Claude 3.7 Sonnet Gets 128K Extended Thinking and Better Tool Use
Anthropic has updated Claude 3.7 Sonnet to support 128K token budgets for extended thinking and improved tool use reliability, available immediately via API and Claude.ai. The update makes sustained, multi-step reasoning tasks significantly more capable without requiring a model version change.
Original sourceAnthropic quietly shipped a meaningful capability update to Claude 3.7 Sonnet, raising the extended thinking token budget ceiling from its previous limit to 128K tokens. Extended thinking is Anthropic's mechanism for giving the model dedicated scratchpad space to reason through complex problems before producing a final answer — the higher the budget, the more intermediate reasoning steps the model can take before committing to a response.
Alongside the thinking budget increase, Anthropic addressed tool use reliability, a persistent pain point in agentic workflows where the model calls external functions or APIs. The update improves consistency in how the model handles tool call formatting, retries, and multi-turn tool sequences — the class of bugs that tend to surface when extended thinking and tool use are combined in the same request.
The update is a drop-in change for existing claude-3-7-sonnet users. No model version pin change, no migration guide, no new API surface. Developers who have already built on 3.7 Sonnet get the capability increase automatically. For teams running long-context reasoning pipelines or multi-step agent loops, this is a material improvement delivered with minimal friction.
The 128K thinking budget puts Claude 3.7 Sonnet in competitive range for tasks like multi-document legal analysis, long-horizon code refactoring, and complex multi-step research workflows — use cases where earlier token budget constraints were a real bottleneck. Whether the reliability improvements in tool use hold up under production-scale agentic workloads remains to be stress-tested by the developer community.
Panel Takes
The Builder
Developer Perspective
“The primitive here is clean: more scratchpad tokens plus better tool call consistency, delivered as a silent upgrade to an existing model endpoint. The DX bet Anthropic made — no version pin change, no migration, no new surface area — is exactly the right call. Where this earns the ship is the tool use reliability fix, because the failure mode it addresses (malformed tool calls mid-chain when thinking is enabled) is the specific thing that was forcing developers to write defensive retry wrappers around every agentic step.”
The Skeptic
Reality Check
“128K thinking tokens is a real number, but Anthropic hasn't published benchmarks showing where the quality curve actually inflects — at what budget does reasoning quality plateau, and how much of this is token spend with diminishing returns? The tool use reliability claim is the more interesting one to pressure-test: 'improved' is doing a lot of work in that sentence, and production agentic pipelines will find the edge cases within weeks. What kills this in 12 months isn't a competitor — it's that OpenAI and Google ship equivalent reasoning budgets as table stakes, and the differentiation evaporates.”
The Futurist
Big Picture
“The thesis here is falsifiable: long-context reasoning is the bottleneck for autonomous agents completing real knowledge-work tasks, and removing that ceiling unlocks a qualitative step change in what agents can close-loop on without human intervention. The dependency that has to hold is that token budget — not architecture, training data, or task decomposition strategy — is actually the binding constraint in current agentic workflows. The second-order effect if this thesis is right: the unit of agent work shifts from 'task' to 'project,' and the tooling and pricing infrastructure built around short-horizon interactions becomes obsolete faster than the market expects.”
The PM
Product Strategy
“The job-to-be-done this update directly serves is 'complete a complex reasoning task without hitting an artificial ceiling mid-thought' — and that's a real job that was causing real workflow breakage for enterprise users running document analysis and code review pipelines. The product decision that earns the ship is the zero-migration delivery: existing users get better without doing anything, which is the rarest outcome in API-land. The gap that remains is observability — users still have no native way to inspect whether their thinking budget was the binding constraint or whether they're spending tokens they don't need, which means optimization is still manual and opaque.”