Spanvero separates a model call's sticker price from the expected cost of completing the task. Every probability, retry, evaluation, and recovery assumption remains visible.
Outcome Cost = first attempt + expected retries or fallbacks + evaluation + (final failure probability × recovery cost). Success rates are conditional: a retry's pass rate means “given that earlier attempts failed.” Presets are scenarios; a rate becomes measured only with a named TaskSpec, success test, sample, and run record.
A GPU host earns more when you rent more; a router earns a cut of every token. Spanvero can show when hardware you already own has the lower marginal cost because it sells none of those routes. Electricity and allocated hardware cost remain editable assumptions.
Non-estimated rent-a-GPU and API rates are captured from provider-published sources. A local route may use $0 only as an intentional marginal-compute assumption on hardware you already own; electricity and allocated hardware cost can be entered instead. Spanvero never adds a cut or a fee.
Non-estimated API and rental rates are captured from provider-published listings or endpoints and stamped with a date (current set: 2026-08-03). A missing sourced rate is shown as a clearly labeled “est.” size-based estimate; a retired or unavailable upstream rate is marked “last-known.” The label states the evidence we have, not live provider verification.
Published tables rank by the stated price field. Outcome scenarios rank by visible receipt inputs. Referral payouts never change either order, and a provider cannot pay to improve a success rate. When no measured rate exists, the tool asks for an assumption instead of inventing a quality score.
The local route is shown next to paid options with a calculated VRAM estimate and an explicit hardware-fit caveat. When the hardware is already yours, you can compare $0 marginal compute or enter electricity and allocated hardware cost instead.
Since June 20, 2026 we log the full cost landscape once a day into an append-only, version-controlled record that is never overwritten — a neutral, public time-series of what AI really costs. No vendor can produce a credible version of this; we can, because we're independent.
By default your provider API keys live in your own browser and never touch our servers — a “your own key” call goes straight from you to the provider. We're an interface and an advisor, not a middleman holding your credentials or compute.
The one-time Decision Pack sells a polished report workflow while the calculator and JSON stay free. Flat-fee Outcome Cost audits are available now; dated data packs will only be sold once the evidence record supports them. Disclosed vendor referrals and Amazon category-search links may also pay us, but never affect a price, route, or ranking. The public price feed remains free to cite.
Open the Outcome Lab → · Measured audits · All models · Download the data
Prices as of 2026-08-03. $0 markup, your own accounts, we never resell or hold compute. © 2026 Cynosure LLC.
A short email of real AI price moves, straight from the daily log — no hype. We're collecting the list now; the first issue goes out when it opens. Unsubscribe with one click.