AI model retirement checker: paste code, get dates

Check your model IDs

Paste anything that names a model: a source file, a .env file, a YAML config or a plain list of IDs, then click Check models. To see a result first, click Use an example.

What it checks and how to read the results

The checker reads every word in your paste that could be a model ID, up to 200 of them, and looks each one up on our retirement calendar. It matches the ID exactly as the provider writes it, ignoring case, so a hit is a date a provider published, never a guess. It also tries the other spellings the same ID takes in code:

• Without a Region prefix: us.anthropic.claude-sonnet-4-5-20250929-v1:0 is also looked up as anthropic.claude-sonnet-4-5-20250929-v1:0.
• Without a path: models/gemini-2.5-flash and openai/gpt-4o-2024-08-06 are also looked up as the part after the last slash.
• Without a version after the @: claude-sonnet-4-5@20250929 is also looked up as claude-sonnet-4-5.

A full Bedrock inference profile ARN isn't matched as a whole, so paste the part after its last slash instead.

Reading a result

• Model ID is the ID as you wrote it, with the line of your paste it's on.
• Where is the platform the date applies to, such as the OpenAI API, Microsoft Foundry, Amazon Bedrock or Vertex AI. One ID gets a row for each platform that lists it, soonest first.
• Date says when it shuts down and how many days are left. "No sooner than" means the provider gave an earliest date, not a fixed one. When the provider names a replacement, the row says what to replace it with.

No date on record

Words shaped like a model ID that have no date on the calendar are listed after No date on record. That isn't the same as safe. It means no provider has announced a shutdown date for that exact ID that the calendar has, and it's worth a look by hand when it's one of these:

• A model that shut down a while ago. The calendar lists dates from about 90 days ago onward, so an ID that stopped working before then lands here.
• A name you chose. On Azure, your code sends a deployment name, and the deployment sets the model and its version (Microsoft Learn). Look up the deployment's model version in Azure.

The page downloads the model dates from our public calendar and does the matching in your browser. What you paste isn't sent anywhere.

Why one model has different dates on each platform

A shutdown date belongs to a model on a platform. The model maker sets the date for its own API, and each cloud sets its own for the copy it serves. The same model can stop working on one platform months before another, which is why the checker shows a row per platform.

• Claude on Microsoft Foundry follows the Claude API. Claude's deprecations page says its dates apply there too (Claude docs), and Foundry lists claude-sonnet-4-5 retiring on November 30, 2026 (Foundry schedule).
• Bedrock's lifecycle can differ by AWS Region. Read the modelLifecycle field from GetFoundationModel in each Region you call (AWS).
• Vertex AI became part of Gemini Enterprise Agent Platform on April 22, 2026 (Google Cloud release notes). Clients and hosts still carry the old names, such as AnthropicVertex and aiplatform.googleapis.com (Claude docs).

Where model IDs hide in a repo

The ID in your main client call is easy to find. These are the places that get missed:

• .env files, CI secrets and your host's environment settings. git grep only reads tracked files, so check these by hand.
• Config files for gateways, agents and evals, which often set a different model per task.
• Infrastructure code, such as Terraform or CDK, and IAM policies that scope Bedrock access to one model ARN.
• Tests, fixtures, recorded responses and notebooks.
• Azure deployments, where the model version lives on the deployment, not in your code.

To collect candidates, run this from the repository root and paste the output into the checker. It catches the common families, so add the ones you use, such as mistral or llama:

git grep -h -i -E \
  'gpt-|claude|gemini|embedding|"o[0-9]' \
  | sort -u

The -h leaves out file names, so search again for any ID the checker flags, or scan the repository to get every file and line.

Use it from an agent or CI

The checker runs on our public calendar API, which is free and needs no key (API docs). This lists every model with a shutdown date in the next 90 days:

curl -s https://api.since.dev/v1/calendar \
  -G -d kind=model -d within=90d -d limit=500

kind takes model, api or runtime. provider narrows it to one provider, such as openai. within keeps dates from today through that many days ahead. limit goes up to 500 a page, and nextCursor gets the next one.

In CI, this prints every line in your repo that names one of those models:

curl -s https://api.since.dev/v1/calendar \
  -G -d kind=model -d within=90d -d limit=500 \
  | jq -r '.entries[].modelId // empty' | sort -u \
  | while read -r id; do
      git grep -n -w -F "$id" || true
    done

• For a calendar app, subscribe to https://api.since.dev/v1/calendar.ics?kind=model to get each shutdown as an event. /v1/calendar.rss is the same list as a feed. Both take the same filters, except limit and cursor.
• For a coding agent, our MCP server has a check_code tool. Give it a model ID and it answers with the model's lifecycle and every change to it. It needs a free account and uses credits.

Where Since.dev fits

Since.dev tracks model retirements on each platform that serves them, from the publishers' own pages. This checker covers what you paste. For a repository you scan or connect, Since.dev finds every file and line that names a retiring model and checks it against the platform your code calls, so a Bedrock date only flags code that calls Bedrock.

• Scan any public GitHub repository, no sign-in needed.
• See every date on the retirement calendar, or subscribe to it as above.
• On Team and up, it opens a fix pull request that swaps the model, ready for your review. Nothing merges or deploys.

Questions

Is the model retirement checker free?

Yes. It needs no account, and the dates come from our public calendar API, which is free and needs no key.

Does my code leave my browser?

No. The page downloads the model dates from the calendar and matches them in your browser. What you paste isn't sent anywhere.

What does No date on record mean?

The word looks like a model ID, but the calendar has no shutdown date for it. Usually no provider has announced one yet. It can also be a model that shut down more than about 90 days ago, or a name you chose, such as an Azure deployment name. It doesn't mean the model is safe.

Why does the same model show different dates?

Each platform sets its own. gpt-4o-2024-08-06 shuts down on the OpenAI API on January 14, 2027 (OpenAI) and on Azure on April 14, 2027 (Foundry schedule). Use the row for the platform your code calls.

Which platforms does it cover?

Every platform the calendar has dates for: the model makers' own APIs, such as OpenAI, Claude, Gemini and Mistral, plus Microsoft Foundry, Amazon Bedrock and Vertex AI.

Related guides