Claude 3.7 Sonnet and Claude Code
Anthropic's first hybrid reasoning model makes thinking a runtime parameter you tune per call, not a separate product tier — and ships a command-line coding agent in the same announcement.
Most reasoning models are a separate product tier. You pick the smarter SKU, pay the premium, accept the latency. Anthropic built reasoning into a single model and made the thinking budget a runtime parameter — same price, toggleable at call time.
Source
Claude 3.7 Sonnet and Claude Code
By: Anthropic
Published: Feb 24 2025
Type: Engineering article
Saved because this is the first time extended thinking shipped as a toggleable mode on a general-purpose model, and because Claude Code signals Anthropic moving directly into the terminal-based developer workflow.
One model, two modes — and what that changes for agentic coding
The standard split in AI product design puts reasoning behind a separate tier. You pick the smarter SKU, pay the premium, accept the latency. Claude 3.7 Sonnet breaks that framing. It is both a fast completions model and a reasoning model within a single package. The switch is a runtime parameter, not a product decision.
In extended thinking mode, the model reflects before answering. The API exposes a thinking budget — a token ceiling settable to any value up to 128K — and the model spends up to that amount reasoning before it responds. Set the budget to zero and it behaves like the standard Claude 3.5 Sonnet upgrade it also represents. Pricing stays flat either way: $3 per million input tokens, $15 per million output tokens, with thinking tokens counting against the output price. A deeply reasoning response costs more than a shallow one, but the base rate does not change, which keeps the mode accessible to test.
For agentic systems, this matters more than benchmark numbers alone suggest. Multi-step coding workflows rarely need uniform reasoning across every step. Step three might require tracing complex dependency logic; step seven might need a quick file lookup. A variable thinking budget per API call lets you express that in application code — decided per task rather than locked to a model tier.
On SWE-bench Verified, Anthropic used deliberately minimal scaffolding: a bash tool, a file editing tool, and a planning tool. No best-of-N rejection sampling, no Monte Carlo Tree Search. Under that setup, Claude 3.7 Sonnet reached 62% pass@1 on the verified subset. With parallel sampling and a scoring model to select the best attempt, that rose to 70.3%. Both numbers placed the model at the top of the leaderboard at launch. On TAU-bench, which evaluates agents on complex real-world tasks involving user and tool interactions rather than synthetic problems, it again led. The deliberate choice to use minimal scaffolding means the results are more portable to realistic deployment conditions than benchmarks that rely heavily on orchestration tricks.
External partners who tested the model before release confirmed the gains held on production codebases. Cursor described it as best-in-class for handling complex repositories and advanced tool use. Cognition found it significantly better at planning code changes and managing full-stack updates. Canva reported fewer errors and better design judgment in generated code. These are vendor testimonials, not independent evaluations, but the consistency across different codebase types and use cases makes the pattern worth noting.
Alongside the model, Anthropic launched Claude Code — a command-line tool in limited research preview. It can read and search a codebase, edit files, write and run tests, commit and push to GitHub, and call arbitrary command-line tools. Each step keeps the developer in the loop rather than running unsupervised. The design is intentionally collaborative: Anthropic was transparent at launch that tool call reliability and support for long-running commands were still being improved, and described the preview as a mechanism for collecting real developer usage data to feed back into model training.
Internally, Anthropic’s team had already been using Claude Code for test-driven development, debugging, and large-scale refactoring. The stated result was some tasks completing in a single pass that would otherwise take 45 or more minutes. That claim cannot be independently verified, but the categories listed — TDD, debugging, refactoring — map directly to scenarios where AI coding assistance has shown the most consistent early returns, and to Claude Code’s current feature set.
The GitHub integration expanding to all Claude plans is separate from Claude Code. It connects repositories to the Claude.ai chat interface for context-aware question-answering and documentation without requiring the terminal workflow.
Two safety areas in the system card are directly relevant to agentic use. The model makes better distinctions between harmful and benign requests, with unnecessary refusals down 45% compared to Claude 3.5 Sonnet — which matters when building tools that need Claude to act without excessive hedging. The card also documents evaluation of prompt injection in computer use scenarios, a threat surface that grows more important as agents interact with external, untrusted content. On reasoning transparency, the card is careful: visible thinking steps do not guarantee the reasoning is truthful or representative of what the model computed internally.
My take
The interesting signal here is not which benchmark it topped. It is that thinking budget is an explicit API parameter — the quality/latency/cost tradeoff belongs in application code, decided by the person who knows the task. Claude Code shipping to external developers in the same announcement suggests Anthropic intends to use real developer workloads as the primary test environment for agentic capability going forward.
Where this is useful
- Deciding whether to enable extended thinking on an existing Claude integration before switching to a higher model tier
- Designing multi-step agentic pipelines where individual steps have different reasoning requirements
- When NOT applicable: latency-sensitive, cost-optimized tasks where consistent fast response times matter more than deeper reasoning
Apply this: Before upgrading to a more expensive model tier, set a thinking budget on your current model for the tasks where output quality matters most and measure the difference first.
What I want to remember
Remember: Reasoning is a dial you tune per call — not a different model you switch to.