Workload before modelNo universal winner
The right model depends on what counts as success, how hard failure is to spot, and how much human cleanup remains. Start with that decision—not a leaderboard.
Explore the test desksWorkload explorer
These are research ladders, not published recommendations.
Decision question: where does a lower-cost model stop producing a verifiable repository change without disproportionate steering or repair?
Decision question: when does a stronger model materially improve source coverage, contradiction handling, calibration, or the cost of checking the brief?
Decision question: when does paying more remove enough structural, factual, or voice-preservation editing to change the economics?
Decision question: where do cross-page relationships, schema pressure, and plausible silent errors justify a stronger tier?
Decision question: how much responsibility can move from human to model before supervision, verification, and recovery erase the benefit?
Publication rule: a ladder becomes a buyer’s guide only after its acceptance bar, exact model records, repeated trials, and limitations are visible.
Before choosing a tier
What artifact must exist, and what can be checked deterministically?
Is a failure obvious and cheap, or plausible and expensive to discover?
Include prompting, supervision, correction, verification, and recovery.
Test the cheapest serious option first. Escalate on evidence, not status.
Current externally researched guide
A practical starting choice and escalation rule using current API terms—clearly separated from future firsthand verdicts.
Repository coding
Compare Luna, Terra, Fable 5.1, and Astra; define the acceptance bar; then make a stronger model prove where it changes the outcome.
Read the full buyer guide →What this page will not become
Workload guides will inherit conclusions from actual review records. They will not convert model cards, affiliate economics, or benchmark ranks into a universal recommendation.
See the review desk