Overview
TensorTrace brings AI models, independent evaluations and inference offerings into one library. Use the website to discover a model; use the public API to build your own search, comparison or research tools.
Start here
The public API base URL is https://tensortrace.net/api/v1. Public catalog reads do not require a TensorTrace API key. They read locally collected, prepared data; a request does not contact inference providers or run a model. News and research read first-party ingestion tables in the same TensorTrace database.
curl --get 'https://tensortrace.net/api/v1/models' \
--data-urlencode 'q=gemini' \
--data-urlencode 'limit=1'
Use the returned canonical_model_id to retrieve the same model across prices, benchmarks and evidence. Models explains searching and detail sections; Prices explains comparable rates and complete billing conditions.
Choose the right object
| Object | What you can build |
|---|---|
| News | Recent AI topics and related model/developer coverage |
| Models | Search, model cards, specifications and capability filters |
| Model Developers | Catalogs organized by the labs and authors that train models |
| Inference Providers | Serving-platform catalogs and official service status |
| Benchmarks and Evaluations | Rankings, charts and original evaluation observations |
| Evidence and Sources | Auditable links, freshness, receipts and original documents |
| Users, Favorites and Comments | Reserved identity/community contracts; currently unavailable |
The left directory includes deployment variants, operational metrics, platform availability, safety and shared Data Types.
Identity and provenance
- Canonical model ID: TensorTrace’s stable identity for a model. It can differ from the developer’s official API identifier.
- Official model ID:
official_model_id, where evidenced, names the developer’s model using only the final/segment. Complete observed identifiers remain inofficial_identity.valueand source profiles. Unknown identifiers remain unknown; usecanonical_model_idfor URLs and global identity. - Source model ID: a collecting platform’s own identifier. An inference route, quantization, moving alias or deployment version is not automatically a new canonical model.
- Source ID and receipt: identify who supplied a fact and when it was observed. The destination of an external link can differ from its collecting source.
Preserve source IDs, conditions and observation dates in your application. Benchmark scores from different boards, versions, units or workloads are not interchangeable. Historical or archived datasets retain their snapshot dates and links to current sources.
Pagination and freshness
Model pages default to 48 results and accept at most 100. Other objects have their own limits, documented on each endpoint. Offset pagination uses limit and offset; user timelines use opaque cursors. A total_count is the number of matching rows, not the number on the current page.
Workers acquire sources serially and incrementally prepare query data. Each request reads committed local data. There is no global snapshot token: different page requests may see updates between requests. Dates and receipt identifiers tell you the scope of each observation. An unprepared catalog returns 503 rather than pretending to be empty.
A missing rate or specification means unknown, not zero or unsupported. Prices retain currency, unit, platform and conditions. Basic token price selection compares Input first, Output second, and prefers the developer when both rates tie; cached, batch or media rates remain separate details.
Errors and access
Errors normally include an error string. Read the endpoint’s response table for its declared status codes. 400 indicates invalid input, 404 an unknown resource, and 503 unavailable configuration or prepared data. Account mutations require a browser session, the session-bound X-CSRF-Token and an exact Origin; see Authentication.
Account/community contracts are retained for compatibility, but their old Rust HTTP implementation is retired. The current JavaScript runtime reports providers disabled and returns private/no-store 503 for other identity APIs. Restoring them requires a separate implementation and acceptance, not a configuration toggle. Administration is CLI-only; public admin paths remain 404.
Machine-readable contract
Download OpenAPI 3.1 for client generation and validation. Data Types lists request and response fields. Request identifiers in examples are illustrative and must be replaced with IDs obtained from your own catalog; sample responses demonstrate the schema and are not live measurements.
On this page
- Read catalog statistics ·
GET /overview - Resolve a social sharing image ·
GET /social-card
Read catalog statistics
GET /api/v1/overview
Current counts and latest preparation timestamps from incrementally maintained local read models. This endpoint does not pin a catalog version.
Access: public read; no TensorTrace API key.
Parameters
This operation has no path or query parameters.
Request example
curl 'https://tensortrace.net/api/v1/overview'
Responses
| Status | Meaning | Body type |
|---|---|---|
200 | Incrementally prepared local data | CatalogOverview |
400 | Invalid query | ErrorResponse |
404 | Not present in the prepared catalog | ErrorResponse |
503 | Catalog preparation has not completed yet | ErrorResponse |
Response example · 200
Illustrative values; schema-valid and not a live catalog measurement.
{
"total_models": 1,
"free_models": 1,
"paid_models": 1,
"multimodal_models": 1,
"reasoning_models": 1,
"tool_calling_models": 1,
"model_count": 1,
"family_count": 1,
"distribution_count": 1
}
| Field | Type | Required | Meaning / constraints |
|---|---|---|---|
total_models | integer | Yes | |
free_models | integer | Yes | |
paid_models | integer | Yes | |
multimodal_models | integer | Yes | |
reasoning_models | integer | Yes | |
tool_calling_models | integer | Yes | |
best_context_window | integer or null | No | |
average_capability_score | number or null | No | |
latest_refresh_at | string · date-time or null | No | |
model_count | integer | Yes | Minimum: 0 |
family_count | integer | Yes | Minimum: 0 |
distribution_count | integer | Yes | Minimum: 0 |
Resolve a social sharing image
GET /api/v1/social-card
Choose a canonical model ID or a public section path. With S3 configured, current content and bundled template/font determine an immutable image key. Existing objects are reused across restarts; missing objects are rendered and uploaded before a 302 to the public S3 endpoint. Redirect responses are never cached by browsers or CDNs. Without storage configured, the development fallback serves a PNG directly. X and Facebook use the same 1200 by 630 image. Legacy /social/*.png routes resolve through the same service.
Access: public read; no TensorTrace API key.
Parameters
| Name | Location | Required | Type / limits | Meaning |
|---|---|---|---|---|
id | query | No | string | Canonical model identifier or resolvable model alias; mutually exclusive with page. |
page | query | No | string Default: "/" | Public section route such as /, /models or /benchmarks; defaults to the homepage when id is absent. |
Request example
curl 'https://tensortrace.net/api/v1/social-card'
Responses
| Status | Meaning | Body type |
|---|---|---|
200 | Storage-disabled development fallback: PNG image bytes. | string · binary |
302 | Current image is available in public S3 storage. Follow Location to read PNG bytes directly from S3. | No declared JSON body |
304 | Storage-disabled fallback conditional image hit. | No declared JSON body |
307 | Model alias resolves temporarily to the canonical image API URL. | No declared JSON body |
400 | Choose id or page, not both. | No declared JSON body |
404 | Unknown model or section. | No declared JSON body |
503 | Storage could not confirm or upload the image; no redirect to a missing object. | ErrorResponse |
Response example · 200
The response contains original bytes or an image. Content types: image/png. Preserve the declared encoding and checksum; do not parse it as JSON.