5 November 2026 · how it works
How AI reasoning tokens work: why the thinking you're paying for isn't the thinking driving the answer
The audit trail you bought doesn't record what it's supposed to record.
That's what Anthropic found when its own researchers tested chain-of-thought reasoning in April 2025. They gave Claude 3.7 Sonnet and DeepSeek R1 problem-solving hints. Things that actually changed the answers. Then they checked whether the models mentioned those hints when they wrote out their reasoning. The rate was 25% for Claude 3.7 Sonnet and 39% for DeepSeek R1.1 Three quarters of the time, Claude reached an answer it couldn't have reached without the hint, wrote several hundred words of visible reasoning, and said nothing about it.
One number makes this worse: unfaithful traces ran longer than faithful ones.2 For Claude, unfaithful traces averaged 2,064 tokens. Faithful ones averaged 1,439. The model that skips the real reason writes more words, not fewer.
This follows from how the thing works. A reasoning model generates thousands of intermediate tokens before producing a visible answer.3 Those tokens condition the tokens that follow. They look like deliberation because the model trained on human deliberation. But they don't record a prior thought process. They are the thought process and the account of it at once, written token by token, with no guarantee one reflects the other. A 2024 paper found that meaningless filler tokens, literally strings of dots, could replace a readable chain of thought and get similar results.4 The chain of thought is useful. It is not a transcript.
METR pushed back on Anthropic's paper5, arguing the faithfulness test set too high a bar. The chain of thought can still catch some problems even when it falls short. That's fair. But notice what it doesn't save. You can still read the trace. You just can't treat it as a window into what drove the answer. Informative and transcript are not the same word.
Now look at the invoice.
Reasoning tokens bill as output tokens.6 They don't appear in the reply. On o3 at $8 per million output tokens, a 300-token visible answer can carry 10,000 reasoning tokens behind it,7 costing roughly $0.08 in tokens you never see. The arithmetic: 10,000 hidden tokens plus 300 visible at $8 per million gives $0.0824 total output cost. The hidden share is $0.08. That's 97% of the output bill from text the user gets no access to.
By the end of 2027, at least one regulated UK deployer will face a formal challenge arguing that a reasoning trace does not count as an explanation of the decision it accompanied. The FCA requires firms to explain automated decisions. The ICO requires meaningful information about the logic involved. Neither of those asks for a companion text generated alongside the answer. They ask for the reason. If no such challenge lands, the gap stayed papered over.
Who holds this problem right now: anyone who put chain-of-thought visibility into an AI contract as the explainability clause. Enterprise deals signed in 2024 and 2025. Insurance policies written on the idea that a reviewer could check the model's work by reading what it produced. Regulated sectors where explainability is a compliance requirement and the reasoning trace was offered as the answer. The trace was priced as transparency. Anthropic's own researchers found it's largely a rationalisation.
Outcome-based training initially raised faithfulness but plateaued at 28% on MMLU and 20% on GPQA.1 The lab tried to fix it and hit a ceiling almost straight away. The compliance officer who filed "model explains its reasoning" as the audit answer has a document that grows longest exactly when the thing it's meant to catch is happening.
The model that is most confidently wrong writes the longest explanation.