On Wednesday, April 22, GitHub paused new signups to Copilot Pro and removed Opus models from the tier. The company said it plainly on the corporate blog: agentic workflows changed Copilot’s compute demand, and the long, parallel sessions consume far more than the original plans were built to carry.

Anthropic moved the same day, but through the back door. It pulled Claude Code from the $20 monthly Pro plan, labeled the change a “2% test,” and let pricing pages, support articles, and customer-facing comms reflect it as if the move were global. Simon Willison published the token math that night, The Register covered it the next morning, and only then did Anthropic reverse.

Two days earlier, xAI had taken the third route. Grok 4.3 Beta launched directly behind the $300 monthly SuperGrok Heavy tier with no entry-level way in. TechSifted noted in its launch review that a year ago $20 a month bought access to xAI’s frontier model; today the same amount only gets you an older one.

What broke that week was not the subscription. It was the business model behind the subscription.

The math that no longer pencils out

Flat-rate pricing worked when AI was autocomplete. You paid your vendor $20 a month for in-editor suggestions, the vendor charged less than the model cost, and the gap was covered by 70% of users who barely opened the product. Same playbook as a gym membership: sell more cards than the room can hold, count on half the holders not showing up.

The playbook breaks when the product moves from autocomplete to an agent that runs overnight on a VM. These are not three-token suggestions anymore; they are six-hour sessions with hundreds of thousands of input and output tokens, tool calls, file reads, and code execution inside a sandbox. Victor Waenerlund did the arithmetic on LinkedIn this week: some plans are charging less than the book value of consumed tokens by up to 10x. The 70% of casual users no longer subsidize anything when the active 30% consume ten times what the plan charges.

Anthropic’s Managed Agents launch at $0.08 per session-hour two weeks ago was the first signal that runtime was moving to variable pricing. This week is the second signal: the developer-facing layer is moving too. OpenAI’s move to $5/$30 per million tokens for GPT-5.5 was the third. All three point at the same ceiling. Flat-rate pricing as an adoption subsidy is over.

The pattern is not new. The speed is.

I have been in this business for thirty-six years. I started in 1990, age fifteen, programming on a Commodore 64 and a Texas Instruments. Since then I have watched the “subsidize adoption, then lock the customer in, then charge the real cost” sequence run so many times I can almost call the month it breaks.

The first cycle I watched up close was long-distance, in the late nineties. International calls were billed by the minute and worked as a barrier for small businesses to even consider going abroad. The digital telcos that came in to break that model started with flat plans, then “unlimited night minutes,” then a flat tariff for everything. Calling looked free. Three years later, the billing category was gone entirely; everything had moved to mobile data, metered to the megabyte.

Unlimited SMS was the next one to fall. It held up until the iPhone made every user send a thousand times more messages in a month than they used to send in a year. Carriers responded with caps, then with per-message tiers, then by bundling SMS into data plans where the picture-messaging app effectively cost more per send than three voice calls.

AWS came along around 2008 with its “free forever” tier: twelve gigs of S3 free, the first EC2 instance free, small applications hosted without ever pulling out a credit card. Ten years later, the same customer that started free pays six figures a month between spot pricing, reserved instance commitments, and committed-use discounts. Nobody feels conned. Nobody remembers the original promise either.

The most recent one is SaaS with unlimited seats. Slack, Notion, and Asana launched with “unlimited” plans because they wanted viral adoption. In four years, every one of them had moved to per-seat with volume discounts, and in the last twelve months the conversation has shifted again, this time to per-message-handled, per-document-processed, per-action. The billing unit keeps coupling tighter to actual usage.

Each of those cycles took four to seven years. The cycle GitHub, Anthropic, and xAI just ran took nine months. The shape of the curve is the same. The time between peaks just collapsed.

The real damage is not the price. It is the opacity.

The best line in the Victor Waenerlund thread did not come from him. It came in the comments from Silvia Adlesic Holmgren, a European pricing consultant: “Buyers do not object to paying more. They object to feeling played.”

That sentence captures why Anthropic reversed and GitHub did not, even though both moves had the same underlying economic justification. GitHub published the why on the same day, on its own blog, with numbers: agentic workloads consume more, the previous plans were not sized for that, new signups pause while the team resizes. Anthropic tried to make the same change without saying it: updated the pricing page, updated the support docs, hid it behind “this is a 2% test.” The price itself was not the problem; the how was.

Price integrity is an asset that weighs more than price level. Once a customer feels that the vendor is changing rules quietly, every future “we are simplifying our plans” email reads as a warning. Anthropic just paid several months of trust for what amounts to a few-week change in quarterly revenue. GitHub paid less. xAI put the sticker on the front page and owes nobody an explanation.

The same Monday this happened, Jaya Gupta posted a 2-million-view essay on X arguing that experience is now a tax, that senior CIOs hide behind “judgment and taste” because they cannot afford to be publicly wrong. Three vendors just demonstrated that the CIOs who signed multi-year contracts on flat-rate plans six months ago are about to learn which of their judgments was real pattern recognition and which was just protecting an old decision. That conversation deserves its own post, and tomorrow’s enters there.

The conversation your CFO should be having this week

If your company signed AI contracts in the last twelve months expecting fixed cost per seat, five fronts are worth closing before the end of the month.

The first one is concentration. How much of your AI spend really sits with a single vendor? If 80% goes to one of them, a single pricing decision moves your budget before you have time to react. With spend distributed across three or four providers, you keep negotiating room.

After concentration comes real usage versus the plan limit. A team using its plan at 30% capacity probably saves money switching to consumption pricing. If usage is already at 200% because the agents are heavy, a forced upgrade is in your pipeline whether you know it or not; better to plan it now.

Third front, the contract clauses themselves. Some enterprise deals carry a twelve-month price lock at the current rate. Others quietly include “the vendor reserves the right to update pricing with 30 days notice.” The difference between those two clauses is the difference between a controlled hit and a fire drill across procurement, finance, and engineering.

The fourth is the exit path if the vendor reprices overnight. Are your prompts, skills, and workflows portable to a harness you own, or do they live entirely inside the vendor SDK? Owning the harness stopped being overengineering six months ago; today it is insurance.

The fifth is the easiest to surface and the one almost nobody checks: the vendor’s recent record with refunds or credits the last time they repriced. GitHub offered refunds inside the cancellation window. Anthropic announced none. That information sits in public threads, and it is worth checking before the next renewal signature lands on the desk.

Those five fronts are the spine of the audit we run as the first step of AI Maestro. It is not a new service. It is the same product we have been building, now with a use case that used to be optional and just became urgent.

What we do at IQ Source about this new runtime price

IQ Source does not sell a plan-comparison dashboard or a cost-tracking tool. That is not what the buyer needs. What the buyer needs is an informed conversation before the next renewal signature.

Two service lines touch this moment from different sides. AI Maestro is the discovery layer: we sit with your team and map where the spend really is, which workloads are sensitive to a price move, which contract clauses expose you, and what portability exists across vendors. The output is a written map for the next renewal negotiation, not a PDF that goes into a folder. Technology Partner applies when your own software company has pricing that depends on an upstream vendor that just moved: there the conversation is about margin redesign, not just an audit.

In both lines the principle is the same. The analysis stays inside the company when the contract ends. The difference from a vendor selling an optimization dashboard is that the dashboard vendor will also reprice when their flat-rate era runs out. Same curve.

If your next conversation with the CFO includes the phrase “we do not know how exposed we are to a single AI vendor,” that is the conversation. It is the first step of AI Maestro: two hours with your team on the contract and consumption portfolio; the output is a written map plus three concrete moves. No quote attached. The email is info@iqsource.ai.

The new price is not the problem. The opacity is the problem. The vendors that survive the next two quarters will be the ones that reprice in public, with numbers, on the same blog where they announced the original plan. The rest will lose not customers, but trust, which is the most expensive asset to recover in a market where the cost of switching providers is dropping every month.