---
name: token-budget-saturation
description: Estimate the reasoning-token point where additional model budget yields diminishing quality gains relative to latency and cost. Use for inference-budget planning, model evaluations, and adaptive reasoning policies.
---

# Token Budget Saturation

Find a practical reasoning budget from observed quality at several token limits.

## Workflow

1. Collect quality, latency, and cost measurements at three or more reasoning-token budgets.
2. Sort measurements by budget and calculate marginal quality gain between adjacent points.
3. Normalize marginal gain by added tokens, latency, and cost.
4. Locate the first point where the marginal gain remains below the chosen threshold.
5. Check whether difficult task subsets saturate later than the overall benchmark.
6. Recommend a default budget plus an escalation rule for uncertain cases.

## Output

Return the saturation estimate, marginal-gain table, default budget, and escalation condition.

Do not extrapolate beyond the measured range without clearly labeling the estimate.

## Research basis

[Token Budget Saturation and Mechanistic Early Detection of Reasoning Non-Convergence in Chain-of-Thought Models](https://arxiv.org/abs/2607.21433) — Oladri, Jawahar, and Mohamed, arXiv (2026). The paper introduces and empirically studies the exact phenomenon.