🌅 llm-sunset

Your app is about to break because an AI provider is shutting down the model it uses. llm-sunset tells you before it happens.

⏰ OpenAI shuts down gpt-4, gpt-3.5-turbo, o1, o3-mini and ~30 more models on Oct 23, 2026. If any of those are hard-coded in your repo, you have 16 days. Run one command and find out.

OpenAI, Anthropic, Google, Groq, Cohere and xAI retire models constantly, often with only a few months' notice. The model ID sits hard-coded in some config file and nobody notices until production starts returning 404 model_not_found.

llm-sunset scans your code, finds every AI model ID, and checks it against a daily-updated list of official deprecation notices.

pipx run llm-sunset     # no install needed

$ llm-sunset
  error  gpt-4o-2024-05-13  OpenAI     in 16d          2026-10-23
         → gpt-5.6-sol
         app/llm.py:12:23 · app/summarize.py:8:11

  warning  claude-sonnet-4-5  Anthropic  in 54d        2026-11-30
           → claude-sonnet-5-5
           worker/.env:2:7

  error  claude-2.0           Anthropic  retired 443d ago  2025-07-21
         → claude-opus-4-8
         config/models.yaml:3:8

2 models in 3 files  ·  2 errors  ·  1 warning  ·  fail-within 90d · data: bundled snapshot

• Zero dependencies. Pure Python standard library, Python 3.8+.
• Works offline. Ships with a bundled snapshot; uses live data when it can reach it.
• No API keys, and your code never leaves your machine. It only downloads a public JSON file.
• CI-ready. Comes with a GitHub Action, a pre-commit hook, SARIF output for GitHub code scanning, JSON and Markdown.
• Readable output. Colourised, column-aligned and grouped by model, with the shutdown date, how long you have left and the replacement on one block per model. Auto-detects your terminal width and never spams a 3,000-character line; --color always|never and NO_COLOR/FORCE_COLOR override the default.
• Low noise. It matches exact model IDs (so gpt-4o-mini is not reported as gpt-4o, and gpt-4.1 is not reported as gpt-4). Generic one-word IDs (command, ada, whisper, davinci…) only count next to model context — a model:/engine= key, a provider or SDK name, or a provider/ prefix — so everyday words like the "command" key in MCP configs don't trigger it. Dot-directories (.zcode/, .qwen/…), logs (*.log, *.jsonl), lockfiles and node_modules are skipped, and .gitignore is respected.

Install

pipx install llm-sunset     # or: pip install llm-sunset / uvx llm-sunset

Usage

llm-sunset                          # scan current directory
llm-sunset src/ config/             # scan specific paths
llm-sunset --provider openai        # only OpenAI notices (repeatable)
llm-sunset --fail-within 30         # only fail on models retiring within 30 days
llm-sunset --group-by file          # group output by file instead of by model
llm-sunset --color always           # force color (also honours NO_COLOR / FORCE_COLOR)
llm-sunset --format json            # also: markdown, sarif, github
llm-sunset info gpt-4o-2024-05-13   # look up a model
llm-sunset upcoming --days 90       # every shutdown in the next 90 days
llm-sunset fix --dry-run            # preview rewrites of deprecated model IDs
llm-sunset fix --replace OLD=NEW    # rewrite, picking the replacement yourself
llm-sunset baseline --write .llm-sunset-baseline.json   # record current findings
llm-sunset --baseline .llm-sunset-baseline.json         # only report new ones

Exit code is 1 if any model is already retired or retires within --fail-within days (default 90). Use --no-fail to only report.

Ignoring things: put llm-sunset: ignore in a comment on a line, or llm-sunset: ignore-file anywhere in a file. Use --exclude 'tests/*' for paths. Markdown/RST/TXT files are skipped unless you pass --include-docs. Files ignored by .gitignore are skipped unless you pass --no-gitignore. Editor history/cache directories, backups (*.bak, *.save, …), logs and lockfiles are skipped unless you pass --no-default-excludes; hidden directories are skipped unless you pass --include-hidden.

Config file: permanent settings live in .llm-sunset.toml (or a [tool.llm-sunset] section in pyproject.toml):

exclude = ["tests/*"]
providers = ["openai", "anthropic"]
fail_within = 60

CLI flags override the file; exclude lists are combined.

Auto-fix: llm-sunset fix rewrites deprecated IDs to the provider's suggested replacement (or your --replace OLD=NEW pick). Always preview with --dry-run first — replacements are provider suggestions, not guaranteed drop-in equivalents. Medium-confidence matches (generic words like command) are skipped unless you pass --include-risky.

JSON output is agent-friendly: every finding carries its line_text, a confidence level (high for distinctive IDs, medium for generic words matched via context), the replacements list and a replacement_note.

Azure, Vertex AI and Bedrock publish their own retirement dates for models they resell, which often differ from the original provider's dates. To avoid false alarms these are off by default. Turn them on with --provider azure or --provider all.

GitHub Action

# .github/workflows/llm-sunset.yml
name: llm-sunset
on:
  push:
  pull_request:
  schedule:
    - cron: "0 8 * * 1"   # also re-check weekly: deprecations are announced while your code sits still
jobs:
  check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: Ashveil1/llm-sunset@v0
        with:
          fail-within: 60          # optional
          # provider: "openai anthropic"
          # exclude: "tests/*"
          # gitignore: "false"     # scan git-ignored files too

Findings appear as inline annotations on the PR, along with a summary table on the run page.

pre-commit

repos:
  - repo: https://github.com/Ashveil1/llm-sunset
    rev: v0.3.1
    hooks:
      - id: llm-sunset

GitHub code scanning (SARIF)

      - run: pipx run llm-sunset --format sarif --no-fail > llm-sunset.sarif
      - uses: github/codeql-action/upload-sarif@v3
        with:
          sarif_file: llm-sunset.sarif

Where the data comes from

Deprecation data comes from deprecations.info (source, MIT). It scrapes the official deprecation pages of OpenAI, Anthropic, Google Gemini and Vertex AI, AWS Bedrock, Azure AI Foundry, Cohere, Groq and xAI every day. llm-sunset caches it for 24 hours in ~/.cache/llm-sunset/ and falls back to a bundled snapshot that is refreshed weekly.

If you find a missing or wrong entry, please report it upstream at deprecations-rss. For false positives or negatives in matching, open an issue here.

Support the project

If llm-sunset saved you from a production outage, please consider sponsoring ❤️

Found a false positive or a model it missed? Open an issue — false positives are the #1 thing holding this back, and each report makes the tool better for everyone.

Built with the daily-updated feed from deprecations-rss. If an entry is wrong or missing, the fix belongs upstream.

License

MIT