Token Effectiveness

7d window • generated 2026-10-13T06:25:28Z

⚠ configured (estimated_cost_usd). Wire real pricing via hermes configure for accurate dollar figures.

Cost / success

$0.0021

1223 successes (317 msgs + 906 crons)

Useful-token ratio

0.0055

output ÷ (input+output+cache_read) = 2,362,740 / 433,043,098

Tokens / success

1,366,066

input+output+cache_read per delivered msg

Cron success rate

90.5%

906 / 1001 cron runs

1 — Cost per successful task

Successes = delivered messages (Slack delivery_obligations.state=delivered) + cron completions. Cost = sessions.estimated_cost_usd when configured, fallback rate otherwise.

Source	Sessions	Cost USD
cli	171	0.0000
cron	154	0.0000
slack	38	2.5477
subagent	3	0.0000

configured (estimated_cost_usd)

2 — Useful-token ratio

High (>0.10) = model is delivering dense output. Low (<0.02) = heavy context load, sparse output — likely over-researching or large skill dumps.

Source	Sessions	Input	Output	Cache read	Ratio
cli	171	1,904,386	226,788	3,646,598	0.0393
cron	154	8,224,926	1,085,879	114,549,007	0.0088
slack	38	14,131,315	1,020,627	287,133,235	0.0034
subagent	3	132,824	29,446	958,067	0.0263

3 — Tokens to solution + retries

Tool-call distribution: n=366, median=4, p90=52, p99=132, max=569, zero=32.2%

Zero-tool-call sessions are typically single-shot lookups. p90/p99 reveal runaway loops.

4 — Compression A/B

Prior data point (2026-09-03): 29% reclaim on clawy 166-msg = 97,000 tok, 22 tool-result messages replaced.

Current window input: 24,393,451 tok. Estimated reclaim: 7,074,100 (29%) or 12,196,725 (50%).

To run the actual A/B: 1 wk compression OFF (compression.proactive_prune_tokens=0) → 1 wk ON (=48000), same workload. Compare total_in deltas.