Claude gets a million-token memory
Anthropic rolled out a million-token context window across Claude models — effectively letting agents hold massive histories in one session. That reduces context-chopping workarounds and changes how teams design episodic memory and long-horizon orchestrations.
[Anthropic announced]aihola.com on March 13, 2026 that Claude Opus 4.6 and Sonnet 4.6 support a 1,000,000‑token context window, details that are reflected on Anthropic’s model pages for Opus and Sonnet. anthropic.com The [company dropped]byteiota.com the long‑context surcharge that had applied above 200K tokens, while keeping Opus base pricing at $5 input/$25 output per million tokens and Sonnet at $3 input/$15 output per million tokens on its pricing pages. anthropic.com Anthropic’s documentation lists prompt‑caching savings of up to 90% and batch‑API discounts near 50%, mechanisms the platform recommends to reduce the effective cost of repeated large‑context requests. anthropic.com Opus 4.6 adds operational controls—“Adaptive Thinking,” a Compaction API, and four granular “effort” levels—that Anthropic describes as tools to mitigate context rot and trade off intelligence for speed and billing. anthropic.com Anthropic’s rate‑limits page records quotas measured as requests per minute (RPM), input tokens per minute (ITPM) and output tokens per minute (OTPM), and notes 1M‑token contexts are gated to higher usage tiers—metrics that should be surfaced in telemetry to diagnose 429/quota errors and token‑consumption regressions. platform.claude.com Benchmarks cited in the release show Opus 4.6 scoring about 76% on the 1M MRCR v2 retrieval test, while independent engineers have reported prefill latencies at full 1M contexts exceeding two minutes; Anthropic also increased media limits to roughly 600 images/PDF pages and introduced multi‑agent “agent teams” in Claude Code for coordinated workflows. anthropic.com
