Proof
Public evidence for the figures on this site. Every claim below links to the artifact that substantiates it — the provider usage dashboard, and the underlying usage export.
Context Cache Management
Middle-removal context editing normally breaks prefix caching. Harmony holds a 98% input-token cache hit on the long-context primary — 98.0% in June 2026, and 98.1% in August, held through middle-removal editing and new-project churn. Flash swarm fan-out carries distinct contexts per call, so there is nothing to cache.
Verifies: provider usage dashboard (volume, requests, model split) + usage export (cache-hit rates) — June and August 2026, below.
June 2026 — Scale
DeepSeek Platform usage, 1–30 June 2026, all keys. 23,522 requests · 2,564,252,156 tokens (2,551,474,411 input + 12,777,745 output). Primary model input-token cache hit rate: 98% (light blue = cached).
The CSV export removes the account ID and pricing columns; these dashboard captures are unedited, so DeepSeek's displayed rates at the time of capture remain visible. Rates change — verify at DeepSeek's pricing page
August 2026 — Currency
DeepSeek Platform usage, 19–25 August 2026, all keys. 2,527 requests · 383,201,160 tokens. Primary model input-token cache hit rate: 98.1% — held at ~98% through middle-removal editing.
The CSV export removes the account ID and pricing columns; these dashboard captures are unedited, so DeepSeek's displayed rates at the time of capture remain visible. Rates change — verify at DeepSeek's pricing page
Cache-Hit Export
The underlying usage export, with the account ID and pricing columns removed. Summing the input_cache_hit_tokens and input_cache_miss_tokens columns reproduces the 98% rate above.
| Period | Model | Input cache hit | Input cache miss | Input hit rate |
|---|---|---|---|---|
| June 2026 | deepseek-v4-pro (primary) | 2,496,870,912 | 50,119,623 | 98.0% |
| June 2026 | deepseek-v4-flash | 37,504 | 4,446,372 | 0.8% |
| Aug 19–25 2026 | deepseek-v4-pro (primary) | 369,985,792 | 7,088,208 | 98.1% |
| Aug 19–25 2026 | deepseek-v4-flash | 20,608 | 475,682 | 4.2% |
Recordings
Screen recordings of the dashboard being navigated (secondary evidence).
Scope
Eternal Harmony integrates many AI providers and gateways through The Harmony Extension, plus local models. This page documents one: DeepSeek, whose usage export is the cleanest record we can publish. The figures below are a floor, not the whole picture.
What this does not establish
These figures cover a single provider. They verify volume and cache behaviour — not how the work was done, or whether any of it was good. A cross-provider total would add operational detail without strengthening the evidence; for further verification, we will share additional details with prospective partners, researchers, journalists, and collaborators on reasonable request.