modelalive 0.4.0

Pre-flight check: is this LLM model ID still alive?

Model Alive

Pre-flight check before every LLM API call: is this model ID still alive?

Hardcoded model IDs break silently until production fails. Anthropic retired Claude Sonnet 4 and Opus 4 on June 15, 2026. Mythos Preview retires June 30, 2026. Model Alive answers in one call — with source links and breaking change notes.

Install

pip install modelalive

Quick examples

# Retired model → exit 1
modelalive check claude-sonnet-4-20250514

# Resolve alias + replacement
modelalive resolve claude-3-5-sonnet-latest

# List all retired Anthropic models
modelalive list --status retired --provider anthropic

# Validate registry integrity
modelalive validate --strict

# Pre-flight: print safe model ID (auto-replaces retired)
modelalive ensure claude-sonnet-4-20250514

# Production CI: fail on unknown models too
modelalive check "$MODEL_ID" --strict-unknown --warn-deprecated

import modelalive

# Pre-flight: auto-replace retired models
model_id = modelalive.ensure("claude-sonnet-4-20250514")
# → "claude-sonnet-4-6"

result = modelalive.alive("claude-mythos-preview")
print(result.days_until_retirement)  # 2 (as of 2026-06-28)
print(result.source_url)             # Anthropic deprecation docs
print(result.confidence)             # "verified"

HTTP API

uvicorn api.main:app --port 8787

Method	Path	Description
GET	/v1/alive?model=	Lifecycle check (410 if retired)
POST	/v1/alive/batch	Check up to 100 models
GET	/v1/ensure?model=	Pre-flight: return safe model ID
GET	/v1/resolve?model=	Best model ID + breaking changes
GET	/v1/registry	Filter by status, provider
GET	/v1/sources	Official doc URLs + last checked date
GET	/v1/validate	Registry integrity report
GET	/v1/health	Version + model count

CI gate (GitHub Actions)

- uses: Ahaozandaburada/modelalive@main
  with:
    models: claude-sonnet-4-20250514 gpt-5-2025-08-07
    warn-deprecated: "true"

Or plain shell:

- run: pip install modelalive && modelalive check ${{ env.MODEL_ID }}

Trust & accuracy

Every verified entry includes:

• source — provider key (anthropic, openai, google)
• source_url — official deprecation page (returned in API/SDK)
• source_checked_at — date registry was last synced with docs
• confidence — verified (in registry) or unknown (not listed)

Registry rules enforced in CI:

• Retired/deprecated models must have replacement + retired_at
• Past retired_at auto-marks model dead even if status lags
• Aliases resolve before lookup (claude-3-5-sonnet-latest → dated snapshot)
• modelalive validate --strict fails on errors before release

Canonical registry: edit registry/models.json, then:

python scripts/sync_registry.py   # copies to package bundle
modelalive validate --strict
pytest

Sources (as of 2026-06-28):

• Anthropic model deprecations
• OpenAI API deprecations

Response shape

{
  "queried_model": "claude-sonnet-4-20250514",
  "canonical_model": "claude-sonnet-4-20250514",
  "alive": false,
  "status": "retired",
  "provider": "anthropic",
  "retired_at": "2026-06-15",
  "replacement": "claude-sonnet-4-6",
  "breaking_changes": [],
  "source_url": "https://platform.claude.com/docs/en/about-claude/model-deprecations",
  "source_checked_at": "2026-06-28",
  "confidence": "verified"
}

Coverage

Provider	Active	Deprecated	Retired
Anthropic	Sonnet 4.6, Opus 4.8, Haiku 4.5, Mythos 5, …	Opus 4.1, Mythos Preview	Sonnet 4, Opus 4, 3.x series
OpenAI	gpt-5.5, gpt-4o, …	gpt-5 snapshots, gpt-5.2-chat-latest, …	gpt-4-0314, gpt-3.5-turbo-0301, …
Google	gemini-3.5-flash, …	2.5 Flash, embeddings, image previews	gemini-2.0-flash (June 1), gemini-3-pro-preview

Unknown models return status: unknown, confidence: unknown, alive: true — we never block what we haven't verified.

Docker (self-host API)

docker compose up --build
curl "http://localhost:8787/v1/alive?model=gemini-2.0-flash"

Deploy to Fly.io:

fly launch --copy-config --yes
fly deploy

Keep registry fresh

pip install modelalive[http]
python scripts/refresh_sources.py   # fetch provider docs, update checked_at
modelalive validate --strict
python scripts/sync_registry.py

Weekly automated refresh runs via .github/workflows/registry-refresh.yml.

Sources (as of 2026-06-28):

• Anthropic model deprecations
• OpenAI API deprecations
• Google Gemini deprecations

License

MIT