How AIWire sources its data
Every shutdown date and price on this site is read from the provider's own documentation. Code checks each fact before it is published, and each model page links to the pages its facts came from.
What is checked, and when
All times Eastern (ET)Every day at (06:00 UTC), AIWire fetches the 36 provider pages listed below: each company's deprecation, pricing and model-list pages, read as plain markdown, HTML or JSON. A new copy is stored whenever a page changes.
A page is set aside ("frozen") for the day if it can't be fetched, redirects to another site, comes back in an unexpected format, or yields fewer than 70% of the rows it gave last time. Nothing from a frozen page is changed, and the site keeps showing when it was last checked. Pages that failed are tried again an hour later, at .
Requests identify themselves as aiwire.dev-bot/1.0 with a link to this page.
Sources
36 pagesCloud platforms
Retirement dates from Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry and Groq. They are shown beside a model's own status and never change it.
| Platform | Page | Address | Last checked | State |
|---|---|---|---|---|
| Google Cloud Vertex AI | open models | https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deprecations/open-models | OK | |
| Google Cloud Vertex AI | partner models | https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deprecations/partner-models | OK | |
| Google Cloud Vertex AI | model versions | https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions | OK | |
| Amazon Bedrock | lifecycle legacy | https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle-legacy.md | OK | |
| Amazon Bedrock | model cards | https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards.md | OK | |
| Microsoft Foundry | retirements | https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule?accept=text/markdown | OK | |
| Groq | Deprecations | https://console.groq.com/docs/deprecations.md | OK |
How a fact is checked
A fact is published only when a deterministic check confirms it. Dates and prices are read from the providers' tables by code, not by a language model.
A few cells are written as prose the code can't read, such as some Gemini prices that change on a given date. For those cells only, an AI model reads the cell's text. Its answer is used only if every amount it returns appears in that cell as a dollar figure, and every date it returns appears in the cell. Otherwise the cell waits for a person to review it.
Some changes are held for review even when the check passes:
- a price that moves by more than 50% in either direction;
- a shutdown date that moves earlier;
- an exact shutdown date that becomes a weaker label (see below), or the reverse.
A model name that doesn't match a known API ID is also held. A model that disappears from a provider's page is noted only after it has been missing twice in a row, and that alone never changes its status.
Dates found in changelog prose are always checked by a person before they are published.
Features, limits and aliases are read from the companies' model pages by code. Some are only written as sentences, for example that a model rejects temperature. For those, an AI model suggests the fact together with the exact sentence it came from. A person checks the sentence before anything is published, and the sentence must appear word for word on the page.
Benchmark scores a company reports come from its own launch post. A person either pastes them from the post, or an AI model reads the post and lists the scores it states. An AI-read score is kept only if its number and benchmark name are in a sentence that appears word for word in the post, and it is published only after a person checks it against the post's text. Each score links to the post.
A feature with no published value shows as "not published", never as "no". Each fact shows when AIWire first saw it, or the company's own date where one is given.
Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry and Groq set their own retirement dates. AIWire reads each platform's own lifecycle pages every day and shows those dates beside the company's own, never instead of them. A platform model ID is matched to a model only when the company lists it, or when its form leaves one answer; anything else waits for a person.
Free tiers come from each company's own pricing, model and help pages. Limits are shown as the company writes them. Where the numbers are only in a signed-in console, the free-tier page says where to find them.
Some companies don't publish shutdown dates at all. For those, AIWire watches the model list and flags a model that disappears; the model page says so.
What the shutdown labels mean
Providers don't all promise the same thing, so each shutdown date carries one of these labels.
| Label | Meaning |
|---|---|
| exact | the provider gave a firm date |
| earliest | the earliest possible date; it may run longer |
| not before | the provider guarantees availability until at least this date |
| month only | only the month is known |
| to be announced | deprecated, but no date yet |
| none | no shutdown date announced |
Google's deprecations page describes its shutdown dates as the earliest possible dates, so Gemini dates are labelled "earliest". Anthropic publishes "not sooner than" dates for some current models; those are guarantees, and the models stay active.
Only exact dates get a countdown. A countdown runs to 00:00 UTC on the shutdown date, the earliest moment the model can stop serving requests. That is the evening before on the Eastern clock: 8 p.m. EDT or 7 p.m. EST.
Active, deprecated, retired
Where the provider states a model's status, as Anthropic does, that status is used. Otherwise a model with a shutdown date is deprecated until the date passes, then retired. A model listed as deprecated with no date yet is deprecated. Everything else is active. Retired models keep their pages.
Each breaking or heads-up change is also posted on the wire as a short item written from a template, without AI, linking to the model's page. Those items are in the RSS feed.
Reusing the data
The data AIWire collects (shutdown dates, replacements, prices, features, limits, aliases, cloud-platform dates and free tiers) is licensed under CC BY 4.0. You may reuse it, including commercially, with credit to AIWire and a link to aiwire.dev.
News headlines and the articles they link to belong to their publishers. A summary of this site for coding agents is at /llms.txt.
Benchmark scores from Arena (CC BY 4.0) and Epoch AI (Creative Commons Attribution) are their data, credited on each page and not covered by our licence; scores a company reports are quoted from its own post, which each score links to.
Report a mistake
If a date, price or status here doesn't match the provider's page, email corrections@aiwire.dev. Include the address of the page on this site and what the provider's page says.