FLUX — verified LLM cost reduction, billed only on savings you can recompute
FLUX is a drop-in wrapper for your OpenAI or Anthropic client that reduces what you spend per completed task — and proves it. You pay a percentage of verified savings. If your traffic is already optimised, FLUX tells you not to pay us.
What FLUX saves, by traffic type
Every figure below is measured per completed task with quality held (at least 36 of 40 tasks fully correct and complete), on a simulated harness using each provider's published prices and cache rules — except where marked live. A first real-traffic export exists (30 requests, measured 9–25% per agent); your own pilot replaces every band with your number before any fee exists. How these numbers are made →
| Your traffic looks like | Measured saving | Label |
|---|---|---|
| Multi-turn Claude cache-only comparison | 37.6% vs the same J prompt stack with cache disabled; both arms clean 38/40 | simulator-measured; cache effect isolated |
| Full J + 5m cache stack vs raw uncached baseline | 57.1% lower cost, −5 pp clean-task quality | fails the agreed −2 pp quality limit; not a qualified win |
| Background / batch / latency-tolerant jobs | 57.7% (net of assumed tier fallbacks; 64.3% ceiling) | simulator-measured |
| One-shot API calls on a shared template | 45.5% — up to 72.8% where latency-tolerant | simulator-measured |
| Interactive agents, sparse traffic | 34.0% | simulator-measured, caveats apply |
| Interactive agents, busy traffic | 25.9–29.4% | simulator-measured |
| Already-optimised stacks | 0–5% — do not pay us | measured on fixtures |
The spread is the honest part: savings are a property of your workload's shape, not of FLUX. The first thing FLUX does is tell you which row you are — from your own logs, in about five minutes, before any contract exists. Get a quote from your own export →
How it works, in one paragraph
Your bill is tokens × price-per-token × execution tier. FLUX attacks all three at once: it removes tokens that carry no information (oversized tool results capped at injection, exact duplicates, schema bloat) without ever breaking your provider's prefix cache; it repairs prompts so cached pricing (10× cheaper input) actually applies; and it routes traffic you flag as latency-tolerant to the provider's own discounted tiers (50% off, same model, same answer). Risky changes are enabled only after a replay gate proves, on your own recorded requests, that quality held.
Why you can trust the number on your invoice
- Savings are computed from your provider's own reported usage, never our estimates.
- Every request writes a hash-chained ledger event; your invoice is a pure function over those events that you can recompute yourself, to the cent.
- Savings you already had — your provider's cache, your gateway's cache — are itemised and never billed.
- A changed answer is a failed task, and failed tasks void their savings.
Methodology · Per-segment detail · Honest comparison with alternatives · FAQ · Every number's provenance