Model Lifecycle Policy

last updated

Lifecycle states

Every model on the platform is either active (runnable through its provider adapter today) or retired (its provider shut it down, or we removed its adapter). There is no in-between: if a battle cannot be run against a model today, users cannot gather new evidence about it, and stale evidence should not occupy the default board.

What retirement means

Retired models are excluded from the default leaderboard and from the random battle pool. Their vote history is preserved — it still informs other models’ Bradley-Terry scores through the pairwise comparisons they participated in — and their rows remain queryable in the ranking snapshot for anyone auditing historical results.

When we retire a model

A model is retired when its provider removes or shuts down the underlying API model, when the adapter consistently fails with model-not-found class errors, or when a preview model is superseded and disabled by its provider. Retirement is recorded in code with a date and reason, and takes effect on the next ranking snapshot.

Currently retired

modelretiredreason
openai:o4-mini-deep-research2026-07-23Shut down early by OpenAI (jobs fail with model_not_found)
gemini:deep-research-pro-preview-12-20252026-06-30Removed from Google's models list; superseded by 04-2026 previews

Models removed from an adapter without an explicit entry above are treated as retired automatically — the board only medals models that can actually be run.