Apps call AI models by name, and those names don’t sit still. Providers retire models, rename them, shrink context windows, add retirement dates and change token prices. Each change is announced somewhere different: a changelog, a deprecations page, an email that went to someone who left. Most teams find out when a request fails in production. Databases have the same problem in slower motion. Someone runs one manual ALTER on production, outside the migrations, and from then on nobody’s sure what the schema really is.
RadarModel watches both. Two products share one engine. RadarModel AI reads the model lists that AI providers publish and tells you when a model you use is going away, getting smaller or getting pricier. RadarModel Schema reads your Postgres, MySQL, MariaDB or SQLite schema on a schedule and tells you when it drifted.
The Pitch in Four Numbers
- 14 ways a provider’s answer was broken in the tests, each run three times, and none reported as a removed model.
- 2,000 tables scanned on Postgres in 0.41 seconds, with a dropped table found on the next scan.
- 9 AI providers read, plus any OpenAI-compatible endpoint.
- 0 cost at the providers. It reads model lists and never calls a model.
RadarModel AI
It reads OpenAI, Anthropic, Google, OpenRouter, Mistral, Groq, xAI, Azure OpenAI and Amazon Bedrock. Each run takes a snapshot of every list, compares it with the last one and sorts the changes. Breaking: a model you watch was removed or retired, or its context or output limit was cut. Warning: a price rise, a retirement date or a deprecation. Everything else is info. Only models you list can raise warnings, so the noise stays low.
It’s careful about false alarms. A failed scan, an outage, a cut-off or empty response, is never read as a removed model, and a removal only counts as breaking after a second scan confirms it. Alerts go to the terminal, a webhook, Slack or email. check exits 1 on a breaking change, so it works as a CI gate that stops a deploy when a model your app calls is about to vanish. API keys come from environment variables and go only into request headers to the provider’s own host, never into files, URLs, errors or alerts.
RadarModel Schema
It reads live schemas: tables, views, columns, indexes, constraints and triggers. No migration tool needed. The first scan sets a baseline and later scans report what changed. Breaking: a dropped table or column, a changed type, a new NOT NULL. Warning: a dropped index or a changed default. The diff command compares two databases, say staging and production, before a release. It reads only the catalog: Postgres and MySQL in read-only transactions, SQLite opened read-only. A lost connection or a migration running mid-scan counts as a failed scan, never as drift.
Who It’s For
- Teams running AI features in production on one or several providers.
- Anyone routing through OpenRouter who wants warning of price rises and retirements on the models they use.
- Database owners who suspect hand fixes on production.
- Release managers who want staging and production schemas compared before a deploy.
Limits
What it can see depends on the provider. OpenAI’s list has model IDs and little else, so RadarModel can tell you a model’s gone but not that its limits changed. Anthropic, Google and OpenRouter publish more. Most providers need an API key, and Bedrock needs AWS credentials. On MySQL and MariaDB it needs SELECT, SHOW VIEW and TRIGGER rights. It doesn’t call models, change databases or monitor provider outages.
The Four-Question Screening
Every project featured here gets the same four questions, the way you’d screen a startup idea.
1. Where is the money in it?
RadarModel is free under Apache 2.0, with no paid tier. AI model churn is a growing operational headache, and companies spending heavily on AI APIs will pay to avoid a surprise outage or a price jump. A hosted version that watches all providers for you, with a history of price and limit changes and alerts routed to the right team, is the clear product. Schema drift detection is an established need in database operations and pairs well with it for platform teams.
2. What real problem does it solve?
Changes underneath your app that nobody told you about. A retired model breaks a feature. A price rise blows a budget. A manual ALTER breaks the next migration. All of them are cheap to catch early and expensive to discover in production.
3. The technology in plain words
Think of someone who reads the supplier catalogues every morning and checks them against yesterday’s. If a part you use has been discontinued or gone up in price, they put a note on your desk. If someone has quietly rearranged the warehouse shelves overnight, they tell you that too.
4. Moat and differentiation
Provider status pages and changelogs exist, scattered across nine sites. Schema tools usually come bundled with a migration framework. RadarModel reads the source of truth directly, the published model lists and the live catalog, and is conservative about alerts, which is what keeps people from muting it. The moat that can grow is provider coverage, the history of changes it builds up, and its place as a CI gate, which is sticky.
Why It Matters
AI providers move fast, and apps are built on their model names. A cheap, quiet watcher that speaks up only when a model you use is about to break saves a lot of 3 a.m. surprises.
Follow the Project
Read why AI models drift under your app and how schema drift starts with one manual ALTER, start with the quick start, see the CI gate case study, or get the code from the RadarModel repository on GitHub. RadarModel is listed with the other open-source projects on SoftDevStudio.com, and more featured projects are in the Projects section.