Context Awareness ⚠ verify

Definition

A built-in capability where the API injects tags telling the model how much of its context window ("token budget") remains, so it can self-manage long tasks.

Key points

• Present on Sonnet 5, Sonnet 4.6, Sonnet 4.5, Haiku 4.5. Automatic — the API injects the tags; no config needed. ⚠ verify current model coverage (2026-07-08).
• System-prompt injection at start: <budget:token_budget>200000</budget:token_budget> (budget matches the available window: 1M for Sonnet 5 / Sonnet 4.6, 200k for Sonnet 4.5 / Haiku 4.5).
• Post-tool-call update: <system_warning>Token usage: 35000/200000; 165000 remaining</system_warning> (image tokens included).
• Newer models (Opus 4.7+, Fable 5, Mythos 5) do not receive injected tags — use task budgets (beta) to give an explicit budget instead.

Why it matters for the exam

• Tests knowing which models get automatic context awareness vs which need task budgets, and recognising the exact <budget:...> / <system_warning> tag syntax.

Common gotchas

• Opus models (4.7+, 4.8) do NOT get injected context-awareness tags — a common trap. They rely on task budgets instead.
• The budget value equals the model's window, not a fixed 200k.

See also

• Context Windows By Model
• Context Rot
• model_context_window_exceeded behaviour
• Adaptive thinking and effort

Sources

• docs-models-pricing
• netlight-study-guide