Claude 3.7 Sonnet
Anthropic's hybrid reasoning model with a controllable thinking budget
Claude 3.7 Sonnet is a large language model from Anthropic that combines quick replies with optional step-by-step reasoning in a single system. Released in February 2025, it targets coding and real-world business tasks, and it powers the Claude Code command line tool.
- First hybrid reasoning model, per Anthropic, merging instant and deliberative modes
- Thinking budget settable by API callers up to a 128K token output ceiling
- Priced at $3 per million input tokens and $15 per million output tokens
- Anthropic reports a 45% drop in unnecessary refusals versus Claude 3.5 Sonnet
- Launched with Claude Code, a terminal tool for engineering tasks
What it is
Claude 3.7 Sonnet is a general-purpose language model built by Anthropic. The company describes it as its smartest release so far and as the first system on the market to merge two response styles: a fast mode that answers straight away, and a slower mode that works through a problem step by step before replying. That longer deliberation is shown to the user. Anthropic points to coding and front-end web work as areas of particular strength. In its standard setting the model behaves like an upgraded Claude 3.5 Sonnet; switching on the longer mode lifts results in maths, physics, instruction following and programming.
How it came about
Anthropic announced the model alongside Claude Code, a terminal-based helper for engineering work that was offered as a limited research preview. The company says Sonnet has been the model developers favour most since June 2024, and that this release is meant to push that further. Claude Code grew out of Anthropic's own engineering practice and is intended to show how developers use the model for programming, so that future versions can be improved. Planned upgrades for the tool included more reliable tool calls, support for long-running commands and better in-app rendering.
Inside the model
Rather than shipping a separate reasoning model, Anthropic folded reflection into one system. Users choose per request whether the model replies immediately or thinks first. Through the API, callers can also cap the thinking budget at any number of tokens up to the 128K output limit, trading speed and cost against answer quality. Prompting works much the same in either mode. Anthropic says it deliberately spent less effort on maths and computer science contest puzzles and more on tasks resembling how businesses actually use language models. Benchmark runs used extra scaffolding, including a planning tool and a raised step cap.
What it can do
Anthropic reports leading results on SWE-bench Verified, which measures fixes to real software problems, and on TAU-bench, which tests agents on multi-step tasks involving users and tools. The company also claims strong instruction following, general reasoning, multimodal handling and agentic coding, with the longer mode adding most in maths and science, and says it beat earlier models in its own Pokémon play tests. External firms including Cursor, Cognition, Vercel, Replit and Canva reported favourably on its coding and agent work. Anthropic notes the TAU-bench figure for Claude 3.5 Sonnet was re-run after dataset changes.
How to use it
The model is served through every Claude plan, from Free to Enterprise, though the longer thinking mode is withheld from the free tier. It is also reachable through Anthropic's developer platform, Amazon Bedrock and Google Cloud's Vertex AI. Pricing matches the earlier Sonnet releases at $3 per million input tokens and $15 per million output tokens, with thinking tokens counted as output. Claude Code was released to a limited research preview.
Safety and testing
Anthropic says it tested the model with outside specialists on security, safety and reliability. Its system card reports Responsible Scaling Policy evaluations across several risk categories, looks at prompt injection attacks tied to computer use and how the model is trained to resist them, and considers whether reasoning traces can be trusted as an account of how decisions are made. The company claims the model tells harmful from harmless requests more finely than its predecessor, cutting unnecessary refusals by 45 per cent.
Reception and debate
Firms given early access described the model in positive terms. Cursor called it best-in-class for practical programming, Cognition said it planned code changes and full-stack updates better than rivals, Vercel praised its accuracy on complex agent workflows, Replit used it to build web apps and dashboards where other systems stalled, and Canva reported production-ready output with fewer errors and better design judgement. No lawsuits, bans or public disputes are recorded here.
Why it matters
Anthropic presents the release as a move towards systems that reason deeply, act on their own and work alongside people. The accompanying terminal tool is described by the company as already essential inside its own engineering team for test-driven work, debugging and large refactors, finishing in one pass jobs that would take more than 45 minutes by hand. Feedback from the preview is meant to shape later models.
Questions people ask about Claude 3.7 Sonnet
What makes Claude 3.7 Sonnet different from earlier Claude models?
It folds two behaviours into one system. You can ask for a straight answer or let the model reason step by step first, and through the API you can cap how many tokens it spends thinking. Anthropic says this beats shipping a separate reasoning model.
How much does it cost?
Anthropic charges $3 per million input tokens and $15 per million output tokens, the same rates as the previous Sonnet releases. Tokens used during the longer thinking mode count towards the output charge. The model is on all Claude plans, but the free tier cannot use the longer mode.
Where can I run it?
It is available on the Claude apps and website across Free, Pro, Team and Enterprise plans. Developers can also reach it through Anthropic's developer platform, Amazon Bedrock and Google Cloud's Vertex AI.
What is Claude Code?
A command line helper released alongside the model as a limited research preview. It works with your repositories and terminal, handling engineering tasks such as editing, testing and version control while keeping the developer involved at each step.
How safe is it?
Anthropic says outside experts helped test it, and its system card covers Responsible Scaling Policy evaluations, prompt injection risks from computer use, and whether reasoning traces are trustworthy. The company claims fewer needless refusals than the previous version.
Good reading and sources
There is no English Wikipedia article on Claude 3.7 Sonnet yet, so the facts on this page come from the sources listed above: anthropic.com/news/claude-3-7-sonnet. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.