Provider Model Registry Audit

Summary

A provider model registry audit compares the model IDs, default selections, and hard-coded token limits a tool ships against the provider's current published documentation. This page records one such audit, run on 2026-10-14 against the Google Gemini API model docs (and Vertex/Gemini Enterprise retirement notices) as part of yoyo's issue #1020, which covers "output-ceiling warnings" for the Bedrock and Google arms.

The audit found three classes of drift: (1) a stale default model with a false declared output ceiling, (2) known-model lists containing IDs that do not exist in the provider's catalog, and (3) a Bedrock warning threshold that is a client-side request default rather than a model maximum.

Key Points

• gemini-2.5-flash is yoyo's default Google model (default_model_for_provider in src/providers.rs), but Google now restricts 2.5 access to users who have actively used it before, and directs new projects to 3.5 Flash-Lite or 3.8 Flash. On Vertex / Gemini Enterprise it retires 2026-10-20.
• The declared output ceiling for gemini-2.5-flash is wrong. yoyo warns at 8,192 tokens labelled as a declared output ceiling; the documented output limit is 65,536 (input limit 1,048,576).
• yoyo's known-model lists are stale. src/providers.rs and src/commands.rs list gemini-3.0-pro and gemini-3.0-flash; neither ID appears in Google's catalog.
• The Bedrock warning is a client default masquerading as a model limit. yoyo warns at 16,000 for anthropic.claude-sonnet-4-6, which is yoyo's own request default, not the model's maximum.

Concepts

• Declared output ceiling — a token cap the tool assumes for a model and warns against. Audits must verify these against provider docs, not client configuration.
• Known-model list — the hard-coded set of accepted model IDs (here in src/providers.rs and ). It must track provider catalog changes or it will accept nonexistent IDs.