Anthropic Launches Claude 4 with 1M-Token Context Window

Full summary

Anthropic has released Claude 4, available in Sonnet and Opus tiers. The headline change is a 1 million token context window — five times longer than the previous generation — paired with what the company calls "persistent memory", a capability that lets the assistant retain durable state across multiple sessions without manual context-stuffing. Claude 4 also introduces refinements to the Model Context Protocol (MCP), with smaller payloads and better error recovery during long agentic runs. Coding benchmarks show meaningful gains on multi-file refactor tasks. The Sonnet tier is positioned as the new default, with Opus reserved for high-stakes reasoning. Pricing remains in line with the previous generation.

Why it matters

Long-context plus persistent memory together remove a real ceiling: codebases that previously needed clever retrieval can now be reasoned about end-to-end, and ongoing projects don't need to rebuild context every session.

Technical explanation

Context window scales to 1M input tokens; output remains capped at 128k. Persistent memory is implemented as a per-account vector store with selective retrieval — Anthropic emphasizes it's not training data but per-user runtime state.

Business impact

Enterprises evaluating AI for legal review, M&A diligence, and codebase-spanning tasks now have a credible long-context alternative to RAG pipelines. Tool vendors building on Claude can rely on more durable agent state.