What is the billable workload?
Measure representative input, output, cached, tool, image, and audio usage. A headline token rate rarely describes the complete invoice.
Purpose guide · Cost-sensitive API use
How should you evaluate AI models for cost-sensitive API use?
Catalog current through July 14, 2026
Evaluation framework
A source-bounded evaluation plan for cost-sensitive deployments. No cost shortlist is published because Epoch's catalog does not include pricing.
Selection rule
No models are shortlisted or ranked. Epoch's catalog identifies access type but does not include the normalized input, output, cached-token, batch, tool, or media pricing needed for a defensible cost comparison.
Evidence dependency
Evidence first. Your evaluation next.
The required evidence is identified before any comparison is attempted.
No cost shortlist published
Pricing dependency: add current first-party pricing artifacts with effective dates, units, currencies, tiers, and model aliases before publishing any cost comparison.
Decision criteria
Questions the source can help you ask.
These checks narrow an evaluation plan. They do not replace hands-on testing.
What must pricing evidence contain?
Capture the official provider URL, effective date, currency, unit, tier, model alias, batch or cache discounts, and retrieval artifact for every price used.
What is cost per successful task?
Combine verified usage cost with your own quality, retry, latency, and reliability tests. A lower unit price can cost more when a workflow needs more calls or human correction.
Evidence boundary
What this guide does not know.
No task-level quality, price, context window, latency, or reliability measurement is present in the Epoch catalog. We therefore publish no ranking or “best” label.
Catalog source: Epoch AI Data on AI Models, retrieved 2026-07-14, CC-BY. Artifact sha256:3c663b17ec989f00ef5f994214b878a914118e2d1fb6077e23329a318209eb86.
Each listed model links to its model-level source above. See the publication source ledger and methodology for the wider evidence policy.