GitHub paused new signups for Copilot Pro, Pro+, and Student plans on April 20. A week earlier, Opus 4.6 Fast was retired from Pro+. Opus was removed from the Pro tier entirely. A leaked internal plan, surfaced by Ed Zitron, has Microsoft migrating all Copilot subscribers to token-based billing on June 1: Business at $19/month plus $30 of pooled AI credits, Enterprise at $39 plus $70. Last day for refunds is May 20. After August 31, the promo credit window closes and the credits drop to match the seat price.
The third panel of a trilogy locked in over the weekend.
— Joe Binder, GitHub VP of ProductIt’s now common for a handful of requests to incur costs that exceed the plan price.
That’s the closest any of the three vendors has come to telling the truth about what is happening. It is not a feature update. It is an invoice catching up.
The Euphemism Trilogy
Three companies, three plan changes, one underlying fact. None will say “we are raising prices.”
- Anthropic pulled Claude Code from the $20 Pro tier on Apr 21 and characterised it as “running a small test on ~2% of new prosumer signups.” Reversed under pressure. Mythos itself was already framed as a “preview release” gated behind nine-figure partner credit commitments.
- Cursor cut rate limits roughly twentyfold overnight alongside its Cursor 3 agent-first redesign. The framing was “compute limits” and “closer to real cost.” The apology that followed was for “communication,” not the price.
- GitHub paused signups, killed Opus on Pro, and is moving everyone to tokens. The framing is “agentic workflows have fundamentally changed Copilot’s compute demands.”
Every phrase is structural-sounding. Every phrase reframes the user’s existing plan as the thing that became wrong, not the price as the thing that went up. Test. Compute limits. Compute demands. Three euphemisms, one event.
The pattern is the story.
What Actually Changes
The pre-change Copilot Individual lineup, still on the website as of this writing under a “temporarily unavailable” banner:
- Free: $0 / 50 agent requests + 2,000 completions per month
- Pro: $10/month / 300 premium requests
- Pro+: $39/month / 1,500 premium requests
- Overage: $0.04 per premium request
The June 1 model, per the Zitron leak:
- Business: $19/user/month + $30 pooled AI credits (promo through August; reverts to $19 in credits afterward)
- Enterprise: $39/user/month + $70 pooled credits (promo through August; reverts to $39 afterward)
- Per-token rates: GPT-5.4 around $2.50 in / $15 out per million tokens. Opus 4.7 around $5 in / $25 out per million.
The math for a heavy individual user has moved. A modest agentic session burning roughly 50k tokens in and out on Opus 4.7 costs about $1.50. The $30 credit budget covers around 20 of those. The $10 flat Pro plan covered as many as the user could fit in 300 “premium requests,” which the agentic-coding pitch was actively encouraging them to fit a lot of. Heavy users move from $10-20 flat to $40-100+ effective monthly spend. Light users stay roughly flat or save.
There is no path published for how Pro and Pro+ subscribers migrate. Pro is paused. Pro+ is paused. Business self-serve was paused two days later.
The Mid-Term Rug Pull
The interim period (Apr 10 to Apr 30) is the part that should bother anyone on an annual subscription. Opus 4.6 was retired from Pro+ on Apr 10. Opus 4.7’s multiplier was set to 7.5x against Pro+ allowances, vs Opus 4.6’s earlier 3x: roughly 250% more expensive for the same model class. Annual subscribers lost the Opus model they paid for, mid-term, without consent.
Refunds are available. The deadline is May 20. The refund is one-way: signups are paused, so a user who refunds cannot resubscribe. GitHub’s own community FAQ thread on the change sits at 1,082 downvotes against 10 upvotes.
Decide before May 20. The refund window is one-way and signups are paused. If your workflow depends on Opus, the model on Pro is now Opus-shaped on the marketing page and not on the API, and Pro+ is throttling Opus 4.7 hard enough that “premium request” is doing a lot of work as a unit.
The Steelman
Token billing is, in isolation, the more honest pricing model. A “premium request” was a 100x-variable unit: same SKU, same multiplier, wildly different underlying compute depending on which model and how long it ran. Per-token pricing is auditable. Light users stop subsidising agent abusers. Idle background loops stop being a free coupon.
Simon Willison made the case directly: “Coding agents consume a lot of compute… heavy LLM users were burning an order of magnitude less tokens just six months ago.” The flat-rate model was already insolvent at the unit-economic level. Pretending otherwise was the temporary state, not the fix.
The steelman is real. It is also not what GitHub said. GitHub said “compute demands.”
The Pattern
All three vendors are reacting to the same underlying fact. Agentic coding broke the flat-rate-per-seat model originally priced for chat-style autocomplete. The agentic pitch (autonomy, parallelism, long-running tasks) is exactly what makes the customer-acquisition coupon insolvent. The companies that sold the agentic future the loudest are now the ones discovering they cannot afford to sell it at the price they used to acquire users.
What’s notable is that none of them will name it. There is a version of the announcement that says: “We priced this for chat. We sold you on agents. Agents don’t fit the chat price. Here is the new price.” That announcement does not exist. The denial is so consistent across three vendors in two months that it has to be coordinated by genre. The genre is “structural change to compute demands.” The genre is not “price increase.” Word goes around the comms team and the comms team is good at its job.
The flat $20/month coding-AI subscription died in April 2026. Only the obituaries differ.
Closing the Trilogy
I wrote yesterday about Anthropic’s Mythos posture collapsing on a fifteen-day clock, and the trilogy was incomplete then. It is complete now. Three vendors, three plan changes, three euphemisms, one fact. Whatever the new normal looks like, it is not the $10 Pro plan with unlimited Opus that GitHub was selling on Friday. It is also not, anywhere, “we raised the price.”
A handful of requests now incurs costs that exceed the plan price. That sentence is the only honest one in the cycle. It just happens to also be the one nobody at any of the three companies wanted to put on the marketing page.
Related Posts
