EOSOptimizeA worked example, static
Spend less on the models you already have
Declare what your organization already has access to and the work you actually run; every option is priced against that workload and ranked, with the ones it would not use and why, and the choice stays with the person who made it. Worked through below on fixed, illustrative figures. This optimizer is not part of the EOS service: it was served by a separate AssetShop service (the estate API), which this page no longer calls, and it is not in the platform either. No model is called and no price is invented: every figure is a declared input or arithmetic on one.
A fixed example on illustrative figures. Nothing on this page signs in, calls a service or calls a model; interactive tools live in the platform.
1 · A declared catalog, one workload, the plan Modeled, a fixed example
Three options an organization already has, at the rates it declared per million tokens, in USD. The workload: supplier note summarization, 200,000 calls a month, 1,800 input tokens per call of which 1,200 repeat on every call, 250 output tokens per call, and answers that may be tolerant. The named baseline is Option C, what runs today.
| Option | Tier | Input, cached, output per M | Per call | Per month |
|---|---|---|---|---|
| recommendedOption B | small | 0.80, 0.08, 4.00 | 0.001576 | 315.20 |
| Option A | standard | 3.00, 0.30, 15.00 | 0.005910 | 1,182.00 |
| Option C, the baseline | frontier | 15.00, 1.50, 75.00 | 0.029550 | 5,910.00 |
Per call is 600 new input tokens at the input rate, 1,200 repeated tokens at the cached rate and 250 output tokens at the output rate; per month is per call times 200,000. Against the baseline, Option B would cost 5,594.80 less a month, which stays a claim about the future until a month is recorded.
2 · What actually happened Simulated: a recorded month, never projected
A person adopts Option B. Then a month is recorded: 204,600 calls for 322.40 actually spent. The baseline would have cost 204,600 times 0.029550, or 6,045.93, so that month cost 5,723.53 less, recorded rather than projected.
How the two layers feed each other
Activity in the spatial layer produces records; EOS reconciles them, keeps the evidence and routes the decision; what EOS settles is meant to return to the spatial layer as a capability people can use. Each stop below states what it is today, not what it could become.
Read this literally: the two EOS stops are live in the service and tested; one spatial stop is a scaffold and one is planned. Nothing here claims a live connection between the two layers, because none is deployed.
Boundary: this page signs nobody in, calls no service and calls no model; it is a fixed example, and the optimizer it shows is not part of the EOS service. A recommendation is a proposal until a person with the right role adopts it, and a saving is a claim about the future until a month is recorded.
The cash flow impact simulator: terms, MOQ and lead time, worked through on fixed figures →