Anthropic ships Claude Sonnet 4.5 with extended thinking over 200K context for coding agents

What Anthropic shipped

Anthropic has released Claude Sonnet 4.5, an update aimed squarely at developers building coding agents. According to a dev.to report on the launch, the headline feature is not a raw benchmark score but the pairing of a 200,000-token context window with a new "extended thinking" mode built for agentic workflows.

The 200K window itself is not new — earlier Sonnet models already offered it, as the dev.to post notes. What has changed is how effectively the model can use that capacity while working through multi-step tasks that span an entire codebase.

Interleaved reasoning, tunable by budget

Extended thinking lets Sonnet 4.5 mix reasoning with action. Instead of producing one plan and executing it blindly, the model can pause mid-task, examine intermediate results and revise its approach before continuing.

In the Claude API this is exposed through a thinking parameter with a budget_tokens setting that caps how much reasoning the model performs per request. That turns reasoning depth into an explicit dial: a lightweight agent that only reads files can run with a small budget, while a complex multi-file refactor can be given a much larger one.

The dev.to write-up, citing an Anthropic technical report, says the interleaved approach cuts hallucination rates on multi-step agent benchmarks by roughly 30% compared with Sonnet 4. In practice, the model can hold a full repository in context, plan a refactor across dozens of files, and check its own output against the original source instead of inventing plausible-looking diffs.

Where it lands on benchmarks

On SWE-bench Verified, a benchmark assembled from real GitHub issues, the dev.to report places Sonnet 4.5 at about 77.2%, up from roughly 68% for Sonnet 4.

Against GPT-4.1 and Gemini 2.5 Pro on the same benchmark, the model reportedly leads on multi-file refactoring tasks while trailing slightly on single-file bug fixes — a trade-off that favors whole-codebase work, which is precisely the kind of job autonomous agents are increasingly handed.

Implications for agent builders

For teams working with frameworks such as LangGraph, CrewAI or AutoGen, the practical effect is that more complexity can stay inside a single agent loop instead of being fragmented across orchestrated sub-tasks. The dev.to post points to Anthropic's Claude Cookbooks repository, which is said to already include examples of Sonnet 4.5 holding a coherent plan across more than 50 tool calls in one conversation.

The cost model is the main caveat. Reasoning tokens are billed like any other output, so extended thinking raises the price of a request. The report suggests tuning the budget per workload: reserve deep reasoning for tasks that genuinely need planning and skip it for simple fetch-and-respond loops.

Why it matters

Autonomous coding is where the major labs are currently competing hardest, and the failure modes that break coding agents are well known: losing track of dependencies across files, hallucinating changes that look right, and over-spending on reasoning for trivial steps. If the reported numbers hold up, Sonnet 4.5 addresses all three at once — long context for the first, mid-task reflection for the second, and an explicit reasoning budget for the third.

The release also marks a shift in how reasoning models are positioned: rather than a fixed property of the model, depth of thought becomes something a developer dials per task and pays for accordingly. Teams already running agents in production will need to decide where that dial sits for each workload, since the same capability that enables a verified multi-file refactor can quietly inflate bills on routine calls.

One caution: every figure here arrives through a single community report on dev.to rather than Anthropic's own release materials, so the benchmark numbers and the claimed 30% hallucination reduction are best read as reported claims pending official confirmation.