Problem#
An alias and the model it points to are two unlinked rows. aliases is set on 4 rows across openai, anthropic, gemini, grok, groq, mistral and byteplus (3 Anthropic, 1 Gemini). sameAs is null on every row of those seven. Snapshot 2026-11-21.
#112 added aliases for Anthropic. The other providers still have the gap:
- mistral: 18
-latest rows, for example mistral-large-latest, codestral-latest, ministral-8b-latest, magistral-medium-latest. None names the dated snapshot it resolves to, and no dated row lists its alias.
- openai:
gpt-4o and gpt-4o-2024-08-06, gpt-4o-2024-11-20 are separate rows with no link. The same holds for other dated snapshots.
- gemini:
gemini-pro-latest, gemini-flash-latest and gemini-flash-lite-latest are their own rows. Only gemini-3-flash-preview lists an alias.
Impact#
A consumer cannot tell that two ids are one model. TanStack AI has to guess with id patterns (a date suffix, -latest) to avoid listing a model twice.
Ask#
- For each alias row, state the row it resolves to. For each resolved row, list its aliases. Use whichever of
aliases and sameAs is meant for this.
- Where the provider does not say what an alias resolves to, add a
factSources entry that says so.
Acceptance#
- Each Mistral
-latest row and each OpenAI dated snapshot is linked to its counterpart, or has a factSources entry.
Fixed by #720, merged and deployed. The forge did not close this on merge although the link was set, so closing by hand. Alias groups are served in
aliases; an alias row names its target inaliasOfwhere the provider states one (OpenAI model pages, Gemini changelog) and carries anunstatedsource where it does not (Mistral states the group, not the direction).