Claude 4 Sonnet Launches with Extended Thinking and 200K Context
Anthropic has released Claude 4 Sonnet with general availability via the Claude API and claude.ai, adding extended thinking mode, a 200K token context window, and improved tool use reliability for agentic workflows.
Original sourceAnthropic has made Claude 4 Sonnet generally available to developers and users through both the Claude API and claude.ai. The model introduces an extended thinking mode, which allows Claude to work through complex problems with explicit intermediate reasoning before producing a final response — a capability previously limited to the Opus tier. The 200K token context window matches the upper range of what competitors like Gemini have offered at scale, enabling processing of large codebases, legal documents, or lengthy research corpora in a single pass.
The release also emphasizes improved tool use reliability, targeting multi-step agentic workflows where previous models occasionally dropped tool calls or mishandled sequencing across long chains. Anthropic positions this as a meaningful improvement for developers building autonomous agents that need consistent, predictable behavior across dozens of steps — not just in demos, but in production deployments.
Claude 4 Sonnet slots into Anthropic's model lineup as the middle tier between Haiku and Opus, but with extended thinking now available at this tier, the capability gap between Sonnet and Opus narrows considerably for reasoning-heavy tasks. Pricing and rate limit details are available through the Anthropic API console, with the model accessible immediately to existing API customers without a waitlist.
The extended thinking feature is notable because it externalizes the chain-of-thought process, giving developers visibility into how the model reasons — which can be valuable for debugging agent behavior, building trust in high-stakes outputs, or simply understanding why a model reached a particular conclusion. Whether the thinking tokens are billed separately or bundled into the output token count is a detail developers will want to verify before scaling.
Panel Takes
The Builder
Developer Perspective
“The primitive here is straightforward: a reasoning model with an inspectable thinking trace, 200K context, and more reliable tool-chaining — all at the Sonnet price point. The DX bet is that exposing thinking tokens gives developers a debugging surface for agent failures, which is the right call because silent failures in multi-step workflows are currently a nightmare to trace. I need to verify whether thinking tokens are billed separately before I'd commit to a pricing model for anything at scale, but if the tool use reliability claims hold up in production, this earns a real look for agent infrastructure.”
The Skeptic
Reality Check
“Extended thinking has been available in OpenAI's o-series and Google's Gemini thinking models for a while now, so Anthropic is catching up at the Sonnet tier, not breaking new ground — calling it out matters because the framing implies otherwise. The specific claim that tool use reliability has meaningfully improved is exactly the kind of thing that needs third-party evals, not an announcement post, because every model release says this and roughly half of them are wrong about it in production. What would flip me from skeptical to shipping is published benchmark methodology on multi-step tool call accuracy at 50-plus steps, from someone other than Anthropic.”
The Futurist
Big Picture
“The thesis embedded in this release is that inspectable reasoning at a mid-tier price point becomes the baseline expectation within 18 months — not a premium feature. The second-order effect worth watching is that visible thinking traces shift power toward developers who can audit and steer model reasoning, which changes how trust gets established in regulated industries like legal and finance where 'the model said so' was always a dead end. Anthropic is riding the trend of reasoning transparency moving down-market, and they're roughly on time — not early, but not late enough to cede the developer trust narrative to OpenAI.”
The Founder
Business & Market
“The strategic move here is compressing the capability gap between Sonnet and Opus, which protects revenue at the mid tier against developers who might have been upgrading to Opus purely for reasoning depth — now they might not need to. The moat question is the same as always with Anthropic: the Constitutional AI brand and safety positioning buy credibility with enterprise buyers, but the underlying model improvements are replicable, and Google and OpenAI both have comparable reasoning tiers with deeper distribution. The business survives if enterprise compliance requirements create a defensible buyer who values Anthropic's safety narrative over raw capability comparisons, and that buyer exists — the question is whether they're large enough.”