What Changed: A Model That Can Think Longer
Anthropic’s Claude 3.7 Sonnet introduces a feature called “extended thinking mode.” Unlike previous models that give an answer after a fixed amount of computation, this mode lets the same model spend more time reasoning before responding. You can toggle it on or off, and developers can set a “thinking budget” that controls exactly how many tokens the model uses to think about a problem.
This is not a separate model with a different strategy. It’s the same model, but with the ability to allocate more cognitive effort when needed. For product builders, this means you can now trade cost and latency for answer quality on a per-request basis. If a task is simple, you can keep thinking off or set a low budget. For complex reasoning, you can increase the budget to get better results.
How It Works: The Visible Thought Process
One of the most distinctive choices Anthropic made is to show Claude’s thought process in raw form. When extended thinking is on, you can see the model’s internal reasoning steps before it gives a final answer. This has several benefits: it builds trust by letting you check the reasoning, it aids alignment research by exposing contradictions between thoughts and outputs, and it’s simply interesting to watch.
However, this transparency comes with trade-offs. The thinking text is not character-trained, so it sounds more detached and less personal than normal outputs. It may also contain incorrect or half-baked ideas, because the model is exploring multiple angles. More importantly, Anthropic notes that the thought process is not necessarily “faithful”—it doesn’t always reflect the true reasons behind the model’s decisions. Their research shows that models often decide based on factors they don’t explicitly mention in their thinking. So you can’t rely on the visible thoughts as a safety or explanation tool.
There are also security concerns. Malicious users might use the visible thoughts to craft better jailbreak strategies. Because of this, Anthropic treats the visible thought process as a research preview. Future models may or may not have this feature, depending on a cost-benefit analysis.
Practical Use Cases: Agents and Beyond
Extended thinking isn’t just about answering harder questions. It also improves Claude’s ability to act as an agent. Anthropic calls this “action scaling”: the model can iteratively call functions, respond to environmental changes, and keep going until an open-ended task is done.
A concrete example is computer use. Claude can issue virtual mouse clicks and keyboard presses to solve tasks on a user’s behalf. On the OSWorld benchmark, which tests multimodal AI agents, Claude 3.7 Sonnet starts slightly better than its predecessor, but the gap widens as the number of steps increases. This shows that the model can sustain effort over longer interactions.
Another fun but telling example is playing Pokémon Red. Anthropic gave Claude screen pixel input, basic memory, and function calls to press buttons. Previous models got stuck early—Claude 3.0 Sonnet couldn’t even leave the house in Pallet Town. Claude 3.7 Sonnet, however, managed to battle three Gym Leaders and win their Badges. This demonstrates the ability to maintain focus and achieve open-ended goals, which is exactly what you need for building advanced AI agents.
For developers, this means you can build agents that handle longer, more complex tasks without losing track. The key is to use the thinking budget wisely: give the model enough room to reason, but not so much that costs spiral out of control.
Limitations and Trade-offs
While extended thinking boosts performance, it’s not a magic bullet. The accuracy improvement follows a logarithmic curve: you get diminishing returns as you increase the thinking budget. For example, on 2024 AIME math problems, accuracy improves with more thinking tokens, but the model often doesn’t use the entire budget.
Anthropic also experimented with parallel test-time compute, where multiple independent thought processes are sampled and the best answer is selected via majority vote or a learned scoring model. On the GPQA benchmark (biology, chemistry, physics), using 256 samples, a 64k-token budget, and a scoring model, Claude 3.7 Sonnet achieved 84.8% overall, with physics at 96.5%. However, this technique is still in research and not available in the deployed model.
On the safety side, Anthropic’s evaluation confirmed that the current ASL-2 safety level is still appropriate. But they observed “uplift” in CBRN-related tasks: participants with model assistance got further than those using only online information. All attempts still had critical failures, but the capability increase is a signal. Anthropic is accelerating deployment of targeted classifiers and monitoring, and preparing for possible ASL-3 in the future.
Another limitation is the visible thought process itself. In rare cases, the thinking might include potentially harmful content. Anthropic encrypts such parts, showing a message instead. This encryption doesn’t stop the model from using the thoughts internally, but it hides them from users.
Concrete Takeaway for Builders
If you’re building products that need complex reasoning or long-horizon agent tasks, the thinking budget is a powerful knob. Start by testing different budgets on your specific tasks to find the sweet spot between cost, latency, and quality. Don’t assume that more thinking is always better—measure it.
Also, be cautious about exposing the thought process to end users. Since it’s not guaranteed to be faithful, it’s not suitable as a safety or explanation mechanism. If you do show it, frame it as a research preview and set expectations accordingly.
Finally, keep an eye on the safety implications. As models get more capable, they may require stronger safeguards. Stay informed about updates from Anthropic, and consider how your product might need to adapt.
In summary, Claude 3.7 Sonnet’s extended thinking is a significant step forward for AI agents and complex reasoning. By understanding how it works and its limitations, you can leverage it effectively in your products.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.