Styxis
Company
Knowledge

Align token classes and price baselines before comparing public-sector AI usage costs

An offer’s discount becomes interpretable only after its model, token class, environment and price baseline are defined.

Do not jump from discount to cost

A first AI-cost table can begin by ranking the largest advertised discount. Different models, token classes, environments and price baselines can make the same percentage answer different questions. Truthound can check units and categories in an approved cost extract. It does not locate the current cheapest offer, evaluate model quality or establish the savings a particular institution actually achieved.

Retain the scope of a particular offer

GSA announced an expanded Anthropic OneGov agreement on 27 October 2026, describing offer-specific token-based discounts and ordering conditions. Those terms are neither universal prices for every customer nor an institution’s realised savings. Do not assume another provider has identical conditions. Start by distinguishing the announced offer’s scope from the contract actually approved for the institution whose invoice is being examined.

Read one invoice item with its definitions

A proposed introductory table retains model identity, token type, approved environment, price baseline, rate unit and billing period. Staff confirm each field against the contract and invoice. Similar model names do not establish identical charging categories. If a rate cannot be confirmed for the current contract, do not borrow one from another offer merely to fill the missing value in a comparison table.

Make a small comparison about units first

A numeric column mixing usage and currency cannot support a cost comparison from magnitude alone. Combining different token classes into one average rate also needs an explicit interpretation. Use units defined by the actual contract. The first objective is an explanation of one row’s cost basis, not a saving percentage calculated before the underlying model, charge and denominator can be understood.

Apply a limited table check

Truthound Core Python SDK 3.3.4 (open-source SDK, Polars backend) documents ColumnPairInSetValidator for approved model and charging-category combinations. BetweenValidator can check caller-defined bounds and boundary inclusion for a specified unit. The operator supplies pairs, ranges and missingness policy. A plausible numeric value does not establish contractual applicability or invoice validity; it identifies only whether that record meets the configured input conditions before responsible reconciliation.

Retain an existing cost contract where correct

Pandera 0.35.0 (MIT open-source Python DataFrame validation library; pandas, Polars and other supported backends) custom checks can express the same cost-unit and category requirements. Keep a maintained contract that already handles them correctly, considering Truthound only for a control still missing. Neither product’s capability documentation compares AI-model efficiency or institutional total cost of ownership. A validation-interface choice therefore supplies no universal conclusion about which AI vendor offers the better service or financial outcome.

One explainable row is a useful beginning

Connect one invoice item to the approved model, environment and unit, returning anything unexplained to its owner. Pause a simple comparison when periods or baselines differ. Expand under the same rule after that first explanation works. Trustworthy cost reporting begins with a record of who applied the rate and on which grounds, rather than an attractive discount whose practical scope cannot be established.

Conditions

Define contract-approved model, token class, environment, price baseline and billing period.

Alternatives

Retain a Pandera 0.35.0 contract for equivalent cost-unit and boundary checks.

Limitations

  • An offer-specific discount is not a universal current rate, realised saving or model-performance result.

Evidence

Event checked
Subject checked
Modified