Why "Turn On Thinking" Is Now an Error Message

For about a year, using AI well meant making two decisions: pick a model, then decide whether to switch on "thinking" — the mode where the model reasons through a problem before it answers.

That advice is now wrong, in a way you can check in about ten seconds. On Claude Sonnet 5 and Opus 5, asking for extended thinking manually does not get you deeper reasoning. It gets you an error.

What actually changed

Anthropic is blunt about it in its own documentation. Claude Sonnet 5 runs with adaptive thinking on by default, and manual extended thinking is removed — it returns a 400 error, the same as on Opus 4.8 and Opus 4.7. The same release also stopped accepting temperature, top_p and top_k. The knobs did not get better. They got taken away.

What is there instead is a single parameter: effort, with five levels — low, medium, high, xhigh, max. High is the default, and setting high is exactly identical to setting nothing at all.

The part that matters is what effort touches. It is not a thinking dial. Per Anthropic's docs it affects all tokens in a response: the answer, the thinking, and the tool calls. Turn effort down and the model does not just think less — it makes fewer tool calls, skips the preamble, and gets terser. Turn it up and it explains its plan first, calls more tools, and writes you a summary of what it did.

It is also a suggestion rather than a ceiling. The docs describe effort as a behavioral signal, not a strict token budget: at low effort on a genuinely hard problem, Claude still thinks, just less than it would at high. You cannot accidentally switch off its brain.

The one difference worth knowing

Both labs now use the same five words. They disagree about where to start.

Claude Sonnet 5 and Opus 5 default to high.
OpenAI's GPT-5.6 defaults to medium — and adds a none level below low, which Claude does not offer.

That is not trivia. It means "just use the defaults" buys you meaningfully more deliberation on Claude than on ChatGPT — and if you have been comparing the two head-to-head on default settings, you have not been comparing like with like.

The framework

Here it is, and it is shorter than you would expect: model tier tracks how much judgment the task needs. Effort tracks how expensive it is to be wrong.

Those are independent, and collapsing them into a single "how hard is this?" question is the mistake almost everyone makes. The genuinely useful question is a different one: if this answer were subtly wrong, when would I find out?

Immediately. Code that either compiles or does not. A number you will sanity-check. Copy you are going to read line by line anyway. Low or medium effort — you are the verification step, so do not pay a model to be one too.
Eventually. A report someone acts on next week. An analysis feeding a decision. High effort, the default. Leave it alone.
Maybe never. Architecture you will build on for months. A ten-step chain where step seven quietly depends on step three. Anything where you are the only reviewer and do not know the subject well enough to catch a subtle error. Xhigh or max, on the best tier you have access to.

That last row is where people systematically underspend, for an uncomfortable reason: a confidently wrong answer looks exactly like a right one.

Two footnotes that save real time. On Opus 5, effort controls thinking volume but not visible response length — if you want shorter answers, ask for shorter answers, because turning effort down will not do it. And if you depend on prompt caching, choose an effort level at the start of a conversation and keep it there; changing it mid-thread throws the cache away.

The honest shortcut

Most people need none of this. Run everything on your mid-tier model at the default, and escalate only when an answer actually disappoints you. Model and effort selection works far better as a reaction to a bad result than as a prediction you make in advance — and the minutes spent staring at the dropdown usually cost more than simply re-running the task one tier up.

Control	Claude (Sonnet 5 / Opus 5)	OpenAI (GPT-5.6)
Effort levels	low · medium · high · xhigh · max	none · low · medium · high · xhigh · max
Default	high	medium
Manual thinking toggle	Removed — returns a 400 error	Not exposed; effort is the control
What effort changes	All tokens — answer, thinking and tool calls	How much reasoning is applied in the chosen mode
Lowest setting still reasons?	Yes — low still thinks on hard problems	No — none exists as a latency baseline
Extra lever	Opus 5 only — disabling thinking is blocked at xhigh/max; Sonnet 5 has no such limit	reasoning.mode: standard or pro

The tell that this shift is real and not cosmetic: Anthropic now describes Claude Fable 5's lower effort settings as often exceeding xhigh performance on previous models. The floor keeps rising underneath everyone.

Which is the actual lesson here. The winning habit was never memorizing this year's settings — it is noticing when the question itself has changed.