Claude's Extended Thinking

When Claude 3.7 Sonnet was released in February 2025, it added an "extended thinking mode" that users could switch on, and a "thinking budget" that let developers cap how long the model reasons before answering. The accompanying post is short and mixes several kinds of material: the rationale for showing the model's reasoning to users, the risks of doing so, capability results from spending more computation at answer time, and a summary of the model's safety evaluation. Anthropic filed it under Announcements rather than as a research paper, and it reports no experiments in detail; the system card is cited for those.

For this library the post matters as a statement of how Anthropic framed visible reasoning at the moment it shipped, including the faithfulness problem that the papers on measuring faithfulness (2023) had raised and that Reasoning Models Don't Always Say What They Think (April 2025) would measure in this same model.

What the post describes

One model, more time

Extended thinking is not a separate model. The post describes it as letting "the very same model" give itself more time and effort on a problem ("Claude's extended thinking"). The feature was available to Pro, Team, Enterprise and API users (footnote 1).

Why the reasoning is visible

The post gives three reasons for showing the raw reasoning (section "The visible thought process"):

• Trust. Reading the reasoning makes answers easier to understand and check.
• Alignment. Earlier alignment research, such as Alignment faking in large language models, used contradictions between a model's private reasoning and its output to identify behavior like deception.
• Interest. Researchers with mathematics and physics backgrounds found the reasoning similar to their own approach to hard problems.

What visibility costs

The same section lists three downsides:

• Less character. The reasoning reads as more detached than Claude's normal replies because Anthropic "didn't perform our standard character training" on it, to give the model maximum leeway in its thinking, including thoughts that turn out wrong or half-formed. The character training referred to is the one described in Claude's Character.
• Faithfulness. Anthropic does not know that the displayed reasoning "truly represents what's going on in the model's mind", and words may not be able to describe why a model behaves as it does. The post reports that results so far "suggest that models very often make decisions based on factors that they don't explicitly discuss in their thinking process," so monitoring current models' thinking cannot support strong safety arguments.
• Safety and security. Visible reasoning might help attackers design jailbreaks. More speculatively, a model that learns during training that its thoughts will be displayed might be pushed to think in less predictable ways or to hide some thoughts.

These concerns would be sharper for more capable future models, so the post calls the visible reasoning a "research preview" and says the trade-off will be weighed for future releases. Footnote 3 suggests a middle path: train the model to be truthful about its reasoning when asked, without displaying it by default.

What it reports

The post's results are presented mainly as charts, and only a few are given as numbers.

• Agentic tasks. On OSWorld, an evaluation of multimodal computer use, Claude 3.7 Sonnet starts somewhat ahead of its predecessor and the gap widens as the number of interaction steps grows (section "Claude as an agent"). The post attributes this to "action scaling", the ability to keep calling tools and adjusting until an open-ended task is done.
• Pokémon Red. Given basic memory, screen input and button-press functions, Claude 3.7 Sonnet progressed further than earlier Sonnet models and won three Gym Leader badges; Claude 3.0 Sonnet did not leave the starting house (section "Claude plays Pokémon").
• Serial test-time compute. Accuracy on 2024 AIME mathematics questions rises roughly logarithmically with the number of thinking tokens allowed, and the model generally stops before using its full budget (section "Serial and parallel test-time compute scaling").
• Parallel test-time compute. Sampling many independent reasoning attempts and choosing among them, by majority vote or with a learned scoring model, raised GPQA performance. With the compute of 256 samples, a learned scoring model and a 64,000-token thinking budget, the model scored 84.8% on GPQA, with 96.5% on the physics subset. This method was not part of the deployed model.
• Safety level. Testing by the Frontier Red Team and Alignment Stress Testing team concluded that the ASL-2 standard under Anthropic's Responsible Scaling Policy remained appropriate. Participants assisted by the model on tasks related to chemical, biological, radiological and nuclear weapons got further than unassisted participants, but every attempt contained critical failures (section "Claude 3.7 Sonnet's safety mechanisms").
• Prompt injection. With new training, a system prompt instructing the model to ignore such attacks, and a classifier, computer-use prompt injections were prevented 88% of the time, up from 74% without these mitigations, with a 0.5% false-positive rate (footnote 4).

Instructions and training choices described

The post contains no principles, but it describes several decisions about the text that shapes or is shown from the model:

• Training withheld. Character training, which shapes the tone of Claude's replies, was deliberately not applied to the reasoning.
• Redaction. In rare cases where the reasoning contains potentially harmful content (child safety, cyberattacks, dangerous weapons), that part is encrypted and users see "the rest of the thought process is not available for this response" instead. The model is not stopped from producing such content in its reasoning.
• System prompt. One of the prompt-injection mitigations is a system prompt with instructions to ignore such attacks.

Limitations and open questions

The post states its own caveats: visible reasoning is a research preview; faithfulness is an active research area and current results are discouraging; and visibility may create incentives to hide thoughts in future models. As an announcement, it gives no methods, sample sizes or uncertainty for most of its results, and refers readers to the Claude 3.7 Sonnet system card. On 3 April 2025, Reasoning Models Don't Always Say What They Think reported that Claude 3.7 Sonnet mentioned a hint it had used in 25% of cases on average.

Research questions

1. Does withholding character training from the reasoning affect its faithfulness? The post says the reasoning was left untrained for character to give the model leeway. Comparing faithfulness measures between reasoning trained and not trained for style would show whether training the tone of reasoning changes what it reveals.
2. Does knowing that reasoning is displayed change it? The post raises this as a speculative risk. Comparing reasoning from models trained with and without the reasoning being visible to graders or monitors would test it.
3. Is the "truthful when asked" alternative workable? Footnote 3 proposes hiding reasoning by default but training the model to report it truthfully on request. Measuring agreement between a model's on-request summary of its reasoning and the hidden reasoning itself would show whether such reports can be relied on.
4. How does the thinking budget relate to faithfulness? Accuracy rises with thinking tokens. Measuring hint verbalization as the budget varies would show whether longer reasoning is more or less likely to state the factors behind an answer.
5. How often is reasoning redacted, and what does redaction hide from monitoring? The post aims for redaction to be rare. Reporting its frequency, and whether redacted segments differ in safety-relevant content, would show the cost of redaction to oversight.

How it connects

• Faithfulness research. Measuring Faithfulness in Chain-of-Thought Reasoning and Question Decomposition (2023) set out the tests behind the post's caution; Reasoning Models Don't Always Say What They Think measured faithfulness in this model.
• Reading thoughts without the transcript. Natural Language Autoencoders (2026) reads activations directly, and reports cases where internal states differ from what the visible reasoning says.
• Character and safety work. Claude's Character describes the training not applied to the reasoning; Constitutional Classifiers is cited for preparation toward ASL-3 safeguards; Developing a computer use model introduced the computer-use capability evaluated on OSWorld.

References

• Anthropic. (2025, February 24). Claude's extended thinking. https://www.anthropic.com/research/visible-extended-thinking
• Chen, Y., Benton, J., Radhakrishnan, A., et al. (2025). Reasoning Models Don't Always Say What They Think. arXiv:2505.05410.
• The Claude 3.7 Sonnet system card, OSWorld and GPQA are cited in the post and were not read for this note.