Anthropic describes extended thinking’s benefits and limits

Anthropic describes extended thinking’s benefits and limits

Anthropic’s page, published February 24, 2025, describes extended thinking in Claude 3.7 Sonnet: users can turn it on or off, and developers can set a thinking budget. Anthropic says the same model spends more time and effort on an answer, with its thought process visible in raw form. The company presents visible thoughts as a way to inspect answers and study model behavior, but warns that they may include incorrect or misleading reasoning and may not faithfully represent why the model behaves as it does. Anthropic says monitoring current models’ thinking cannot support strong arguments about their safety. That gap matters because visible reasoning may look informative without revealing the factors behind a decision. Anthropic reports improved results on evaluations and describes the model using computer controls and playing Pokémon Red. It also reports gains from parallel test-time compute on GPQA, while noting that this approach was not available in the deployed model. These evaluation results do not establish everyday reliability. The post describes safeguards for visible thoughts and computer use, calls visible extended thinking a research preview, and says Anthropic will weigh its trade-offs for future releases. How well can visible thoughts help assess a model’s behavior if they may not represent the factors behind its decisions?

Publication: 9e65f45c-4135-5975-3ec4-236944acb76e

attribution — pass: The reading consistently attributes the descriptions, warnings, benchmark results and safeguard figures to Anthropic's own post about Claude 3.7 Sonnet. The page is a single first-party article, not a listing.

With the new Claude 3.7 Sonnet, users can toggle “extended thinking mode” on or off

summary — pass: The reading leads with the actor and the change: a toggleable extended thinking mode with a developer-set thinking budget and a visible raw thought process. It conveys the post's central tension, that visible reasoning helps trust and research but may not be faithful, and it reports Anthropic's research-preview framing. Test-time compute is defined beside the reading. GPQA is not explained but is not central. The question follows from the source's faithfulness concern. Omitting the ASL-2 determination does not distort the reading's focus.

Another issue is what’s known as “faithfulness”—we don’t know for certain that what’s in the thought process truly represents what’s going on in the model’s mind

date — pass: The stated date, February 24, 2025, matches the UTC day of originalPublished. The page was edited after publication, but the cited passages concern the Claude 3.7 Sonnet launch and are plainly part of the original piece. The limitations disclose the later retrieval and the edit. No hindsight framing appears, and the plans for future releases remain attributed intentions.

We’ll weigh the pros and cons of revealing the thought process for future releases3.

support — pass: Each claim matches the source text: the toggle and thinking budget, the incorrect or misleading thoughts, faithfulness, the safety-monitoring limit, GPQA 84.8% under the stated setup, prompt injection prevention of 88% versus 74%, and the research preview. The line "That gap matters" stays close to the source's own point that decisions rest on undiscussed factors. The everyday-reliability caveat is appropriately cautious.

Thus far, our results suggest that models very often make decisions based on factors that they don’t explicitly discuss in their thinking process.

media — pass: No media was selected, and no verified media candidates were supplied, so no invalid or unverified media is attached.