Anthropic's 2024 Claude model card and system prompt evaluation suite found that pushing Claude past 80% context utilization drops instruction-following accuracy by up to 35% on multi-step coding tasks. That is not a rounding error. It is the difference between a clean refactor and a broken import chain that compiles but fails at runtime.
The model does not run out of tokens. It runs out of attention budget. When the context window fills, the effective search space for the next token expands across too many competing signals. The system prompt, the file tree, the conversation history, and the current diff all fight for the same limited attention head capacity. Precision on multi-step logic collapses because the model starts skipping boundary checks and hallucinating function signatures.
I saw this last week on a 14-file React migration. The prompt was 18k tokens, well under the 200k limit but sitting at 92% utilization after system instructions and prior turns. Claude generated a new useAuth.ts hook that referenced validateSession from auth.utils.ts, except that function had been renamed to verifySession three turns earlier. The model had the information. It simply lost the binding. Dropping the active context to 65% by summarizing old turns and pruning the file map fixed the drift immediately.
Keep a running token counter in your orchestration layer. When you cross 75%, archive old messages to a summary block and strip inactive file contexts. The pattern is simple: treat the context window like a cache, not a log. Your pass rate recovers.
Top comments (0)