AI Inference Market Analytics — Surplus Intelligence

Market analytics

Aggregate marketplace usage and realized pricing.

Last 28 UTC calendar days · generated 22:39 UTCYour activity →

28-day usage

81,739,063Requests

4.1TInput tokens

75.1BOutput tokens

3.2TCache tokens

Input is gross: each UTC day's bar is total input tokens, split into uncached input and cache (read + write), so the composition shows the cache share. Output is charted on its own axis alongside. Current day is partial; the faded cap shows the projected full-day total.

Input tokens

uncachedcache

240B120B0

Sep 17Sep 30Today

Output tokens

5B2.5B0

Sep 17Sep 30Today

Daily requests

5M2.5M0

Sep 17Sep 30Today

Daily savings

$60K$30K0

Sep 17Sep 30Today

Token usage by model

deepseek-v4.1-flashglm-5.3-flashglm-5.3gpt-6-lunaGPT-5.6 LunaOther

Sep 17Sep 30Today

Token methodology

usage_log.input_tokens is gross: it includes cache. The stacked bar splits it into uncached = input_tokens − cache_read_tokens − cache_write_tokens, and cache = cache_read_tokens + cache_write_tokens. Output is on its own axis because it is about 30× smaller. The model split shows each day's tokens across the top 5 models by volume, with the rest folded into Other.

7-day realized pricing

26,974,643Eligible requests

89.5%Mean discount

95%Median bucket

95–100%Peak bucket

Realized discount distribution

Share of eligible requests per 5% bucket

mean: 89.5%

median bucket: 95%

<1%

15%

59%

<0%0–55–1010–1515–2020–2525–3030–3535–4040–4545–5050–5555–6060–6565–7070–7575–8080–8585–9090–9595–100100%+

Discount methodology

Realized discount = 1 - buyer_cost_usdc / direct_cost_usdc. The distribution uses the 7 UTC days ending today and includes paid non-fallback rows where direct and buyer costs are both greater than zero. Median is reported as the lower edge of the median 5% bucket to keep high-volume windows fast and stable. Mean discount is the exact request-weighted mean; the median bucket is the 5% band that holds the middle request.