Anthropic announced Claude 3.7 Sonnet on February 24, 2025. The release introduces an opt‑in “extended thinking mode” that lets users instruct the same underlying model to spend more time and compute on harder problems. Developers can also set a “thinking budget” to control how long Claude is allowed to reason for a given task.
Extended thinking does not switch to a different model or strategy; rather, it enables the same model to perform more sequential reasoning steps before producing its final answer. Anthropic says this capability boosts performance but also raises questions about how models are evaluated and how their safety should be managed.
Making the thought process visible
Claude 3.7 Sonnet can expose its internal reasoning in raw form. Anthropic lists several advantages:
- Trust: observing the thought process can help users understand and verify outputs.
- Alignment: previous Alignment Science work used mismatches between a model’s internal thoughts and its external outputs to flag concerning behaviors such as deception.
- Interest: researchers with math and physics backgrounds have noted that Claude’s visible reasoning—exploring branches, checking alternatives—can resemble human problem solving.
There are also drawbacks. The revealed thoughts often sound less personal because the internal outputs were not subject to the usual character training; they can include incorrect, misleading, or half‑formed lines of reasoning. Anthropic also stresses a faithfulness concern: it is not certain that the displayed words accurately describe the real causal factors behind a model’s behavior. Their research indicates models frequently make decisions based on factors they do not explicitly articulate in their thought traces, so monitoring current thought outputs is not sufficient to draw strong safety conclusions.
Finally, visible thought traces create security risks. Malicious actors could exploit them to devise improved jailbreaks, and there is a speculative risk that models trained to have their internal thoughts exposed might change how they think or attempt to hide thoughts. Anthropic says these concerns will be weighed for future, more capable models; for now, the visible thought process in Claude 3.7 Sonnet is a research preview.
New tests and agentic capabilities
Claude 3.7 Sonnet benefits from what Anthropic calls “action scaling”: improved ability to iteratively call functions, react to environmental changes, and continue until an open‑ended task is complete. One practical example is virtual computer use: Claude can generate virtual mouse clicks and keyboard presses to complete tasks on a user’s behalf, and the Sonnet model can allocate more turns, time, and compute to such tasks than previous versions, often yielding better results.
This improvement is reflected in OSWorld, an evaluation of multimodal AI agents: Claude 3.7 Sonnet begins with an advantage and the performance gap grows with continued interaction compared with earlier Sonnet versions.
Playing Pokémon as a demonstration
Anthropic demonstrated the improved agentic behavior by having Claude play the Game Boy classic Pokémon Red. They provided basic memory, pixel screen input, and function calls for button presses so Claude could play continuously beyond typical context limits, sustaining tens of thousands of interactions. Earlier Sonnet versions stalled very early (Claude 3.0 Sonnet reportedly couldn’t even leave the house in Pallet Town), whereas Claude 3.7 Sonnet advanced significantly, defeating three Gym Leaders and earning their Badges. According to Anthropic, the model’s tendency to try multiple strategies and question assumptions helped it progress.
While Pokémon is a playful demonstration, Anthropic says the same abilities—maintaining focus and accomplishing open‑ended goals—have practical applications for building state‑of‑the‑art AI agents.
Serial and parallel test‑time compute scaling
When using extended thinking, Claude effectively performs serial test‑time compute: it takes multiple sequential reasoning steps and increases computation as it goes. Generally, performance improves predictably; for example, accuracy on math problems increases roughly logarithmically with the number of allowed “thinking tokens” (Anthropic illustrates this with results on the American Invitational Mathematics Examination 2024 problems).
Researchers are also experimenting with parallel test‑time compute: sampling multiple independent thought traces and selecting the best answer without access to the ground truth. Selection methods include majority/consensus voting, using another language model (for example, a second Claude) to check outputs, or a learned scoring model. Similar strategies have appeared in evaluations of other AI systems.
Anthropic reports striking gains on the GPQA benchmark (a set of difficult biology, chemistry and physics questions). Using the equivalent compute of 256 independent samples, a learned scoring model, and a maximum 64k‑token thinking budget, Claude 3.7 Sonnet achieved a GPQA score of 84.8%, with a physics subscore of 96.5%. Anthropic notes continued benefit from scaling beyond majority‑vote limits. They also emphasize that parallel test‑time compute is not yet available in the newly deployed model, but remains an active research direction.
Safety mechanisms for Claude 3.7 Sonnet
Under Anthropic’s Responsible Scaling Policy, models are trained and deployed only when appropriate safety and security measures are in place. Anthropic’s Frontier Red Team and Alignment Stress Testing teams ran extensive evaluations to determine whether Claude 3.7 Sonnet required the same AI Safety Level (ASL‑2) safeguards as earlier models or stronger protections.
Their comprehensive assessment concluded that the ASL‑2 standard remains appropriate, although the model demonstrated increased sophistication across domains. In controlled studies of tasks related to Chemical, Biological, Radiological and Nuclear (CBRN) processes, Anthropic observed some performance uplift among model‑assisted participants versus unassisted participants—participants got further than they would using only publicly available information—but all attempts contained critical failures that blocked end‑to‑end success. Expert red‑teaming produced mixed feedback: some experts saw knowledge improvements in certain CBRN areas, but the frequency of critical failures was too high to allow reliable end‑to‑end completion.
Anthropic is accelerating development and deployment of targeted classifiers and monitoring systems to strengthen ASL‑2 safeguards, and says future model capabilities might require upgrading to ASL‑3. Work on Constitutional Classifiers and other jailbreak mitigations is intended to support meeting ASL‑3 requirements when needed.
Handling visible thought content and encryption
Even at ASL‑2, the visible extended thinking feature requires safeguards. In rare cases Claude’s thought process may include potentially harmful content (topics such as child safety, cyberattacks, and dangerous weapons). Anthropic will encrypt the relevant portion of the thought trace in such cases: the model may still reason about that content internally (which can be important for producing a benign final response), but the user will see the message “the rest of the thought process is not available for this response” instead of the sensitive internal text. Anthropic intends to apply this encryption rarely and only when the potential for harm is high.
Computer use protections
Anthropic has also strengthened defenses for Claude’s computer‑use capability against prompt injection attacks—situations where a malicious party hides instructions in content the model might see while using a computer. With new training to resist prompt injection, a revised system prompt instructing the model to ignore such attacks, and a classifier that flags potential injections, Anthropic reports preventing 88% of prompt injection attempts with these mitigations, up from 74% without them. The system currently produces a 0.5% false‑positive rate (the safeguards trigger when no prompt injection is present); Anthropic is working to reduce that rate.
Anthropic provides a fuller summary of safety work, analytic results and examples in its System Card.
Availability and feedback
Claude 3.7 Sonnet is available now at Claude.ai and via Anthropic’s API. The extended thinking and visible thought process options are available to Claude Pro, Team, Enterprise, and API users. Anthropic invites feedback to feedback@anthropic.com.
