Last Orders (working name)

Find retiring AI model IDs in your code before they break. One command scans your repo for OpenAI (and, as a preview, Anthropic) model IDs that are being shut down. For each one it tells you the shutdown date, the replacement the provider recommends, and the code changes the new model needs.

Heads-up: OpenAI removes GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini from the API on 14 January 2027 and recommends gpt-6-luna. In our tests, gpt-6-luna rejects max_tokens, and it rejects tool calls on Chat Completions unless reasoning_effort is "none". So changing the model name isn't enough. The scan flags both.

The scan is free. You don't need an account, a licence or an API key, and your code never leaves your machine.

Install

pipx install lastorders        # or: pip install lastorders

You need Python 3.10 or later.

Scan

lastorders scan .              # scan the current repo
lastorders scan path/to/repo --out scan-results
lastorders scan . --offline    # no network at all (uses the bundled migration pack)

You get scan.md, a readable report, and scan.json, the same data for tools. They go in lastorders-out/, or wherever --out points. Each finding includes:

• File, line and model ID, from code (Python, JS, TS), .env files, YAML, JSON, TOML and docs
• Shutdown date and recommended replacement, from the provider's published deprecations
• Risk level. High means the date is close or has passed, or the new model needs a code change your code doesn't have yet, such as max_tokens becoming max_completion_tokens, reasoning_effort for tools and temperature, or an unsupported endpoint
• Fine-tuned models (ft:...), listed separately. They shut down with their base model and can't be swapped automatically.
• Models chosen by environment variable (os.getenv("OPENAI_MODEL")), so you can check the values you have deployed
• SDKs detected: the official openai and anthropic clients, LangChain and plain HTTP

Useful flags:

Fix (patch preview)

lastorders fix .                 # writes lastorders-out/fix.diff + fix-plan.md; changes nothing on disk
lastorders fix . --verify        # also checks the patch on a scratch copy (see below)
git apply lastorders-out/fix.diff

fix swaps the retiring model IDs and then makes the code changes the new model needs at each API call:

• Renames max_tokens, to max_completion_tokens on Chat Completions or max_output_tokens on Responses.
• Adds reasoning_effort: "none" where gpt-6-luna needs it: Chat Completions calls with tools, and calls that set temperature, top_p or logprobs. On Responses it adds reasoning.effort.
• Flags anything it can't decide safely in fix-plan.md: legacy functions calls, Assistants and legacy Completions calls (these stay on their current model in the patch), fine-tunes and fine-tuning jobs, LangChain wrappers, calls that pass **kwargs or a spread object, conflicting values, and calls whose model isn't visible in the code (use --assume-model).

Every change comes from the migration pack's rules. Python is edited with libcst, so your formatting and comments are kept. JS and TS are edited with a built-in parser. Rules the pack marks inferred are only suggested unless you pass --include-inferred.

Old models reject reasoning_effort. We tested gpt-4o and gpt-4.1 against the live API, and both return HTTP 400. Any fix that adds it has to ship together with the model switch. When the model comes from an environment variable, the plan marks the fix "ship with the model switch". --no-model-swap applies only the changes that are safe on the old model.

--verify applies the patch to a temporary copy of your repo, then:

1. checks the changed files still parse;
2. re-runs the fixer and checks it changes nothing more;
3. replays each fixed call twice: once as it was (old model, old arguments) and once as patched (new model, new arguments).

Without --live, the replay uses simulated responses. With --live and your own OPENAI_API_KEY (licence), the calls go straight to OpenAI. A control run also sends the unpatched arguments to the new model, to show each fix was needed. --evals evals/regression.yaml reruns your own suite with the fixed arguments.

GitHub Action

Add .github/workflows/model-scan.yml:

name: Model retirement scan
on:
  pull_request:
  schedule: [{ cron: "0 7 * * 1" }]   # weekly, so new deprecations show up without code changes
permissions:
  contents: read
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: burntandtoast/lastorders@v0     # available once the repo is published
        with:
          fail-on-high: "true"

The Action posts a summary table to the job page, marks findings inline on the PR and uploads scan.json as an artifact. Inputs:

Outputs: total, high and report-dir.

Migration packs: where the data comes from

What the scanner knows lives in migration packs, versioned JSON files with one pack per provider. Each pack holds the model IDs and aliases, shutdown dates, replacements, call-site patterns, and parameter and endpoint rules. Every fact links to the provider's documentation and the date we checked it. Rules we confirmed against the live API are marked live-tested, and anything we assumed is marked inferred.

• OpenAI: full pack
• Anthropic: preview (stub). It covers Claude API retirement dates from Anthropic's deprecations page. Amazon Bedrock and Google Cloud dates are not covered yet.
• Gemini, Azure OpenAI and Bedrock: planned

The tool checks for a newer pack once a day and falls back to the copy bundled with the package. lastorders packs list shows which version and source you're using. The format is documented in docs/PACKS.md, with a JSON Schema in migration_kit/pack_data/pack.schema.json.

Privacy and network use

• The scan runs entirely on your machine. Your code and prompts are never uploaded.
• The tool has no telemetry, analytics or tracking. (Our website counts visits anonymously with Cloudflare Web Analytics: no cookies, no ads, no tracking across sites.)
• The tool only ever contacts us to download the latest migration pack (a plain GET of a static JSON file, with no identifiers) and, for paid features, to check a licence key. That check sends only the key and the repo name (github.com/owner/repo), and the signed answer is cached for a week, with a 14-day offline grace period after that.
• --offline (or LASTORDERS_OFFLINE=1) turns off all network access and uses the bundled pack.
• The paid eval command calls the AI provider directly with your own API key. Nothing goes through us.

Paid features

One licence covers one GitHub repository, monorepos included. It's bound to the repo the first time you use it. Buy one with lastorders licence buy (it prints a Stripe Checkout link for this repo), then pass the key with LASTORDERS_LICENCE (a repository secret in Actions) or --licence. lastorders licence status shows whether it's verified. Paid commands work --offline from the cached check. Repos without an origin remote can set LASTORDERS_REPO=github.com/owner/repo. The licence server lives in server/; see docs/LICENCE-SERVER.md.

init-evals reads the call sites the scan found, including prompts, JSON schemas and tool definitions where they can be read from the code, and writes a starter evals/regression.yaml. You then edit it into 10 to 20 real examples. eval runs them against the old and new model with your own key.

Limitations

• Detection uses pattern matching. Model names built at runtime or stored in a database won't be found; env-var lookups are listed so you can check them.
• A pack is only as current as its updated date. Always check the provider's documentation before acting.
• The scan suggests changes; it doesn't decide them for you. Review every patch.

Made by Burnt & Toast Studios. Not affiliated with OpenAI or Anthropic.