Fable Brain

Context Window Management

5 min read

Context Window Management

Claude Fable 5 runs a 1M-token context window by default, and Anthropic’s primary management strategy is server-side compaction, backed by context editing, curation, and persistent artifacts for multi-window work.

What it is

  • Claude Fable 5, Claude Mythos 5, Opus 4.8, 4.7, and 4.6, Sonnet 5, and Sonnet 4.6 have 1M-token context windows; “For every model with a 1M-token context window, 1M is the default” (https://platform.claude.com/docs/en/build-with-claude/context-windows.md), with no beta header and standard pricing.
  • Claude Fable 5 can generate up to 128K output tokens per request; older models such as Sonnet 4.5 have 200K-token windows.
  • Anthropic frames context as “a precious, finite resource” with an attention budget that depletes as tokens accumulate; the goal is the smallest set of high-signal tokens that maximizes the outcome.
  • Harnesses budget context explicitly: the claude.ai system prompt opens with a declared budget of 190000 tokens inside a budget:token_budget block (System Prompt Export 2026-07, L1-7).

How it works

  • Context rot: “As token count grows, accuracy and recall degrade, a phenomenon known as context rot.” (https://platform.claude.com/docs/en/build-with-claude/context-windows.md). More space is not automatically better.
  • Server-side compaction (beta header compact-2026-01-12, context_management edit type compact_20260112) summarizes earlier conversation into a compaction block when input tokens reach the trigger threshold, default 150,000 and minimum 50,000; Anthropic positions it as the primary strategy for long conversations and agentic workflows.
  • Compaction billing adds an iteration: total tokens are the sum across usage.iterations, not the top-level fields.
  • Client contract: append the entire response.content, including compaction blocks, back into messages; when the API receives a compaction block it ignores all content blocks before it.
  • Context editing (beta header context-management-2025-06-27) clears rather than summarizes: clear_tool_uses_20250919 clears the oldest tool results with defaults of a 100,000 input-token trigger and keeping the 3 most recent tool uses; clear_thinking_20251015 controls preserved thinking turns and must be listed first when the two are combined.
  • Editing executes server-side before the prompt reaches Claude, so the client keeps the full unmodified history.
  • Launch evidence (2025-09-29, with Sonnet 4.5): memory plus context editing improved agentic search 39% over baseline, editing alone 29%, with an 84% token reduction on a 100-turn web search evaluation.
  • Long-context ordering: for prompts of roughly 20K tokens or more, “Place your long documents and inputs near the top of your prompt” (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/long-context-tips.md), above the query, instructions, and examples.
  • Context awareness: Sonnet 5, Sonnet 4.6, Sonnet 4.5, and Haiku 4.5 receive an injected token budget and remaining-capacity updates from the API.
  • Multi-window work: Anthropic’s harness pattern uses an initializer session, one feature per session, and persistent artifacts (a JSON feature list, descriptive git history, a progress file), because “each new session begins with no memory of what came before.” (https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents). Compaction alone is not sufficient for sustained multi-session progress.
  • Sub-agent isolation: focused agents return distilled summaries of roughly 1,000 to 2,000 tokens instead of raw context.
  • March 2026 update: with Opus 4.6, sprint decomposition and context resets could be replaced by one continuous session with automatic compaction.

Automatic context awareness is documented for Sonnet 5, Sonnet 4.6, Sonnet 4.5, and Haiku 4.5; whether Claude Fable 5 receives the same injected budget updates is not attested in this claim pack.

Best practice

  • Rely on the 1M default on Fable 5; no beta header or premium pricing is needed. EVIDENCE-BASED
  • Enable server-side compaction for long agentic runs as the primary strategy. EVIDENCE-BASED
  • Always append the full response.content, compaction blocks included, on the next request. EVIDENCE-BASED
  • Use context editing to clear stale tool results, and list clear_thinking_20251015 first when combining edit types. EVIDENCE-BASED
  • Place long documents near the top of prompts of 20K tokens or more. EVIDENCE-BASED
  • In harnesses that compact, tell Claude its context will be compacted so it does not wrap up work early. EVIDENCE-BASED
  • When a window is cleared, consider a fresh context window with state rediscovered from the filesystem; it can beat compaction. EVIDENCE-BASED
  • For work spanning many windows, maintain persistent artifacts: feature list, git history, progress file. EVIDENCE-BASED
  • Curate aggressively; target the smallest set of high-signal tokens rather than filling the window. EVIDENCE-BASED

Pitfalls

  • Appending only the text of a compacted response silently loses compaction state.
  • Cached prompt prefixes still occupy the context window; caching changes billing, not token accounting.
  • Context rot: large windows degrade accuracy and recall when filled indiscriminately.
  • Reading top-level usage fields under compaction undercounts; totals live in usage.iterations.
  • Clearing tool results invalidates cached prefixes at the clearing point, trading cache hits for space.
  • Assuming compaction alone sustains multi-session projects; the harness guidance says it does not.

Sources

Next actions

  • Verify whether Fable 5 receives automatic context awareness injections.
  • Record compaction trigger tuning results for one long agentic run in this vault.
  • Compare fresh-window filesystem rediscovery against compaction on a real multi-session task.

Graph View

Backlinks