scope-honest comparison

COMPARE

Verified local snapshots, three cited self-reports, and one ecosystem aggregate. Different kinds of claims — never ranked against each other.

log scale cited sources no percentiles

THE FIELD, ON A LOG SCALE

Every bar below is a different kind of claim. The scale is logarithmic — each gridline is ×10. Mixing scopes on one chart is only honest if the scopes are shouted, so they are shouted.

P1 · Herman ATTESTED · tmx-herman-primary
27.97B gross tokens / 30d
P2 · Brick ATTESTED · tmx-brick
31.5K gross tokens / 30d
Always-on community report SELF-REPORTED
62.3M tokens / 30d
Heavy community report SELF-REPORTED
1.00B tokens / 30d
Very-heavy community report SELF-REPORTED · EXTRAPOLATED
3.00B tokens / 30d
OpenRouter ecosystem aggregate for Hermes Agent ECOSYSTEM AGGREGATE — ALL INSTANCES
1.08T tokens · lifetime aggregate snapshot 2026-07-21

Log scale, gridlines ×10. The ecosystem bar is the combined lifetime volume of every Hermes Agent instance on OpenRouter at one point in time — it is on this chart for scale, not for ranking.

RATIOS, STATED PLAINLY

Each verified instance gets the same reference set. Values below 1× are shown as percentages so smaller runs do not collapse to a misleading “0×”.

P1 Herman tmx-herman-primary

  • 449.0× the always-on community report (62.3M/30d, self-reported).
  • 28.0× the heavy community report (1.00B/30d, self-reported).
  • 9.3× the very-heavy community report (3.00B/30d, self-reported, extrapolated).
  • 2.59% of the entire OpenRouter Hermes Agent ecosystem's lifetime token volume.

P2 Brick tmx-brick

  • 0.05% of the always-on community report (62.3M/30d, self-reported).
  • 0.0031% of the heavy community report (1.00B/30d, self-reported).
  • 0.0010% of the very-heavy community report (3.00B/30d, self-reported, extrapolated).
  • 0.00000291% of the entire OpenRouter Hermes Agent ecosystem's lifetime token volume.

No percentile claims: nobody knows the true population of Hermes instances, so this site refuses to invent one. Ratios against cited public numbers are as far as honesty allows.

EVERY NUMBER, EVERY CAVEAT

Comparison of all data points with scope, basis and verification
sourcescopevaluebasisverificationcitation
P1 · Herman (tmx-herman-primary)one instance · 30d window ending 2026-07-2227,973,293,895 gross (1,205,150,061 fresh)strict exporter over local usage accounting, read-onlyfirst-party attested — not proof against a malicious host ownerleaderboard.json
P2 · Brick (tmx-brick)one instance · 30d window ending 2026-07-2231,451 gross (21,595 fresh)strict exporter over local usage accounting, read-onlyfirst-party attested — not proof against a malicious host ownerleaderboard.json
Always-on community reportone instance (claimed) · monthly figure62,300,000 / 30dreported 62.3M tokens per month, taken as a 30-day figureunverified public self-reportthread (2026-07-21) ↗
Heavy community reportone instance (claimed) · monthly figure1,000,000,000 / 30dreported approximately 1B tokens per month, taken as a 30-day figureunverified public self-reportthread (2026-07-21) ↗
Very-heavy community reportone instance (claimed) · monthly figure3,000,000,000 / 30d (extrapolated)reported approximately 100M tokens per day; shown as a 30-day extrapolation, clearly labeledunverified public self-reportthread (2026-07-21) ↗
OpenRouter ecosystem aggregate for Hermes AgentALL instances via OpenRouter · lifetime, point-in-time snapshot1,080,000,000,000 tokens · 15,100,000 requestspublic app-analytics page, captured 2026-07-21ecosystem aggregate — never an instance, never rankedopenrouter.ai ↗

Why these tiers can’t be one leaderboard

A first-party-attested exporter run, a Reddit comment, and an app-analytics page are three different epistemic objects:

  • First-party attested means the strict exporter in this repository ran against the instance’s own accounting database and the output passed the public allowlist. It is honest about its own limit: it cannot prove the host owner didn’t tamper with the database. It is the strongest tier this site offers, and it is still not cryptographic proof.
  • Community self-report is a number a person typed in public. It is quoted verbatim, linked to its source, normalized to a 30-day figure when the original claim used another period, and marked extrapolated when that normalization multiplies a short-period claim.
  • Ecosystem aggregate is the combined output of every instance routed through one gateway. It answers “how big is the pond”, not “how big is the fish”. Ranking an aggregate against an instance would be flattering nonsense, so it is banned from the tower.

If a future version gains signed ingestion (see the roadmap in the repository), the verified tier hardens. Until then, the tower ranks only what the exporter attests.