Anthropic's Claude 3.7 Sonnet merges fast and deep reasoning in one model

Claude 3.7 Sonnet is Anthropic's first hybrid reasoning model, uniting switchable standard and extended thinking modes in a single model and serving as the engine for the new agentic coding tool Claude Code.

Anthropic's Feb. 24 release of Claude 3.7 Sonnet collapses a fragmented reasoning-model landscape into a single offering. Rather than forcing developers to route simple and difficult queries to separate models, the company says the new model handles both fast responses and extended deliberation, with an API-level thinking budget that lets customers buy only as much reasoning as a task requires. The model launches alongside Claude Code, a research-preview command-line agent that can work through a software repository from initial navigation to pushing a commit.

The release lands at a moment when frontier labs are competing on both raw capability and agentic tooling. Anthropic is positioning Claude 3.7 Sonnet as the first model that makes reasoning a switchable property of a single system. The distinction matters to enterprise buyers, who have grown wary of maintaining multiple model endpoints and routing layers for different task types.

Why did Anthropic build a hybrid reasoning model?

Frontier labs have largely treated reasoning as a separate product tier, releasing distinct models for quick chat and for step-by-step problem solving. Anthropic argues that split is artificial. The company said: “Just as humans use a single brain for both quick responses and deep reflection, we believe reasoning should be an integrated capability of frontier models rather than a separate model entirely.” That framing positions Claude 3.7 Sonnet as a consolidation play in a market crowded with specialized reasoning variants.

The hybrid design also reduces operational complexity for developers. Teams no longer need to maintain separate prompts, model routing rules, and fallback logic for fast and slow modes. AWS, which distributes the model through Amazon Bedrock, described the approach as integrating reasoning seamlessly with other capabilities rather than separating quick responses from deeper thought.

What is new in Claude 3.7 Sonnet?

Claude 3.7 Sonnet is available immediately and is Anthropic's most intelligent model to date, according to the company. The model has a knowledge cutoff of October 2024 and was released Feb. 24, 2025. Its defining feature is a switchable reasoning mode: users can keep the model in standard mode for fast answers or turn on extended thinking for tasks that require visible, step-by-step deliberation.

The extended thinking mode is not a separate endpoint. Developers can toggle it on or off per request and, through the Anthropic API, set a thinking budget measured in tokens. Anthropic's extended thinking documentation says users can direct the model to think more deeply about trickier questions and developers can set a thinking budget to control precisely how long Claude spends on a problem.

Anthropic describes the thinking budget as a mechanism for trading speed and cost against answer quality. The budget is expressed in tokens, which makes the trade-off measurable and gives developers a concrete knob to tune in production. This is a departure from models where reasoning depth is fixed at inference time and cannot be adjusted per request.

Claude 3.7 Sonnet at a glance

Attribute	Detail
Release date	Feb. 24, 2025
Knowledge cutoff	October 2024
Reasoning mode	Hybrid: standard and extended thinking
SWE-bench Verified	70.3% (high-compute scaffolded, 489-task subset)
Input price	$3 per million tokens
Output price	$15 per million tokens, including thinking tokens
Availability	Claude plans, Anthropic API, Amazon Bedrock, Google Cloud Vertex AI
Agentic tool	Claude Code research preview CLI

How does extended thinking work?

In extended thinking mode, Claude 3.7 Sonnet produces a chain of reasoning before generating its final answer, and that reasoning is visible to the user. Anthropic says this visibility helps developers audit how the model reached a conclusion. The thinking budget creates a direct trade-off: a larger budget can improve answer quality on complex tasks, while a smaller budget keeps latency and cost down.

Anthropic positions the budget control as a practical lever for production workloads. A customer running a customer-support classifier might set a minimal budget because speed matters more than deliberation. A developer debugging an obscure compiler error might allocate a much larger budget and wait longer for a detailed patch. The same model serves both use cases.

1. Toggle extended thinking on or off for any individual request.
2. Set a token-based thinking budget through the Anthropic API to cap how long the model deliberates.
3. Balance cost and latency by choosing a lower budget for routine queries and a higher budget for complex coding or multi-step reasoning.

The mechanism matters because it treats reasoning as a parameter rather than a product line. Keeping reasoning inside one model lets a conversation move from a quick answer to deep analysis without switching endpoints. It also avoids the context fragmentation that can occur when a fast model and a reasoning model are used in the same workflow.

What is Claude Code?

Claude Code is a limited research preview command-line tool that Anthropic introduced alongside the model. It is designed for agentic coding: the tool can navigate a codebase, edit files, write and run tests, create git commits and pushes, and use command-line tools to complete tasks. Anthropic describes it as a command line tool for agentic coding.

1. Navigating and reading repository files autonomously.
2. Editing files across a codebase.
3. Writing and running tests.
4. Creating git commits and pushing changes.
5. Using command-line tools to complete coding tasks.

The tool is aimed at developers who want an AI agent that can carry a task from specification to committed code. Because Claude Code operates on the command line, it can integrate with existing developer workflows, version control systems, and CI tooling. Anthropic is releasing it as a research preview, signaling that the workflow is still evolving and that the company expects feedback from early users.

Claude Code extends Anthropic's agentic strategy beyond chat interfaces. Instead of asking a model to suggest a patch, a developer can delegate the entire loop of exploration, editing, testing, and committing. The tool's ability to push changes is a step beyond models that only generate code.

What does the 70.3% SWE-bench Verified score mean?

Anthropic reports that Claude 3.7 Sonnet achieves 70.3% on SWE-bench Verified, a state-of-the-art result on the 489-task subset. The benchmark evaluates models on software engineering tasks, and Anthropic reported the result with a high-compute scaffolded setup.

SWE-bench Verified measures whether a model can take a GitHub issue and produce a patch that passes the associated tests. The 70.3% figure places Claude 3.7 Sonnet at the top of Anthropic's reported results and strengthens the case for using the same model as the engine behind Claude Code. For developers, the score is a signal that the model can handle repository-scale tasks, not just isolated code generation.

The benchmark result is also a point of differentiation in the enterprise market. Coding agents have become one of the most commercially visible uses of frontier models, and a strong SWE-bench Verified score is a marketing asset for procurement decisions. Anthropic's decision to bundle that model with a CLI agent gives it a combined software-development offering.

How is Claude 3.7 Sonnet priced and where is it available?

Pricing for Claude 3.7 Sonnet is unchanged from the previous Sonnet generation. Anthropic charges $3 per million input tokens and $15 per million output tokens, and thinking tokens are included in the output-token price. The decision to hold pricing steady while adding extended thinking gives existing customers access to deeper reasoning without renegotiating cost structures.

Availability spans Anthropic's consumer and enterprise surfaces. The model is on all Claude plans, with extended thinking excluded from the free tier, and is available through the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI. AWS confirmed the Bedrock rollout and highlighted the ability to toggle between standard and extended thinking modes within the model.

The broad distribution means Claude 3.7 Sonnet is not locked to a single cloud. Enterprises running on AWS can call the model through Bedrock, while teams on Google Cloud can access it through Vertex AI. Developers already using the Anthropic API can adopt the new model without changing their commercial relationship with Anthropic.

What do experts say about the shift?

Dianne Penn, Anthropic's product research lead, told The Verge that the company sees reasoning as a feature of the AI rather than a separate offering. “We fundamentally believe that reasoning is a feature of the AI rather than a completely separate thing,” Penn said. The quote captures the strategic bet behind Claude 3.7 Sonnet: reasoning should be embedded in the model's core behavior, not bolted on as a separate product.

We fundamentally believe that reasoning is a feature of the AI rather than a completely separate thing.Dianne Penn, Anthropic product research lead

AWS's announcement echoes that framing, noting that Claude 3.7 Sonnet introduces a unique approach to AI reasoning by integrating it seamlessly with other capabilities. For enterprises, that integration can simplify model selection and reduce the need to operate parallel reasoning and non-reasoning pipelines. It also changes how vendors are likely to describe reasoning capabilities in future model releases.

What comes next for hybrid reasoning?

Claude 3.7 Sonnet is the first hybrid reasoning model on the market, according to Anthropic, and the release may reset expectations for how frontier models are packaged. If the hybrid approach gains traction, specialized reasoning-only models could become less central to production AI stacks. The model's visible reasoning and controllable budget also point toward a future where transparency and cost control are built into the inference layer rather than managed externally.

Claude Code, meanwhile, is an early test of autonomous agentic workflows. Anthropic is positioning the command-line agent as a way to move from a single model response to a completed software task. The research-preview label suggests the company expects the tool to change as developers stress it against real repositories. The combination of a hybrid model and an autonomous coding agent gives Anthropic a distinctive position in the frontier-model race.

Observers will be watching how the thinking budget performs in production and whether rival labs follow with their own hybrid designs. Anthropic's move also raises a broader question: if reasoning is a feature rather than a separate model, what other capabilities might be integrated into future frontier releases? The answer will determine whether the reasoning-model market of 2024 becomes a feature set by 2026.