mo | Pro ($20) tested as access floor → reversed |
| Cursor Ultra | $200/mo | SpaceX $60B option to own the compute |
| Windsurf Max | $200/mo | Added March 19, matching the new floor |
| GitHub Copilot | $10-39/mo | Token billing in June. Signups paused. |
| Anthropic Enterprise | Per-token | Already moved. Others expected within 6 months. |
GitHub's June transition is the clearest signal. After a three-month promotional window, Business subscribers ($19/month) will get $19 in token credits. Enterprise ($39/month) gets $39. Credits match price. There is no subsidy in that equation.
The enterprise side has already moved. Anthropic shifted to per-token billing. Google and OpenAI are expected to follow within six months. At those rates, a typical 40-turn coding agent session runs about $4. Ten sessions per day across a team: $40/day. Twenty workdays: $800/month per developer. That's not a subscription. That's an infrastructure cost.
Three Responses to One Problem
GitHub chose transparency — expose the real costs, let users manage budgets, shift to token billing. It's honest. It also means developers will watch their spend like they watch cloud bills, and some will use less.
Anthropic chose to test the boundary, then retreat. The Claude Code Pro removal lasted hours before backlash forced a reversal. But Avasare's quote — "our current plans weren't built for this" — isn't a retraction. It's a forecast. The $20 Pro plan with Claude Code access is living on borrowed time.
SpaceX-Cursor chose vertical integration. If the problem is paying someone else's API rates, the solution is to stop. Colossus provides the compute. xAI provides the model. Cursor provides the product. The margin structure changes completely when you own the stack.
All three responses point in the same direction. The question is no longer whether AI coding tools are worth paying for. It's whether the price you're paying reflects the cost of what you're using. For a year and a half, the answer was no. The bill came due in 48 hours.