Huge context window change

# Huge context window change Anthropic is shutting off an older way developers accessed a one‑million‑token context window, and the cutoff is specific: April 30, 2026. After that date, requests that still rely on the old beta header for Claude Sonnet 4.5 or Claude Sonnet 4 will fail instead of quietly falling back. (platform.claude.com) To understand why this matters, start with the idea of a context window. A context window is the total amount of text a model can keep “on the table” at once while it answers, including your prompt, any attached material, earlier conversation turns, and the model’s own reply. (platform.claude.com) A token is the small chunk of text that systems count internally instead of full words. Anthropic’s documentation describes the context window as capacity that can go up to 1 million tokens, which is why developers use token counts as the real limit when they send large prompts, long chats, or giant documents. (platform.claude.com) That token budget gets used faster than many teams expect. Anthropic says the context window includes both stored conversation history and newly generated output, so a long agent session can fill the window even if no single user message looks especially large. (platform.claude.com) This is where the April change bites. Anthropic’s release notes say the company is retiring the 1M token context window beta for Claude Sonnet 4.5 and Claude Sonnet 4 on April 30, 2026, which means any workflow still depending on that beta access path needs to be updated before the deadline. (platform.claude.com) The practical result is not that one million tokens disappear from Anthropic’s platform entirely. The result is that Sonnet 4.5 and Sonnet 4 go back to being, in normal use, 200,000‑token models unless a team migrates to newer models that support larger windows by default. (platform.claude.com, platform.claude.com) Anthropic’s current documentation shows a split between generations. Claude Opus 4.6 and Claude Sonnet 4.6 have a 1M‑token context window, while other Claude models including Claude Sonnet 4.5 and Claude Sonnet 4 are listed with a 200k‑token context window. (platform.claude.com) That makes this less like a tiny settings tweak and more like a compatibility break. If a product was built around stuffing an entire codebase, contract archive, or research bundle into one request on Sonnet 4 or Sonnet 4.5, the old request format will not survive past April 30. (platform.claude.com, anthropic.com) The most exposed teams are the ones running massive document ingestion or long agent loops. Anthropic’s context-window guide specifically points developers toward strategies like server-side compaction and context editing for long-running conversations, which is a strong hint that “just keep appending everything forever” is not the intended pattern. (platform.claude.com) In plain terms, developers now have three choices. They can split large inputs into smaller chunks, summarize earlier material before continuing, or move to a newer model such as Claude Sonnet 4.6 or Claude Opus 4.6 that supports a one‑million‑token window without the old beta mechanism. (platform.claude.com, platform.claude.com) There is another wrinkle here: Anthropic has also been changing how long outputs work. In release notes dated March 30, 2026, the company said developers can use the output-300k-2026-03-24 beta header to generate longer single-turn outputs, up to a 300,000 max_tokens cap on the Message Batches Application Programming Interface, which is a separate knob from the size of the input context window. (platform.claude.com) That distinction matters because some teams confuse “how much the model can read” with “how much the model can write back.” A workflow can be broken by the loss of a one‑million‑token input path even if output limits improved elsewhere. (platform.claude.com, platform.claude.com) The immediate deadline is simple and concrete: before April 30, 2026, audit any Anthropic integration that still uses the old 1M beta header with Claude Sonnet 4.5 or Claude Sonnet 4. If those requests are not redesigned or migrated, Anthropic says they will error out after the retirement date. (platform.claude.com) For companies that built around giant prompts, this is the kind of infrastructure change that shows up first as a production incident, not a strategy memo. The teams that avoid trouble will be the ones that treat 200,000 tokens as the safe default for older Sonnet flows and rework oversized jobs now, while there is still time. (platform.claude.com, platform.claude.com)