You chose the “affordable” AI model. You did your homework — compared the pricing pages, ran the numbers, felt good about yourself. Then your bill arrived. Surprise: you’re paying 28x more than the model you avoided. Welcome to the thinking tokens scam.
The Study Nobody Asked For (But Everyone Needed)
Researchers from Stanford, UC Berkeley, Carnegie Mellon, and Microsoft just published a paper called “The Price Reversal Phenomenon: When Cheaper Reasoning Models End Up Costing More” and it’s the most damning indictment of AI pricing transparency since, well, forever.
They ran 11,872+ queries across 8 frontier reasoning models and 9 task types — competition math, code generation, science QA, multi-domain reasoning. The conclusion: the prices on the marketing page are functionally fiction.
What Are “Thinking Tokens” and Why Should You Be Furious?
Modern reasoning models — the ones with “thinking” or “o” in their names — don’t just answer your question. They work through it internally, step by step, before they respond. That internal monologue? It costs tokens. Lots of them. And they’re billed as output tokens at full price.
Here’s the problem: these thinking tokens aren’t shown in the listed price. They’re not capped. They’re not predictable. The same query, run twice, can produce up to 9.7x more thinking tokens on the second run. You literally cannot budget for this. It’s like hiring a contractor who charges by the hour but won’t tell you how long they’ll think about your project before starting.
Some models spend 97.9% of their output tokens just thinking — before they write a single visible word of your response. You’re paying for their internal monologue. You’re not even reading it.
The 28x Gut-Punch
Let’s get specific, because the abstract number sounds fake until you see the examples.
Take Gemini 3 Flash. Google lists it at roughly 78% cheaper than OpenAI’s GPT-5.2. Sounds like a no-brainer, right? Run the study’s benchmarks and Gemini 3 Flash’s actual total cost across all tasks is 22% higher than GPT-5.2. The cheaper option costs more. Not slightly more. More.
At the extreme end, comparisons between certain model pairs — like Gemini 3 Flash vs. Claude Haiku 4.5 on MMLUPro — showed cost reversals of up to 28x. You picked the model listed at a fraction of the price and ended up paying twenty-eight times the alternative.
This isn’t a rounding error. This is a structural failure in how the entire industry communicates pricing.
The Correlation Problem
The researchers measured how well listed prices predict actual costs using Kendall’s τ (a rank correlation metric). The result was 0.563. For context, 1.0 would mean “listed price perfectly predicts actual cost.” 0.0 means it’s useless. 0.563 is barely better than a coin flip.
When they stripped out thinking tokens from the analysis? Correlation jumped to 0.873. Almost all the pricing chaos is caused by hidden thinking tokens. Remove the hidden variable, and suddenly pricing makes sense again.
The ranking reversals — cases where cheaper-listed models cost more in practice — dropped by 70% without thinking tokens in the mix.
So the providers know. They know their pricing pages are misleading. They’re not fixing it.
Who’s Doing This?
Everyone, basically. OpenAI (GPT-5.2), Google (Gemini 3 Flash, Gemini 3 Pro), and Anthropic (Claude Haiku 4.5, Claude Opus variants) were all tested. This isn’t a “one bad actor” story. It’s an industry-wide habit of listing the base per-token price while burying the variable thinking token cost in fine print — or not mentioning it at all.
The models affected are precisely the ones being sold to developers as “cost-efficient reasoning” solutions. The budget-tier reasoning models are often the worst offenders, because they’re architected to think longer (compensating for smaller parameters) while being priced to look cheap.
What Developers Are Actually Dealing With
Imagine building a product on top of a “cheap” reasoning model. You do a pricing estimate. Your estimate is based on listed prices. You build, you launch, you scale. Then the invoices start coming in and they don’t match your spreadsheet. At all.
This isn’t hypothetical — it’s what every developer using reasoning models is living through right now. The unpredictability isn’t just financial. Even sophisticated cost predictors — using prompt embeddings, length features, and task classification — only reduce prediction error by 23%. You can throw ML at this problem and still lose.
The paper describes thinking token variation as an “irreducible noise floor.” In plain English: you can’t fix this with a better model or a smarter tool. The chaos is baked in.
What You Should Actually Do
The researchers are diplomatic about it. BluntAI is not.
- Never trust the pricing page. Listed prices are for input tokens. Your actual bill includes thinking tokens that aren’t listed, aren’t capped, and vary wildly by task.
- Benchmark on YOUR workloads. The study ran standardized benchmarks. Your use case might behave completely differently. Run 100 real queries, track total token consumption, then price it.
- Use cost monitoring per request. Tools like Portkey or Helicone give you actual per-call cost visibility — not estimates, not averages. If you’re using reasoning models at scale without this, you’re flying blind.
- Set hard token limits. Every major provider lets you cap max output tokens. Use it. Yes, it might cut off thinking mid-thought. That’s better than an infinite bill.
- Compare “total cost on task” not “price per million tokens.” The study proves these numbers are uncorrelated. Stop using them.
The Uncomfortable Conclusion
OpenAI, Google, and Anthropic are all multi-billion dollar companies with teams of people who understand pricing. They know their listed prices don’t reflect actual costs for reasoning models. They know thinking tokens create unpredictable bills. They haven’t fixed the pricing pages.
That’s a choice.
The researchers call for “transparent per-request cost monitoring” and “cost-aware model selection.” These are polite academic ways of saying the providers should tell you what things actually cost before you buy them. Revolutionary concept.
Until that happens — and given the incentives, don’t hold your breath — the burden is on you to measure, monitor, and set limits. The providers have decided that’s your problem. Now you know why.
Source: “The Price Reversal Phenomenon: When Cheaper Reasoning Models End Up Costing More” — Lingjiao Chen, Chi Zhang, Yeye He, Ion Stoica, Matei Zaharia, James Zou (Stanford, UC Berkeley, CMU, Microsoft Research, 2026)