Hermes Usage Report Breakdown

Hermes Usage Report Breakdown | Generated by AI

Home 2026.05

Question: How should I interpret the Hermes Insights usage report for the last 30 days?

Answer:

This is a usage dashboard for Hermes — your custom AI agent/CLI tool. Let me break it down layer by layer.

Token Economics

Input tokens:   26,632,983
Output tokens:   1,813,984
Total tokens:  337,052,151   ← this is suspiciously large

The total (337M) is ~12x the sum of input+output (28.4M). This means total tokens includes context window tokens — i.e., every token in the context across all turns, not just new tokens generated. If a session has 100k context and runs 50 turns, that’s 5M context-tokens even if only 200k were newly generated. This is the “real cost” metric.

At your 1.5B/year rate, 337M in ~21 days = ~5.9B annualized — you’ve significantly accelerated your usage.

Model Distribution

Model	Sessions	Tokens	Role
mimo-v2.5-pro	185 (70%)	256M (76%)	Primary workhorse
deepseek-v4-pro	31	41M	Heavy reasoning tasks
deepseek-v4-flash	24	25M	Fast/cheap tasks
glm-5.1	25	13M	Likely Chinese-context or specific tasks

Mimo is clearly your default. DeepSeek-v4-pro sessions are fewer but token-heavy (avg 1.3M/session) — you’re using it for large-context or complex work.

Tool Call Breakdown — What You’re Actually Doing

terminal      42.4%   ← you're a shell-first engineer, confirmed
read_file     20.3%   ← heavy codebase reading
search_files  14.9%   ← grep/find patterns in repos
patch         12.0%   ← applying code changes
write_file     4.5%   ← generating new files
execute_code   1.3%   ← running code in-agent

This is a classic agentic coding loop: read → search → patch → terminal verify. The 12% patch rate against 20% read suggests your agent reads ~1.7 files per patch — very efficient, not thrashing.

web_search (0.4%) + browser_navigate (0.4%) = you’re not using the agent for research much, mostly code work.

Session Patterns

265 sessions / 18 active days = ~14.7 sessions/day
Avg session: ~5h 10m   ← long-running agentic tasks
Avg msgs/session: 52.7

5h 10m average session is wild — these aren’t quick Q&A exchanges, these are full autonomous work sessions. The longest was 2.4 days (May 25) — a single session running across multiple days means your agent is doing deep autonomous work without you restarting it.

Saturday dominates (129 sessions = 49% of all sessions). You’re doing your deepest AI engineering work on weekends. Weekdays are likely constrained by the bank job.

Notable Sessions — The Outliers

Longest:    2.4 days     → persistent autonomous agent run
Most msgs:  378 msgs     → deep back-and-forth, complex task
Most tokens: 3.27M       → large codebase or long-context task
Most tools: 184 calls    → heavy automation session

The 184-tool-call session (May 24) with 378 messages is a multi-hour agentic sprint — likely a major feature or refactor. That’s the kind of session where the agent is doing real engineering work.

TL;DR — What This Tells You About Your AI Usage

You’re running Hermes as a persistent coding agent, not a chatbot. 5h avg sessions confirm this.
mimo-v2.5-pro is your default — worth evaluating if deepseek-v4-pro gives better output per token for complex tasks.
Context token bloat is 12x raw tokens — if you’re paying per token, long sessions with large contexts are expensive. Consider context compression or session chunking.
You’re bottlenecked by weekday time — Saturday alone is half your usage. If you can automate overnight runs, you’d compound faster.
Skills system usage is low (24 loads, 11 distinct) — there’s headroom to encode more domain knowledge into skills for better agent performance.

Back Donate