Claude 4.5 Sonnet Arrives with Extended Thinking and 1M Token Context
Anthropic has released Claude 4.5 Sonnet, a new model featuring an optional extended thinking mode, a 1-million-token context window, and benchmarked improvements on coding and agentic workflows. It is available immediately via API and Claude.ai.
Original sourceAnthropic released Claude 4.5 Sonnet today, positioning it as a significant step up from its predecessor across the coding and agentic task benchmarks the company publishes. The headline capabilities are two: an extended thinking mode that lets the model spend more compute on multi-step reasoning before responding, and a 1-million-token context window that roughly doubles what Claude 3.5 Sonnet offered. Both features are available at launch rather than being staged rollouts.
Extended thinking is opt-in at the API level, meaning developers choose when to invoke the additional reasoning overhead rather than absorbing it on every call. Anthropic describes it as particularly relevant for tasks requiring decomposition — code generation across large codebases, long-horizon planning, and agentic pipelines where intermediate reasoning quality compounds across steps. The 1M context window opens the door to ingesting entire repositories, lengthy legal or research documents, or multi-session conversation histories without chunking.
The model is available immediately through the Anthropic API and on Claude.ai, with pricing details published on Anthropic's pricing page. Anthropic notes that extended thinking mode carries a separate token cost tied to the reasoning tokens generated, a model that will be familiar to developers who have worked with OpenAI's o-series or Google's Gemini thinking variants. Competition in this segment is now effectively a four-way race between Anthropic, OpenAI, Google, and an increasingly capable open-source tier.
For teams already running Claude 3.5 Sonnet in production, the upgrade path is a model string swap, though teams relying on predictable latency should benchmark extended thinking mode separately — the additional reasoning pass introduces variable response times that may require timeout and retry logic adjustments in agentic systems.
Panel Takes
The Builder
Developer Perspective
“The primitive here is clean: a thinking budget you toggle at call time, not a separate endpoint or a new SDK to install. The DX bet is that putting complexity at the API parameter level — rather than in a new model tier or a wrapper service — is the right call, and it is. The moment of truth is whether the extended thinking tokens justify their cost on your specific workload, and Anthropic at least gives you the knobs to measure that directly rather than hiding it behind a flat pricing abstraction.”
The Skeptic
Reality Check
“1M context and extended thinking are real features, not marketing — both Google Gemini and OpenAI have shipped comparable capabilities, so this is Anthropic closing a gap, not opening one. The scenario where this breaks is long agentic pipelines at scale: variable reasoning latency plus token costs that compound across steps will blow up cost estimates for teams that didn't build with thinking-mode overhead in mind. What kills this in 12 months is not a competitor but Anthropic's own pricing — if the reasoning token costs don't come down as the underlying compute gets cheaper, developers will fine-tune smaller models for the same coding tasks and route around it entirely.”
The Futurist
Big Picture
“The thesis embedded in 1M context plus extended thinking together is that the bottleneck in agentic systems shifts from 'can the model hold enough state' to 'can the model reason well enough over that state' — and that both constraints need to fall at the same time to unlock the next class of applications. The dependency that has to hold is that long-context retrieval quality doesn't degrade at the tail of a 1M window, which is not a given and Anthropic has not published needle-in-a-haystack data for this release. If both capabilities hold up under real workloads, the second-order effect is that the 'retrieval vs. context' architectural debate in RAG systems gets reframed — teams will stop chunking and start asking whether the retrieval layer is load-bearing at all.”
The Founder
Business & Market
“The buyer here is the enterprise dev team already on Anthropic's API, and the upsell mechanism is elegant: extended thinking mode generates more tokens per call, which means higher spend per task without requiring a seat-count expansion conversation. The moat question is real though — OpenAI and Google are shipping functionally equivalent reasoning toggles, and Anthropic's defensibility rests almost entirely on model quality perception and the trust narrative around safety, neither of which shows up cleanly on a pricing comparison page. If reasoning token costs don't compress faster than the competition's, Anthropic risks being the premium option in a market that is rapidly treating long-context reasoning as a commodity input.”