You Pay Claude to Think, and It Never Shows You the Work

Branden Bell
8 hours ago
5 min read

Somebody sent me their Claude API usage last month and asked why the output token count was so much bigger than the text they actually got back. I gave them a confident answer about setting a thinking budget, and then I went to pull up the docs to show them where, and the setting I was describing doesn't exist on the model they're running.

So here's the corrected version. A chunk of what you're billed for is Claude thinking before it answers, on the current models you don't see any of it unless you go turn a setting on, and even then what comes back isn't the real thing. Anthropic documents all of this openly, in more detail than most vendors bother with. The numbers below are theirs, verified against the platform docs in August 2026.

What changed in the Claude 5 lineup

The manual thinking budget is gone. If your code still sets one on a Claude 5 model, the request doesn't run.

The old mode was called extended thinking. You passed a thinking object with a budget, Claude reasoned against that budget, then answered. It's deprecated on the Claude 4.6 models, and per Anthropic's own docs, Claude 4.7 and later "do not support it and reject requests that use it, returning a 400 error." The migration list is explicit: Opus 4.7, Opus 4.8, Opus 5, Sonnet 5, Fable 5 and Mythos 5 all refuse it.

What replaced it is adaptive thinking, where Claude decides whether and how much to think on each request. On Opus 5, Sonnet 5, Fable 5 and Mythos 5 it's already on with no configuration at all. Haiku 4.5 is the holdout still running the old mode, which is the split I mention in passing in my guide to using Claude in 2026 and never had room to unpack there.

Depth control moved with it. It used to be a token number you picked. Now it's a word, and the parameter is called effort.

The old control was never a real cap

Before anyone mourns the budget, it's worth knowing what it actually did, because Anthropic says so plainly:

The budget is a target rather than a strict cap.

It had rules. The minimum was 1,024 tokens, and it had to come in under your max_tokens, since thinking counts toward the same output ceiling as the response text. But Claude could stop reasoning well short of the budget, and the budget itself never stopped anything from happening. The hard ceiling was max_tokens the whole time.

Which makes the migration less dramatic than it reads. What you had was guidance with a number attached to it, and what you have now is guidance with a word attached to it.

You are billed on thinking you never read

Here's the part that shows up on the invoice. Reasoning tokens are output tokens, whether or not anything comes back to you:

the tokens Claude spends reasoning are billed as output tokens, even when the thinking text isn't returned to you

And on the newest models, nothing comes back by default. The display field defaults to omitted on Fable 5, Mythos 5, Opus 5, Sonnet 5, Opus 4.8 and Opus 4.7, which returns thinking blocks with an empty thinking field. Anthropic doesn't dress up what that setting saves you. It says you're still charged for the full thinking tokens, and that omitting reduces latency, not cost.

Flip display to summarized and you get a summary, not the reasoning. You're charged for the full thinking tokens from the original request rather than the summary tokens you can actually read, so the billed count and the visible count don't match by design. That summary also gets written by a different model than the one you called, and the model doing the thinking never sees it.

Then there's the sentence that settles it:

No display setting returns the raw chain of thought.

That's not a plan tier or a beta header you can go ask for. The reasoning you're paying for isn't available to read at any price.

What you actually control now

You've got three levers and a meter, and only one of the levers is hard:

Effort, which is guidance: it scales how often and how deeply Claude thinks, and at lower settings Claude may skip thinking entirely on easy inputs.
max_tokens, which is the ceiling: Anthropic draws this line itself, saying effort is soft guidance and max_tokens is a strict limit. If you need a spend ceiling you can trust, this is the only one on the list.
The off switch, where it exists: Sonnet 5 accepts thinking disabled. Opus 5 accepts it at high effort or below and returns a 400 error at the two highest levels. Fable 5 and Mythos 5 reject it outright, so on those models thinking cannot be turned off at all.
The meter: usage.output_tokens_details.thinking_tokens reports how many of your billed output tokens were internal reasoning. When you're streaming it only shows up on the final message_delta event.

That last one is the only real visibility you get, and it's a count, not a transcript.

The question to ask whoever builds with this

I run several agents unattended on schedules and I've never once read their reasoning. I read what they produced and check whether it matches what I asked for. That's just the working condition now.

So a guardrail has to be something other than watching. If you can't inspect how the thing decided, the controls that still mean anything are the ones that don't depend on inspection: a hard ceiling on what it can spend, and a written list of things it may never do no matter what it concludes. The agent I built to run my own brokerage account is the clearest version of that second one I've published, and its never-list is on the build page, free, no email. Same thinking behind the tasks I run while my laptop is asleep.

When you're paying somebody to build with Claude, ask them what the ceiling is and what's on the never-list. Those are answerable questions with specific answers, and not having one is worth knowing before the invoice shows up.

If you want that built properly around your business, that's what I do. Consulting runs $250 for the first hour and $75 an hour after that, with the hours estimated up front. When it's done you own it and it runs in your Claude account, not mine.