How we rank AI Gateway models

What the weekly picks mean

• Privacy lane keeps models where the catalog marks both ZDR and no-training as all or some. That matches Vercel's ?zdr=true and ?npt=true filters.
• Open lane ranks the full catalog, including models that train or skip ZDR.
• Capable means tool-use plus at least 128,000 tokens of context. Vision is not required.
• Blended price is 3× input + 1× output, in $ / 1M tokens. Privacy ranks prefer a cheaper ZDR endpoint when one exists.
• Value is 7-day mean token share divided by blended $ / 1M.
• Bang-for-buckis DeepsecBench score divided by that run's cost. Frontier prefers the highest Deepsec score among capable models; we ignore scores below 12 when ranking bang.
• Cheap router blends at or under $0.50 / 1M. Workhorse blends at or under $3.00 / 1M.

Sources

• AI Gateway catalog
• Model adoption leaderboard
• Lab adoption leaderboard
• DeepsecBench results

Frequently asked questions

What does ZDR + no training mean?

ZDR is zero data retention: the provider does not keep prompts. No training means prompts are not used to train models. A privacy pick requires the AI Gateway catalog to mark both `zdr` and `no_training` as all or some, matching the official ?zdr=true and ?npt=true filters.

How is bang-for-buck calculated?

Bang-for-buck is DeepsecBench score divided by that run's cost. Value score is 7-day mean token share divided by blended $/1M, using a 3× input + 1× output mix. Discounted picks use a cheaper ZDR endpoint than list price.

Which models count as capable?

Capable models support tool use and have at least 128,000 tokens of context. Vision is not required. Cheap routers must also blend at or under $0.50 / 1M; workhorses at or under $3.00 / 1M.

Is this an official Vercel product?

No. bestmodels.dev is an independent weekly ranking. Catalog, adoption, and DeepsecBench numbers come from Vercel AI Gateway data licensed CC BY 4.0. We are not affiliated with Vercel.