1m-context-era-token-budget-recalibration-5-10x-from-200k
1M-context era: recalibrate token budgets by 5-10x from 200k-era numbers
Context: many tool-output-offload, subagent-return, and attention-dilution thresholds in 2024-2025 were calibrated for ~200k context windows. Opus 4.7 ships with 1M context. Naively keeping the old thresholds underutilizes the larger window AND creates false-positive throttling.
Rule: every “N tokens” or “N KB” budget threshold calibrated for 200k context should be audited and either (a) raised by 5-10x when the real concern is attention-dilution not raw context headroom, or (b) annotated [RECALIBRATED <date>] with pointer to new canonical number.
Specific recalibrations (2026-04-19 session, applied to arjtech-dev setup):
| Budget | Old (200k calibration) | New (1M calibration) | Rationale |
|---|---|---|---|
| Tool output offload | 2,000 tokens | 10,000 tokens | Below this, context-bombing risk is negligible; attention-dilution is the new concern |
| Sub-agent return budget | 1,000-2,000 tokens | 20,000 tokens | Attention-dilution kicks in around 10-20k per single turn even with 1M headroom |
Non-bottleneck: raw context. Filling 1M with useful data is fine. Real bottleneck: attention-dilution. Beyond ~20k tokens in a single turn, orchestrator reasoning quality degrades — sharp falloff on complex multi-step decisions.
Don’t auto-migrate: preserve evidence-based history. Annotate original thresholds with [RECALIBRATED YYYY-MM-DD] inline pointer to current canonical. Rewriting the original claim loses provenance.
Pattern for other 200k-era budgets:
- Memory file truncation cliffs (platform-enforced, NOT budget-adjustable) → keep
- Line-count targets (signal-density > line-count in 1M era) → raise 2-3x if cohesive
- Context-window utilization alarms (80% of 1M is 800k — different operating point than 80% of 200k) → recalibrate
Applicability: any agent/orchestrator running on 1M-context models (Claude Opus 4.7, etc.). Pattern generalizes beyond Claude — whenever model class shifts, token budgets need explicit recalibration rather than inherited defaults.