Models.dev: An Open‑Source Registry Bringing Transparency to AI Model Pricing and Capabilities

The problem AI developers keep hitting

Every time a team wants to compare a new LLM or multimodal model, they end up digging through scattered documentation, pricing tables, and sometimes even reverse‑engineering API responses. OpenAI, Anthropic, Google, and dozens of smaller providers publish their own specs, but there is no single source that lists model capabilities, token limits, pricing breakdowns, and availability of features such as tool calling or structured output in a comparable format. The result is wasted time, accidental cost overruns, and a reliance on ad‑hoc scripts that quickly become outdated.

Models.dev’s answer

Models.dev is a community‑driven repository that stores model metadata as plain‑text TOML files, one per model, grouped by provider. The repo powers a public JSON API (https://models.dev/api.json) that returns the full catalog, and a lightweight logo service (https://models.dev/logos/{provider}.svg). By exposing the data through a stable endpoint, developers can programmatically fetch the latest pricing, context windows, and feature flags without scraping individual provider sites.

Key design choices that keep the project usable:

• Schema‑driven TOML files – each model must conform to a strict schema (see packages/core/src/schema.ts). This guarantees that the generated JSON is predictable and machine‑readable.
• Extends mechanism – wrapper providers can inherit a canonical definition and only override the fields that differ, reducing duplication and the risk of divergent data.
• GitHub Action validation – every pull request runs a CI job that checks required fields, data types, and value ranges before the changes are merged.
• Open‑source tooling – the front‑end is built with Bun, and the same codebase can be used internally by the maintainers of SST (the company behind the project) via the OPENCODE_MODELS_PATH environment variable.

How contributors add a model

1. Create or locate a provider directory under providers/. If the provider is new, add a provider.toml with its name, npm package, auth environment variables, and optional OpenAI‑compatible endpoint.
2. Add a logo (optional) as an SVG that respects currentColor so it adapts to surrounding UI themes.
3. Define the model in a TOML file named after its ID. The file includes booleans for capabilities (attachment, reasoning, tool_call, structured_output), cost tables (cost.input, cost.output, etc.), limits (limit.context, limit.input, limit.output), and supported modalities (modalities.input, modalities.output).
4. Submit a PR – the CI validates the schema and runs a diff tool (bun run compare:migrations) to ensure only intended changes affect the generated JSON.

Why the community matters

Because AI model offerings evolve rapidly—new versions, price adjustments, and feature toggles appear almost weekly—a static catalog quickly becomes stale. By opening the data to anyone who can read a TOML file, Models.dev turns the maintenance burden into a shared responsibility. Contributors can quickly add emerging providers (e.g., a new Chinese LLM vendor) or update cost changes that would otherwise surprise developers during budgeting.

Potential impact on the ecosystem

• Cost transparency – developers can script cost‑estimation tools that pull the latest per‑million‑token rates, reducing surprise bills.
• Feature discovery – a simple lookup tells whether a model supports tool calling or structured output, helping teams decide which provider fits a given use case without trial‑and‑error.
• Standardisation pressure – as more tooling starts to rely on the Models.dev API, providers may feel indirect pressure to keep their public specs consistent, knowing that discrepancies will be visible to a broad audience.

Funding and sustainability

The repository does not list any external investors or a formal funding round. It is maintained by the creators of SST, who use the data internally for their own products. The project’s sustainability currently hinges on community contributions and the goodwill of its maintainers. No monetisation strategy is announced, which aligns with the open‑source ethos but also leaves the long‑term upkeep dependent on volunteer effort.

Where to start using it

• API quick test: curl https://models.dev/api.json | jq '.providers.openai.models["gpt-4o"]'
• Fetch a logo: curl https://models.dev/logos/openai.svg > openai.svg
• Add a model: fork the repo, add the TOML definition following the schema, and open a pull request.

Closing thoughts

Models.dev is a pragmatic response to the fragmentation that hampers AI development today. By providing a single, version‑controlled source of truth for model specs, it saves developers time and money while encouraging a more transparent market. Its success will depend on the community’s willingness to keep the data fresh and on the maintainers’ ability to keep the validation pipeline reliable. If those pieces stay in place, the catalog could become the go‑to reference for anyone building on top of the rapidly expanding AI model zoo.