Summary: Claude 3.7 Sonnet introduced extended thinking mode, letting one model reason for a developer-controlled token budget before answering. Anthropic also displays the raw thought process, although it warns that visible reasoning is not guaranteed to be faithful. The mode strengthens long-horizon agent work while raising cost, security, and safety-evaluation questions.
The Change: One Model That Thinks as Long as You Let It
With Claude 3.7 Sonnet, Anthropic shipped a capability named “extended thinking mode.” Earlier models committed to an answer after a fixed amount of computation; this one can keep reasoning before replying for as long as you allow. The mode switches on or off at will, and a “thinking budget” hands developers direct control over precisely how many tokens go into thinking through a problem.
No second model sits behind this. It is a single model that can pour extra cognitive effort into a task whenever the task calls for it. For anyone building products, that translates into per-request control: spend more cost and latency, receive better answer quality. Leave thinking off — or set a tiny budget — when a request is simple; raise the budget when the reasoning gets demanding and stronger results matter.
Under the Hood: Thoughts You Can Actually Read
Anthropic’s most unusual decision here is displaying Claude’s reasoning exactly as it occurs, in raw form. With the mode enabled, the model’s internal steps are laid out ahead of the final answer. Three payoffs justify the choice: you can audit the reasoning yourself, which builds trust; alignment researchers gain visibility into places where thoughts and outputs contradict each other; and frankly, following along is just interesting.
That openness carries a price, though. Because the thinking text never went through character training, it reads colder and less personal than the model’s normal replies. Half-formed and flat-out wrong ideas appear in it too, since the model is working through several angles at once. The deeper problem, as Anthropic points out, is “faithfulness”: what gets written down doesn’t necessarily capture the true factors behind the model’s decisions. Their research found that models frequently act on considerations their thinking never spells out — which rules out leaning on visible thoughts as either a safety or an explanation tool.
Security raises another worry: an attacker could mine the exposed reasoning for material to build more effective jailbreak strategies. For that reason Anthropic labels the visible thought process a research preview, and whether later versions keep it public will come down to a fresh weighing of costs and benefits.
Where It Shines: Agent Work and Past That
Harder questions are only part of the story. Extended thinking also strengthens Claude as an agent — Anthropic’s term is “action scaling.” The model can call functions over and over, adjust to whatever the environment throws at it, and stay with an open-ended task until it is actually finished.
Computer use makes this concrete. Claude issues virtual mouse clicks and keyboard presses to complete tasks on a user’s behalf. On OSWorld, a benchmark built for multimodal AI agents, Claude 3.7 Sonnet only edges out its predecessor at the start — yet the margin keeps growing as the step count climbs, evidence that the model can hold its effort together across longer interactions.
Then there is the entertaining data point: playing Pokémon Red. Anthropic wired Claude up with screen pixel input, basic memory, and function calls for pressing buttons. Earlier models stalled almost immediately — Claude 3.0 Sonnet never made it out of the house in Pallet Town. Claude 3.7 Sonnet, by contrast, fought through three Gym Leaders and claimed their Badges. Holding focus long enough to finish an open-ended goal is exactly the quality building advanced AI agents demands.
The practical upshot for developers is agents that can carry longer, more complicated assignments without dropping the thread. The lever, once again, is the thinking budget: leave the model room to reason, but keep it tight enough that costs never spiral out of control.
What to Watch Out For
The performance gains are real but not unlimited. Accuracy rises along a logarithmic curve, so each extra token of thinking budget buys less than the one before it. On the 2024 AIME math problems, for instance, accuracy does climb as thinking tokens increase — yet the model frequently finishes without spending its entire budget.
A second line of experimentation was parallel test-time compute: sample multiple independent thought processes at once, then pick the winner by majority vote or with a learned scoring model. Run against the GPQA benchmark (biology, chemistry, physics) with the compute of 256 independent samples, a 64k-token thinking budget, and a scoring model in the loop, Claude 3.7 Sonnet scored 84.8% overall and 96.5% on physics. The catch: the technique remains at the research stage and has not reached the deployed model.
Safety evaluation confirmed the current ASL-2 level is still the right one. Even so, Anthropic recorded “uplift” on CBRN-related tasks: participants working with model assistance got further than participants limited to online information. Critical failures blocked every attempt, but the capability gain reads as a signal — so Anthropic is speeding up deployment of targeted classifiers and monitoring while preparing in case ASL-3 becomes necessary later.
The visible thought process has a limitation of its own. In rare cases the raw thinking may carry potentially harmful content. Anthropic’s handling is to encrypt those parts and show a message in their place. The model still puts those thoughts to use internally; users simply never see them.
What Builders Should Actually Do
Products that lean on complex reasoning or long-horizon agent tasks should treat the thinking budget as a first-class knob. Start by benchmarking several budget levels against your own tasks and locate the sweet spot among cost, latency, and quality. More thinking is not automatically better — measure before you believe it.
Think twice before exposing the thought process to end users. Its faithfulness is not guaranteed, which disqualifies it as a safety or explanation mechanism. If you do surface it, present it as a research preview so users know what they are looking at.
And keep safety on your radar. As models grow more capable, they may demand stronger safeguards. Follow Anthropic’s updates and periodically ask how your product would need to adapt.
The bottom line: extended thinking in Claude 3.7 Sonnet marks a significant step forward for AI agents and complex reasoning. Grasp both how it works and where it falls short, and your products can put it to effective use.