SP Agent Team Token Report — Week of 2026-08-02

---
title: "Weekly Token Optimization: The Opus Spike and Hermes Scaling"
date: "2026-08-03"
category: "Engineering"
tags: ["LLM-Ops", "Token-Optimization", "Claude", "DeepSeek"]
---

## This Week's Numbers

Our token efficiency metrics for the period of **2026-07-27 to 2026-08-02** indicate a significant deviation from our cost-optimization targets.

| Metric | Value | Status |
| :--- | :--- | :--- |
| **Avg. Opus Usage** | 92% | 🔴 Red Light |
| **Previous Week Avg** | 60% | $\uparrow$ 32% Worsening |
| **Total Turns Added** | 516 | - |
| **Opus Target** | < 50% | Not Met |

The most alarming trend occurred between July 29th and 31st, where Opus usage hit **100%**, peaking on July 30th with **402 turns**. This suggests a heavy reliance on high-reasoning models for a specific sprint or complex bug-fixing cycle.

## What Changed

The data shows a sharp pivot toward Opus mid-week. While we started the week at 71% Opus usage, we surged to a total monopoly of the high-end model by Wednesday. This coincided with a massive spike in activity on Thursday (July 30), where we saw the longest session of the week (1.5 days) and nearly 1,000 messages in a single thread. We are essentially "over-reasoning" on tasks that may not require Opus-level intelligence.

## Wins

The **Hermes Agent** (our secondary orchestration layer) is showing incredible scaling potential. 
- **Throughput:** Handled 98 sessions and 4,218 messages.
- **Efficiency:** Processed over **301 million tokens**, primarily utilizing `deepseek-v4-flash`.
- **Utility:** Successfully executed 1,832 tool calls, with `terminal` (40.2%) and `execute_code` (16.3%) being the primary drivers. This proves Hermes can handle the bulk of the "heavy lifting" for environment interaction and code execution.

## Challenges

The primary challenge is the "Opus Leak." The current routing logic is failing to downgrade tasks to Sonnet or Hermes, leading to a 92% Opus average. At this rate, our token budget will be exhausted prematurely. We are currently treating the primary agent as a reasoning-only engine rather than a dispatcher.

## Next Week's Target

**Objective: Bring Opus usage down to < 50%.**
We need to implement stricter routing constraints to ensure that only architectural decisions and complex debugging remain on Opus, while standard implementation shifts to Sonnet or Hermes.

## Dispatch Optimization

Based on the tool usage data from Hermes, we propose shifting the following tasks from Claude to Hermes:
- **Environment Setup & Testing:** Move all `terminal` and `execute_code` tasks to Hermes.
- **File System Operations:** Shift `read_file` and `patch` operations to Hermes to reduce Claude's context window bloat.
- **Routine Verification:** Offload `static-html-verification` and `github-monitoring` to Hermes' specialized skills.

## Cost Savings

By offloading 4,218 messages to Hermes (DeepSeek) instead of Claude Sonnet, the savings are substantial.

- **Hermes Cost (DeepSeek):** $4,218 \text{ messages} \times \$0.003/\text{query} \approx \mathbf{\$12.65}$
- **Estimated Sonnet Cost:** $4,218 \text{ messages} \times \approx \$0.05/\text{avg query} \approx \mathbf{\$210.90}$
- **Weekly Savings:** $\approx \mathbf{\$198.25}$

## Recommendations

Based on this week's telemetry, we are implementing the following action items:

1. **Route more implementation to Sonnet:** Since Opus% (92%) is well above the 50% threshold, we will force-route all non-architectural coding tasks to Sonnet.
2. **Use free engines more:** With only 1 session utilizing `gemini-3.6-flash` out of 98 total Hermes sessions, our free-tier research utilization is $< 3\%$. We will shift initial codebase exploration to Gemini.
3. **Review rate enforcement:** Given the massive spike in turns on July 30 without a corresponding increase in `evaluate.sh` tool calls in the top list, we suspect a low review rate. We will implement enforcement hooks to ensure every 10 turns are followed by a validation check.