SP Agent Team Token Report — Week of 2026-09-06

---

title: “Weekly Token Optimization Report: August 31 - September 6”
date: “2026-09-07”
category: “Engineering”

This Week’s Numbers

Our token routing strategy has seen a significant shift this week, moving away from high-cost reasoning models toward more efficient execution engines.

Metric	This Week	Last Week	Trend
Avg Opus Usage	7%	62%	$\downarrow$ 55%
Target Opus%	< 50%	-	🟢 Green
Total Hermes Tokens	135.3M	-	$\uparrow$
Hermes Sessions	97	-	Active

The daily breakdown shows a near-total elimination of Opus dependency on 09-01, 09-02, and 09-06, with only minor spikes during mid-week complex tasks (peaking at 12% on 09-04).

What Changed

The primary architectural shift this week was the aggressive delegation of “execution-heavy” tasks to the Hermes Agent. By utilizing deepseek-v4-flash as the primary driver for secondary agentic workflows, we’ve decoupled high-level orchestration from repetitive file system operations and terminal commands.

Wins

• Drastic Opus Reduction: Dropping from 62% to 7% Opus usage represents a massive optimization in our cost-per-turn without sacrificing agentic capability.
• Hermes Stability: The Hermes Agent handled 4,419 messages across 97 sessions with a high success rate, particularly in CLI-based tasks.
• Tool Efficiency: We’ve successfully offloaded the bulk of terminal (54.8%), read_file (15.9%), and write_file (10.1%) calls to the secondary agent.

Challenges

While the overall trend is positive, we saw a temporary regression on 09-03 and 09-04 where Opus turns spiked to 247 and 267 respectively. This suggests that certain complex project contexts still trigger a fallback to Opus when the secondary agent fails to resolve the initial prompt.

Next Week’s Target

Target: Maintain Opus% < 10%
We aim to stabilize the routing logic to prevent mid-week spikes and further refine the “skill” triggers for the Hermes Agent.

Dispatch Optimization

Based on the high utilization of terminal and read_file in Hermes, we propose shifting the following tasks from Claude Sonnet to Hermes next week:

1. Initial Project Indexing: Move all search_files and initial directory crawls to Hermes.
2. Boilerplate Generation: Route standard file writes and patch applications to the deepseek-v4-flash pipeline.
3. Routine Monitoring: Fully migrate github-monitoring and ad-hoc-verification skills to be Hermes-first.

Cost Savings

The efficiency gain is most apparent when comparing the Hermes expenditure to a hypothetical Sonnet-only baseline for the same volume of tokens.

• Actual Hermes Cost: ~$1.31 (for 135.3M tokens via DeepSeek).
• Estimated Sonnet Cost: Assuming an average blended rate of $5.00 per 1M tokens, the same 135.3M tokens would have cost approximately **$676.50**.
• Total Weekly Savings: ~$675.19

Recommendations

Based on the current data telemetry, our action items for next week are:

1. Maintain Routing Logic: Opus% is well below the 50% threshold (currently 7%); continue with current dispatch rules.
2. Expand Tooling Integration: Hermes is highly utilized (97 sessions); explore moving more “research-heavy” skills into the Hermes framework to further lower Sonnet load.
3. Audit Mid-Week Spikes: Analyze the 09-03/09-04 Opus spikes to identify if specific “Skill” failures forced a model upgrade.