Canonical review recordsSelection is editorial
We would rather publish five defensible reviews than five hundred scraped summaries. Inclusion should mean a model can lower the cost floor, raise the capability ceiling, or change whether an upgrade is worth paying for.
Read the publication barExternally researched profiles
Exact IDs, current API terms, integration constraints, and the evidence that still needs to be collected.
$0.20 input / $1.20 output per million at standard short-context rates.
Open profile →Escalation candidateThe same-generation option with a 10× like-for-like token premium.
Open profile →Premium candidateLong-horizon positioning, distinctive cache rates, and tool restrictions.
Open profile →Frontier candidateCurrent frontier terms, access questions, and long-context costs.
Open profile →Review anatomy
Every published record will separate the thing tested, the setup used, what happened, and the conclusion a buyer can safely draw.
MODEL / EXACT ID / DATE
One plain recommendation—who should use it, who should not, and the job where its premium starts to matter.
Open test register
Status describes the protocol, not model performance.
From an isolated fix to an ambiguous, long-running repository job.
From bounded synthesis to conflicting sources and verification-heavy conclusions.
From schema-bound extraction to cross-document reasoning with silent-failure risk.
Living records
The tested identity and material buying conditions still match the record.
A model, alias, price, tool, or harness changed enough to revisit the decision.
The record remains historical evidence but should not guide a current purchase.
The tested option is no longer available for the reviewed use.
Research is not a rating
The repository-coding guide turns these four candidates into a bounded, evidence-conscious buying sequence.
Read the buyer guide