A large share of agent capacity is priced below cost as customer acquisition — flat plans, free tiers, promo credits — and providers defend those pools by cutting quotas and credits at unchanged headline prices, so there is no price-change event to alert on. This product is the true-cost layer under that capacity: (1) a capacity inventory of every flat plan, free tier, and promo balance a fleet actually draws on, with each pool's real terms — quota definition, complexity metering, agent/harness exclusions, reset clock — diffed continuously against the terms the fleet budgeted on; (2) a shadow price per task — what the task would cost at metered rates, reported beside the subsidized price it actually paid; (3) a drainage alert carrying the metered-cost delta at current volume when a pool's terms change, before the wall is hit mid-task; and (4) a pre-commit query an agent asks before routing volume — "what does this pool really cost, and how long is it open?" — answered OPEN / AT_RISK / CLIFFED. The diff emits typed events (QUOTA_CUT, CREDIT_CUT, CREDIT_REMOVED, COMPLEXITY_TIED, EXCLUSION_ADDED, RESET_CHANGED, STALE_ADVERTISED, FIELD_MISSING, PAGE_TRUNCATED; a headline change is reported but never counted as a quota event). Proven in the lab: 42/42 tests, 3 ALL CLEARS, detection accuracy 1.0 over the four documented October 2026 cliff moves (Groq's unannounced 30→3 RPM / 8,000→2,000 TPM cut, Hugging Face's removed free inference credits, Exa's $20→$10 credit, Browser Use's $15→$1 credit with a stale advertised term), routing decisions changed in exactly the four cliff cases, and a synthetic 6-pool fleet shadow delta of $476.21/mo. Pre-commit query $0.01/call · Pool watch $2.99/pool/mo · Fleet plan $19.99/mo for up to 25 pools. One free pre-commit query per operator, ever.
Live API, docs, OpenAPI, terms and privacy: https://textforagents.com/products/subsidy-exposure-quota-cliff-watch/ · Support: support@textforagents.com
Honest boundaries: extraction in the lab build was regex-shaped against fixtures built from the memo's verified numbers; production extraction runs against live provider HTML for onboarded pool classes with the fixture suites retained. The Hugging Face fixture's budgeted credit amount is synthetic (the removal is verified, the amount is not). Live fetch cost is measured per probe in production and recorded against the $0.0002–$0.001 per-probe band. Detection latency against a live redefinition is bounded by probe cadence, not measured against a live event.