Updated May 30, 2026. In a single month, three of the largest AI vendors raised what you actually pay, and only one of them did it by changing the number on the pricing page. OpenAI doubled the GPT-5.5 sticker. Anthropic left the Opus 4.7 sticker alone and changed the tokenizer underneath it. GitHub kept its plan prices and swapped the billing model. The per-token list price is now the least reliable line in your AI budget, and this piece walks through why, with the numbers.
For three years the AI market trained everyone to compare list prices. Dollars per million input tokens, dollars per million output tokens, pick the cheaper one. That habit broke in May 2026. The same week, three vendors moved the real cost of a request in three completely different ways, and a buyer reading only the pricing pages would have caught exactly one of the three.
The useful way to think about it: list price is what a vendor advertises per token. Billed cost is what lands on your invoice for getting a job done. Those two numbers used to track each other closely. They no longer do, and the gap between them is now where the entire pricing story lives.
Mechanism one: the honest increase
OpenAI did the straightforward thing. GPT-5.5 launched with input moving from $2.50 to $5.00 per million tokens and output from $15 to $30 per million. A clean doubling, printed on the pricing page, impossible to miss.
Except the doubling is not the real number. According to FairMind's analysis of OpenRouter routing data across millions of calls, GPT-5.5 produces 19 to 34 percent fewer completion tokens than GPT-5.4 did on long prompts, because it is terser. So the real, measured cost increase landed between 49 and 92 percent depending on prompt length, not a flat 100 percent. The sticker said "twice as expensive." The invoice said "somewhere between half-again and nearly twice, and you will not know which until you measure your own prompt shapes."
This is the friendly version of the problem. At least the direction and rough magnitude were visible. The other two were not.
Mechanism two: the invisible increase
Anthropic shipped Opus 4.7 at the same advertised price as its predecessor. Nothing changed on the pricing page. What changed was the tokenizer.
Independent measurements reported by FairMind found that Opus 4.7 emits 32 to 45 percent more native tokens for the exact same input text. Same words in, more billable tokens out, identical price per token. The result, by their measurement, is that teams running Opus 4.7 pay roughly 12 to 27 percent more per task on prompts above 2,000 tokens, while very short prompts come out about 1.6 percent cheaper. The headline they landed on: paying about 27 percent more for the same work at an unchanged sticker.
There is a real mitigation here, and it matters. Prompt caching at the standard 90 percent discount absorbs most of the extra tokens once a prompt crosses the large-context threshold. The same analysis estimated caching soaks up around 93 percent of the additional tokens on prompts above 128K. So the teams hit hardest are the ones running medium prompts (a few thousand tokens) without aggressive caching, which describes a lot of agent and RAG workloads. We cover the caching mechanics separately in the prompt-caching optimization guide.
The point is not whether Anthropic intended this as a price increase. The point is that a tokenizer change is invisible on a pricing page and very visible on an invoice, and no amount of list-price comparison would have surfaced it.
Mechanism three: the structural increase
GitHub kept Copilot Pro at $10 a month and Pro+ at $39 a month. The plan prices did not move at all. What moved, effective June 1, 2026, is the billing model: premium requests are gone, replaced by GitHub AI Credits billed on actual token consumption at each model's per-token rate.
For a developer who lives in tab-completion, this is a non-event, because completions stay free and do not touch credits. For a developer running agent loops on frontier models, it is a different story. The model multipliers do the damage: Opus 4.7's multiplier alone reportedly jumps from 7.5x to 27x. FairMind's read is that agentic users can see effective increases of three to four times. We broke down the credit math in detail in the Copilot usage-based billing breakdown.
Notice the pattern across all three. One raised the sticker (OpenAI). One raised the token count (Anthropic). One raised the conversion between usage and dollars (GitHub). Three different levers, one shared effect, and only the first shows up if you are comparing pricing pages.
What the same task cost, before and after
Take a single medium agent task: roughly 8,000 input tokens, 1,500 output tokens, run a few hundred times a day. Here is how the three mechanisms read on that one unit of work.
| Vendor / change | What the pricing page showed | What measurement showed on the task |
|---|---|---|
| OpenAI GPT-5.5 | Input and output rates doubled | Roughly +49% to +92%, partly offset by fewer output tokens |
| Anthropic Opus 4.7 | No change | Roughly +12% to +27% on prompts over 2K tokens, near flat on short ones |
| GitHub Copilot | Plan price unchanged | Up to 3x to 4x for agent-heavy usage after June 1 |
A team comparing only sticker prices would have budgeted for a doubling on OpenAI and zero change on the other two. The actual exposure was a partial increase on OpenAI and meaningful, unbudgeted increases on both of the "unchanged" vendors.
The only number that survives: effective cost per task
If list price no longer predicts your bill, the fix is not a better pricing comparison. It is measuring the thing that actually varies. That number is effective cost per completed task, and you build it from data you already have access to:
- For every request, log input tokens, output tokens, and cached tokens separately. The split matters, because caching and output verbosity are exactly where the sticker and the bill diverge.
- Multiply each by the model's current per-token rate and sum. That is the true dollar cost of the request, not the advertised one.
- Divide total spend by completed outcomes, not by request count. An agent that retries five times to finish one task costs five requests and produces one unit of value. Cost per request will lie to you the same way list price does.
- Track that number per model and per workload over time. When a vendor changes a tokenizer or a multiplier, your effective-cost line moves even though every pricing page stayed still. That movement is the early warning the pricing page will never give you.
This is the entire reason a metering layer exists. You cannot catch an invisible increase by reading announcements, only by measuring per-token, per-model, per-task spend continuously and watching the effective rate. That is what UsageBox is built to do: capture input, output, and cached tokens per request and surface the effective cost per task before the invoice does. The hidden-cost breakdown goes deeper on the components, and the metering platform comparison covers the tooling options.
The takeaway from May 2026 is not that AI got more expensive. It is that the pricing page stopped being where the price lives. Three vendors proved it in three weeks. Budget for the bill, not the sticker.