Claude Sonnet 4: 1M Context Beta, Access & Pricing

A 5x increase, gated behind a beta header

Anthropic announced on August 12, 2025 that Claude Sonnet 4 supports a 1 million token context window on the Anthropic API, up from 200,000 tokens, a fivefold increase. The announcement described the use case as loading entire codebases, including source files, tests and documentation, or processing large sets of documents such as legal contracts and research papers in a single request, and pointed to agents that maintain context across hundreds of tool calls. [3]

The feature launched in public beta and required sending a beta header, context-1m-2025-08-07, with each API request. Anthropic restricted access at launch to customers on API usage tier 4 or with a custom rate limit, a tier that in Anthropic's published system requires having purchased at least 400 dollars in API credits, with wider availability to follow. [2][3]

Atlas interpretation: The gating mattered as much as the headline number. A context window announced as a capability of Claude Sonnet 4 was, on day one, available only behind an opt-in beta flag and only to a spending-qualified slice of API customers, not to Claude.ai subscribers or to every developer with an API key. [2]

A new way to price a prompt

The expanded window came with length-based pricing on the Anthropic API: prompts of 200,000 tokens or fewer stayed at 3 dollars per million input tokens and 15 dollars per million output tokens, while prompts above that threshold cost 6 dollars per million input tokens and 22.50 dollars per million output tokens, roughly double the input price and one and a half times the output price. [3][5]

Atlas interpretation: Simon Willison flagged this as Anthropic's first tiered-by-length pricing on a Claude model, noting Google had already been pricing Gemini's long context this way for some time. Treating a long prompt as a more expensive product, rather than folding the cost into a flat per-token rate, is the kind of change that shows up in a bill before it shows up in a benchmark. [2]

The cloud platforms followed within two weeks

Amazon Web Services published its own announcement the same day, putting the expanded Claude Sonnet 4 context window into public preview on Amazon Bedrock in three US regions, Oregon, N. Virginia and Ohio, with the same roughly doubled input and 1.5 times output pricing above 200,000 tokens. Anthropic's own announcement said the window would also reach Google Cloud's Vertex AI on August 26, 2025. [5][3]

Atlas interpretation: Shipping the same capability across the API and both major third-party cloud platforms inside two weeks reflects how much of Anthropic's Claude Sonnet 4 usage runs through Bedrock and Vertex AI rather than the direct API. A feature that stayed API-only would have missed most enterprise traffic. [5]

Catching up more than leading

Both Willison and independent testing published the same week noted that a 1 million token window was not new to the field: Google's Gemini models had offered context windows at or above that size since 2024, and other providers had similar figures. A comparison published by Every the day of the announcement ran the same long-document and needle-in-a-haystack tasks against Claude Sonnet 4, Gemini and GPT-4.1: Claude answered faster than Gemini Flash on a text-analysis task, 41.8 seconds versus 69.2 seconds, and produced shorter, more concise responses, roughly 500 words against Gemini Flash's 3,372, while Gemini scored higher on a code-analysis task, 91.7 versus 74.6. All three models it tested located facts buried partway through 900,000-word documents. [2][4]

Atlas interpretation: The independent comparison undercuts the framing of this as a capability jump. Claude Sonnet 4 arrived at parity with context windows competitors already shipped, at a list price, 6 dollars per million input tokens above 200,000, that the same test noted was several times Gemini's comparable rate. What Anthropic added was the size number closing a gap, not a new capability nobody else had, and the retrieval task the comparison ran suggests the window's practical value depends on what a workload actually needs pulled back out of it, not just on how many tokens it can hold. [4]