model-availability

Daily snapshots of the model catalogs offered by the major AI providers, using Simon Willison’s git scraping technique. The git history is the dataset: git log -p -- data/ shows exactly when a model appeared, changed, or was removed, and each commit message summarizes that day’s delta.

What’s tracked

Provider	Catalog	Snapshot
AWS Bedrock	Runtime/Foundation models (us-east-1)	data/providers/bedrock/foundation_models.json
AWS Bedrock	Mantle models (us-east-1)	data/providers/bedrock/mantle_models.json
Azure	AI Foundry pay-as-you-go (MaaS) models	data/providers/azure/foundry_models.json
Google	Gemini API models	data/providers/google/gemini_api_models.json
Google	Model Garden third-party PAYG models (us-central1)	data/providers/google/model_garden_models.json
Anthropic	First-party models	data/providers/anthropic/models.json
OpenAI	First-party models	data/providers/openai/models.json

Every JSON snapshot has a sibling .csv for human review. Scope is deliberately pay-per-token: self-deploy / managed-compute offerings (Azure managed compute, Model Garden deployables, Hugging Face mirrors) are filtered out.

How it works

One workflow per provider (.github/workflows/), daily, staggered an hour apart so pushes never race. Scrapers are single-file uv scripts (PEP 723) under scripts/providers/; auth is keyless throughout (GitHub OIDC → provider WIF/role; Azure’s catalog API is anonymous). A run commits only if something actually changed, with a deterministic message from scripts/diff_summary.py.

The design decisions that keep the diffs honest:

• Canonicalization (lib/common.py): dict keys sorted, unordered arrays sorted, so provider APIs that shuffle response order between requests can’t churn the snapshots.
• Volatile fields dropped: server-stamped timestamps, popularity rankings, and other per-request noise are stripped so a diff means the catalog really changed.
• Stable ordering: models are emitted oldest-to-newest (or by a pinned order list), so new models append at the bottom of the file and diffs stay minimal.
• Drift detection: new models are free, but a previously-committed model vanishing upstream fails the run for human review. Azure additionally re-fetches once before trusting a disappearance, since its API occasionally flickers.
• Drift acknowledgement, phone-friendly: the failing run commits a <snapshot>.pending-drift.json marker recording exactly the drift it saw. Re-running the failed workflow (“Re-run failed jobs”, which the GitHub mobile app can do straight from the failure notification) accepts exactly that recorded drift and deletes the marker in the same commit that applies it. Drift the marker doesn’t cover fails again and re-records, and scheduled runs never auto-accept, so unacknowledged drift keeps failing daily until a human re-runs. Locally, --accept-pending is the same acknowledgement; --allow-drift still accepts everything unconditionally.