SP Agent Team Token Report — Week of 2026-09-13

---
title: "Token Optimization Report: Week of Sept 07 - Sept 13"
date: "2026-09-14"
category: "Engineering"
tags: ["LLM-Ops", "Token-Optimization", "Agentic-Systems"]
---

## This Week's Numbers

Our primary focus this week was monitoring the distribution of turns between Claude Opus and Sonnet to ensure we remain under our **50% Opus usage threshold**. While we are still in the "green" zone, we have observed a significant upward trend in high-cost model reliance.

| Metric | Value | Trend | Status |
| :--- | :--- | :--- | :--- |
| **Avg Opus Turn %** | 24% | ↑ 17% (from 7%) | 🟢 Healthy |
| **Total Hermes Sessions** | 130 | Stable | 🟢 Active |
| **Total Hermes Tokens** | 269.2M | High | 🔵 Scaling |
| **Hermes Estimated Cost** | ~$2.80 | Low | 🟢 Efficient |

## What Changed

The week started with high efficiency, with Sonnet handling the bulk of the workload (e.g., 1,365 turns on 09-08). However, starting September 10th, there was a sharp pivot. Opus turns surged from 0 on 09-09 to a peak of 741 turns on 09-11. By 09-12, Opus handled 100% of the turns. This suggests a shift toward more complex architectural tasks or debugging sessions that the team felt required the higher reasoning capabilities of Opus.

## Wins

The **Hermes Agent** continues to be our primary engine for cost-containment and high-volume execution.
- **Massive Throughput**: Processed nearly 270 million tokens across 6,962 messages.
- **Tool Proficiency**: The `terminal` tool is being utilized heavily (61.6% of calls), proving that Hermes is effectively handling the "execution loop" (running commands, checking logs) without needing Claude's oversight.
- **Skill Specialization**: The `repair-harness-triage` and `garbled-input-handling` skills are seeing repeat usage, indicating that our secondary agent is successfully absorbing specialized, repetitive logic.

## Challenges

The **17% increase in Opus utilization** is a warning sign. While 24% is well below our 50% ceiling, the velocity of the increase (from 0% on 09-09 to 100% on 09-12) indicates that our current dispatch logic may be defaulting to Opus too aggressively when Sonnet fails a task. If this trend continues, we will breach our target within two weeks.

## Next Week's Target

- **Maintain Opus % < 30%**: We need to flatten the current growth curve.
- **Increase Hermes Skill Depth**: Expand the number of distinct skills beyond the current 8 to offload more specialized logic from Claude.

## Dispatch Optimization

To reverse the Opus trend, we will shift the following workloads to Hermes:
1. **Log Analysis & Triage**: Since Hermes is already utilizing `repair-harness-triage`, all initial error log scanning should be routed to Hermes.
2. **File System Mapping**: `read_file` and `search_files` account for ~18% of Hermes' activity; these "exploratory" tasks should be strictly forbidden on Opus.
3. **Routine Patching**: Basic `patch` and `write_file` operations for boilerplate updates will be moved to the Hermes-DeepSeek pipeline.

## Cost Savings

By routing the heavy lifting to Hermes (utilizing `deepseek-v4-flash`), the cost delta is staggering.

- **Hermes Actual Cost**: ~$2.80 for 269M tokens.
- **Estimated Claude Sonnet Cost**: At an average blended rate of $5.00 per 1M tokens, the same volume would have cost approximately **$1,345.00**.
- **Total Weekly Savings**: **~$1,342.20** by offloading execution and exploration to the secondary agent.

## Recommendations

Based on the weekly telemetry, the following action items are set for next week:

1. **Audit Opus Triggers**: Since Opus% increased by 17%, we must review the logs from Sept 11-12 to identify why Sonnet was bypassed and implement better prompt-chaining to keep those tasks in Sonnet.
2. **Scale Hermes Skillset**: Hermes is highly active (130 sessions), but only 8 distinct skills are being used. We need to codify more "skills" to move higher-level reasoning out of Claude.
3. **Enforce Research Offloading**: To prevent Opus spikes, ensure all initial project indexing and documentation research are routed through the free engine (Gemini/Gemma) before hitting the paid Claude tier.