How we choose AI models: evidence and methodology
We recommend AI models by the job they can documentably support. Capabilities are sourced and dated; editorial task choices explain their reasoning. We publish benchmark numbers only after running a recorded, comparable test.
Three kinds of evidence
Verified capabilities come from provider documentation or a dated Higgsfield catalog export. Integration-specific settings do not automatically establish provider-wide limits, native resolution or a subscription entitlement. Each capability keeps its own source and verified date.
Editorial verdicts interpret those capabilities for a real creative brief. They are judgments about workflow fit, not measured quality, realism, motion, speed or commercial performance. Recommendations can include models outside our affiliate platform.
Original benchmarks require output files, the exact model ID, prompt, inputs, settings, generation date, duration, dimensions, job ID and actual credits. Our image snapshots publish n=1 outputs, requested and backend model IDs, native sizes and preflight credit estimates. Video benchmarks await original clips.
What the finder measures
The finder applies hard filters for documented input, native audio, portrait framing and required duration. Unknown support fails a hard requirement. It then ranks workflow fit using task-specific editorial shortlists, task tags and clearly labeled draft preferences.
Internal matching points are routing logic, not a model-quality score. Affiliate availability and commission do not enter ranking. If nothing fits every constraint, the finder says so instead of silently relaxing them.
A fair original test
- Keep the subject, prompt, starting image and delivery frame consistent.
- Record the exact variant and settings; disclose unavoidable differences such as 5s versus Veo 4s.
- Retain the first successful output; record failed jobs and rerolls separately.
- Judge sound separately from picture and do not mix native audio with added tracks.
- Inspect physical contact, identity, text, motion and prompt adherence with a written rubric.
- Publish the clips and provenance before assigning any benchmark score.
Freshness and corrections
Capability dates mean that the linked evidence was reviewed on that date, not that a new generation was run. Page revisions track content hashes and change only when substantive source content changes. Prices, model access and promotions require a checkout-time check.
Send a correction with the model, disputed claim and primary source. The contact mailbox is shown but delivery is not yet verified.
Owned dataset
Our structured capability records and editorial metadata are licensed CC BY 4.0 with attribution to What’s That Model? Linked provider documents, trademarks and any third-party source material retain their original rights. The license is not a grant of rights to third-party model outputs.