Anthropic describes extended thinking as 'serial test-time compute': sequential reasoning steps before final output
Anthropic introduced extended thinking with Claude 3.7 Sonnet in February 2025, framing it explicitly in the language of test-time compute research. The company's announcement (2025-02-24) states: "When Claude 3.7 Sonnet is using its extended thinking capability, it could be described as benefiting from 'serial test-time compute'. That is, it uses multiple, sequential reasoning steps before producing its final output, adding more computational resources as it goes."
The toggle framing
Anthropic positioned extended thinking not as a separate model but as a different operating mode of the same model: "Extended thinking mode isn't an option that switches to a different model with a separate strategy. Instead, it's allowing the very same model to give itself more time, and expend more effort, in coming to an answer."
This framing distinguishes their approach from parallel sampling approaches: "They do this by sampling multiple independent thought processes and selecting the best one without knowing the true answer ahead of time." Extended thinking, by contrast, is serial — each reasoning step conditions on all prior steps.
The serial / parallel distinction
Two forms of inference-time compute scaling:
| Form | Mechanism | Anthropic term |
|---|---|---|
| Serial | Sequential chain-of-thought steps; each step builds on prior | "Serial test-time compute" |
| Parallel | Multiple independent samples; select best via verifier | "Parallel compute" |
Snell et al. (arXiv:2408.03314, 2024) study primarily parallel approaches; Anthropic's extended thinking is the serial path. See claim-test-time-compute-can-substitute-parameters.
Trajectory post-launch
As of January 2026, Anthropic made extended thinking automatically enabled by default in Claude 4 models, with a default budget of 31,999 thinking tokens. Claude 4 models "return summarized thinking output rather than full thinking tokens, but you're charged for the full thinking tokens with a condensed summary in the response" — a design tradeoff between transparency and cost to the caller.
Relationship to chain-of-thought faithfulness
Extended thinking makes the model's reasoning visible (or a condensed version of it), which intersects with ongoing questions about whether such visible reasoning is causally connected to the model's outputs. The faithfulness of chain-of-thought traces to actual computation is a live research question. See cot-faithfulness-anthropic-biology and introspection-access-problem for the interpretability angle on what this reasoning trace does and does not reveal.
See also: claim-test-time-compute-can-substitute-parameters, claim-ai-inference-means-running-a-model, cot-faithfulness-anthropic-biology, introspection-access-problem
Source
“When Claude 3.7 Sonnet is using its extended thinking capability, it could be described as benefiting from 'serial test-time compute'. That is, it uses multiple, sequential reasoning steps before producing its final output, adding more computational resources as it goes.”