Scope
AIDb (The AI Database) is a structured directory of AI models: large language models, multimodal and reasoning models, image, video, speech and music generators, embeddings, vision systems, scientific models and the consumer products built on them. Each model is one row with the same 14 groups of fields, one static dossier page and one entry in the machine-readable dataset.
Sources
Every fact in AIDb is taken from the maker’s own publications: model cards, technical reports and papers, developer documentation and API model lists, product pages, official blogs and changelogs, pricing pages, and regulatory filings (SEC, HKEX) or company-issued press releases. The page a row rests on is recorded in its Primary Source URL, and the dossier links to it.
Not used as sources: third-party leaderboards and rankings, benchmark aggregators, news coverage, Wikipedia, analyst estimates, and a maker’s claims about other companies’ models. A description may quote the maker’s own benchmark numbers and compare a model with the maker’s earlier models, but not with rivals.
When no official page can be reached, the cell stays blank or keeps its last verified value and the gap is noted in the changelog. A blank cell beats a guess.
What is in a row
| Model_ID | Stable slug used in every URL: theaidatabase.com/m/{id}.html and ?model={id}. |
|---|---|
| Model Name, Vendor | As the maker names them; a vendor's own capitalisation is kept. |
| Status, Release Date, Release Year | Release information as the maker published it; the Release Source URL column names where the date came from. Retirements are recorded in the description and pricing cells (see Lifecycle). |
| Model Type | Family such as LLM, Multimodal LLM, Image Generation, Video Generation, Vision/CV, Code LLM, Reasoning LLM, Speech Synthesis, Music Generation, Embeddings. |
| Jurisdiction | Where the maker is based; USA, China, USA / UK, France and UK are the largest groups. |
| CEO, Key Staffers, Leadership / Governance Context | Named leadership as published by the maker, and a short note on the CEO's documented public political activity (donations, statements, op-eds) or the absence of any, drawn from public records. |
| Capability flags | Image Gen, Video Gen, PDF Gen, Spreadsheet Gen, App Available, Voice Chat, Music Gen, Dictation, Audio Gen: Yes/No from the maker's documentation. |
| Subscription Cost, Free Tier Available | Current pricing in the maker's own units, with the date of any retirement when a model has left an API. |
| Open Source, Deployment Type | Open weights and their licence, or proprietary; cloud only, self-hosted, on-device. |
| Special Capabilities / Description | Up to 500 characters: what the maker documents the model does, with its own numbers. No rival comparisons, leaderboard placings or superlatives. |
| Known Safety Issues | Documented incidents and maker-acknowledged limitations, with sources. |
| AIDb Rating (1-10), Score Basis | The editorial rating and the four component scores behind it (see Ratings). |
| Data Completeness, Evidence Confidence, Last Verified | How much of the row is filled, how strong its sourcing is (High, Medium or Low), and when it was last checked. |
| Primary Source URL, Release Source URL, Pricing Source URL, Docs / API URL, Benchmark Source URL | The pages each fact rests on. |
Verification
Each row carries a Last Verified date, a Data Completeness percentage and an Evidence Confidence grade (High, Medium or Low) that reflects how directly the primary source supports the row. Dossier pages compute a 26-point A-to-Z completeness score from the fields they can show.
Maintenance runs in numbered accuracy passes. A pass checks a set of rows against their makers’ current pages, rewrites what has changed, records every cell change with its old value, new value, reason and source URL, and republishes the site. The published files are then fetched back from the live server and compared byte for byte with the build before the pass is closed.
Descriptions are capped at 500 characters and written from the source page in plain language: what the model is, what the maker says it does, how it is offered, and what has happened to it since (price changes, deprecations, successors).
Ratings
The AIDb Rating (1 to 10) is an editorial score built from four components that each dossier lists in its Score Basis line:
- Capability: what the model does as documented by its maker, across modalities and tasks.
- Currency: how recent and actively maintained it is; retired and deprecated models score low here.
- Access: how easy it is to use or obtain: free tier, open weights, API and app availability.
- Significance: its weight in the field, judged from its documented adoption, lineage and successors.
Ratings are AIDb’s judgment, not a benchmark result, and are not taken from any leaderboard. Every published score matches the formula exactly; rescoring is recorded in the changelog like any other change.
Changelog
Every dataset release is listed on the changelog with what was added, removed, corrected or rescored, and the maker pages checked. Behind it sits a cell-level log: for each change, the model, the field, the old text, the new text and a note naming the source and the date it was checked. As of October 17, 2026 that log holds 8,195 verified changes across 199 releases.
Lifecycle
AIDb keeps rows for models that have been retired, deprecated or scheduled for shutdown rather than deleting them, and records the maker’s notice in the description or pricing cell. The Retirements & Deprecations page collects those entries (162 at this update, with shutdown dates still ahead for 31), each linked to its dossier and to the maker’s own page; retiring.json is the machine-readable copy.
Data access
- models.json: the full dataset, one object per model with
id,name,vendor,type,jurisdiction,release_year,ceo,key_staffers,description,capabilities,open_source,free_tier,stock_status,subscription_cost,deployment,x_handle,aidb_rating,score_basis,data_completeness,last_verifiedandprimary_source, plus agenerateddate at the top. - retiring.json: retirements and deprecations with ISO dates and sources.
- Dossiers at
/m/{id}.html: static pages with schema.org JSON-LD, the A-to-Z profile, the primary source and a citation line. - sitemap.xml (every dossier) and llms.txt (a plain-text guide for AI agents).
Files are regenerated together on every release, so the generated date in models.json, the dossiers’ Last Verified dates and the changelog agree.
Citing AIDb
For the dataset as a whole: TheAIDatabase.com (AIDb), models.json, generated YYYY-MM-DD, https://theaidatabase.com/models.json
Each dossier carries its own ready-made citation line with the row’s Last Verified date.
Corrections
If a cell disagrees with a maker’s page, send the model, the field and the official URL to @youraidatabase. Corrections are checked against the source, applied in the next pass and recorded in the changelog with the old and new values.