SP Agent Team Token Report — Week of 2026-06-28

---
title: "Weekly Token Optimization Report: Fighting Opus Gravity"
date: "2026-06-29"
category: "Engineering"
---

## This Week's Numbers

Our focus remains on reducing the reliance on Claude Opus in favor of Sonnet and our secondary agent, Hermes. Here is the breakdown for the week of June 22 to June 28:

| Day | Opus Turns | Sonnet Turns | Opus% | Trend |
| :--- | :--- | :--- | :--- | :--- |
| 06-22 | 119 | 5 | 95% | - |
| 06-23 | 78 | 6 | 88% | ↓ |
| 06-25 | 509 | 27 | 86% | → |
| 06-26 | 1307 | 249 | 83% | ↓ |
| 06-27 | 823 | 291 | 71% | ↓ |
| 06-28 | 130 | 30 | 76% | ↑ |
| **Avg** | **80%** | **Target: <50%** | **Status** | 🟡 Yellow |

## What Changed

The most significant event this week was the surge in activity on June 26th, where we saw a peak of 1,307 Opus turns. While we saw a slight dip in Opus% toward the end of the week (hitting a low of 71% on the 27th), the weekly average remains stagnant at 80%, barely moving from last week's 78%. We are currently in a "Yellow Light" state—stable, but not progressing toward our efficiency target.

## Wins

The real victory this week lies in the **Hermes Agent's** performance. Hermes successfully handled a massive volume of work that would have been prohibitively expensive on Claude:
- **Total Token Volume:** 188,362,441 tokens.
- **Model Efficiency:** `deepseek-v4-flash` handled the lion's share, processing over 131 million tokens.
- **Tooling Integration:** Hermes is now a powerhouse for environment interaction, with the `terminal` tool accounting for 50.6% of all tool calls.

## Challenges

The "Opus Gravity" is strong. Despite our efforts to route tasks to Sonnet, the team still defaults to Opus for the majority of complex turns. The gap between our current average (80%) and our target (<50%) indicates that our dispatch logic or developer habits are not yet optimized for Sonnet's capabilities.

## Next Week's Target

**Opus% < 70%**. We need to break the 80% plateau by aggressively shifting implementation tasks to Sonnet.

## Hermes Agent Utilization

Hermes has evolved into our primary "workhorse" for high-token, low-reasoning tasks. 
- **Sessions:** 102 sessions with an average of 39.4 messages per session.
- **Platform Distribution:** The TUI is the dominant interface, contributing 119.6M tokens, proving that our developers prefer the terminal-integrated experience for heavy lifting.
- **Top Models:** DeepSeek-v4-flash (70% of tokens) and GLM-5.2 (29% of tokens).

## Dispatch Optimization

Based on Hermes' tool usage (Terminal: 50.6%, Read File: 13%), we should shift the following from Claude to Hermes next week:
1. **Log Analysis & Grepping:** Move all `terminal` and `search_files` heavy tasks to Hermes.
2. **Initial Code Exploration:** Use Hermes for the first 5-10 turns of any new project to map the file structure before engaging Claude.
3. **Repetitive Patching:** Move standard `patch` and `write_file` operations to the Hermes/DeepSeek pipeline.

## Cost Savings

By routing 4,021 messages through Hermes (primarily DeepSeek) instead of Claude Sonnet, the savings are substantial.

- **Hermes Cost (Approx):** $4,021 \text{ queries} \times \$0.003 \approx \$12.06$
- **Estimated Sonnet Cost:** For 188M tokens, Sonnet would cost approximately $564 (est. $3/M input).
- **Weekly Savings:** $\approx \$552$

## Recommendations

Based on this week's data, the following actions are mandatory for next week:

1. **Route more implementation to Sonnet:** With Opus% at 80%, we are over-relying on the most expensive model for tasks Sonnet can handle.
2. **Increase Free Engine Adoption:** While Hermes is efficient, we must further integrate Gemini/Gemma for initial research to bring costs down further.
3. **Enforce Dispatch Hooks:** Since the Opus% trend is stable rather than declining, we will review the `evaluate.sh` enforcement hooks to ensure task routing is being followed.