{"version":"2.1.295","anchor":"context-1m-refusal-tracking-and-gateway-usage-now-works-in-m","canonical_anchor":"context-1m-refusal-tracking-and-gateway-usage-now-works-in-m","heading":"Claude Code remembers when a model refuses the 1M-token context","tier":"internal","area":"Long Context","scope":"individual","heads_up":false,"url":"https:\/\/changelogs.core-directive.com\/v\/2.1.295\/e\/context-1m-refusal-tracking-and-gateway-usage-now-works-in-m","release_url":"https:\/\/changelogs.core-directive.com\/v\/2.1.295","markdown":"### Claude Code remembers when a model refuses the 1M-token context\n\nClaude Code now keeps track, per model, of models that refused the 1-million-token context window\n\n**What**\n\nSome models can work with a context window, the amount of text the model can consider at once, of 1 million tokens. Claude Code now keeps a per-model record of models that refused that larger window, next to its existing record of refused optional features.\n\nTwo smaller changes ship with it:\n\n- Reading usage from a gateway (a server between Claude Code and the model) now also runs in a case where it was skipped before.\n\n- An update to the instructions MCP servers provide can now include instructions attached to it. MCP servers are add-on tool servers that connect to Claude Code.\n\n**Why**\n\nOnce a model refuses the 1-million-token window, Claude Code can remember that for the model.\n\n- Area: Long Context\n- Tier: Under the hood\n- Useful: 1\/5\n- Signal: 1\/5\n- Scope: individual\n- Heads-up: no"}