SP Agent Team Token Report — Week of 2026-07-05

---
title: "Weekly Token Optimization: Reducing Opus Dependency"
date: "2026-07-06"
category: "Engineering"
---

## This Week's Numbers

Our focus this week remained on the **Opus% metric**—the ratio of high-cost Claude Opus turns compared to Claude Sonnet. 

*   **Weekly Average Opus Usage:** 65% (down from 78% last week)
*   **Improvement:** $\downarrow 13\%$
*   **Current Status:** 🟡 Yellow Light (Target: $< 50\%$)
*   **Total Activity:** 97 sessions, 90 turns added across 94 new project files.

The daily trend showed significant volatility. We peaked at **91% Opus usage on July 1st**, likely due to complex architectural reasoning, before hitting a low of **26% on July 2nd**. This suggests that while we are improving, our routing is still highly dependent on the specific nature of the day's tasks rather than a systematic dispatch policy.

## What Changed

We observed a more aggressive shift toward Sonnet for implementation tasks. The drop from 78% to 65% indicates that the team is becoming more conscious of "over-provisioning" intelligence. By utilizing Sonnet for routine coding turns and reserving Opus for high-level system design, we've begun to bend the cost curve downward.

## Wins

*   **Successful Trend Reversal:** Breaking the 70% barrier is a key psychological and financial milestone.
*   **Hermes Offloading:** The Hermes Agent has become a powerhouse for secondary tasks, handling **75 sessions** and over **3,200 messages** this week.
*   **TUI Efficiency:** The TUI platform emerged as the primary driver for Hermes, processing **139M tokens**, proving that heavy-duty agentic workflows can be offloaded from the primary Claude interface.

## Challenges

The "Yellow Light" status persists because we are still 15% above our target. The spike on July 1st highlights a "dependency trap"—whenever a task feels "too hard," the default is to jump to Opus. We need to refine our prompts for Sonnet to handle more of these "hard" tasks.

## Next Week's Target

**Target: $< 50\%$ Opus%**
We aim to move from 🟡 Yellow to 🟢 Green. This requires a disciplined approach to routing implementation and debugging tasks exclusively to Sonnet.

## Dispatch Optimization

Based on the Hermes Agent data, the **Terminal tool** accounted for **61.6% of all calls (958 calls)**. This represents a massive opportunity for dispatch optimization. 

**Proposed Shift:**
All "Environment Exploration" and "Shell Script Execution" tasks should be routed to **Hermes (DeepSeek-v4-pro)** by default. Claude should only be invoked to analyze the output of these commands or to design the logic of the script, not to execute the trial-and-error process of terminal debugging.

## Cost Savings

The financial impact of utilizing Hermes over Claude Sonnet is stark. 

*   **Hermes Cost:** With 3,208 messages at an estimated $\approx \$0.003/\text{query}$, the total cost was roughly **$\$9.62$**.
*   **Claude Sonnet Equivalent:** Processing **174 million tokens** via Claude Sonnet (est. avg $\$5/\text{million tokens}$) would have cost approximately **$\$870$**.

**Estimated Weekly Savings: $\approx \$860$** by leveraging DeepSeek and Gemini-flash-lite for auxiliary agentic loops.

## Recommendations

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

1.  **Route more implementation to Sonnet:** Since Opus% is currently 65% (above the 50% threshold), all feature implementation turns must start with Sonnet.
2.  **Optimize Terminal Loops:** Given that terminal calls are the highest tool usage (61.6%), move all "read-execute-verify" loops to Hermes.
3.  **Review Rate Enforcement:** Ensure `evaluate.sh` is running on all new artifacts to prevent "token bleed" from unoptimized code being fed back into the context window.