Models
Two API standards that nobody standardized. OpenAI published an API and the industry turned it into a wire format; Anthropic's Messages API became the second dialect a growing share of model hosts speak. Both are profiled here as adopters actually implement them, graded into tiers by measured adoption, with a registry of who implements what and how strongly each claim is evidenced.
| Cohort | reimplementation — independent servers written against someone else's published surface |
|---|---|
| Status | Published — 582 operations graded across two profiles, 195 adopters recorded |
| Repository | https://github.com/api-commons/models |
| Profiles | 2 in this domain |
The OpenAI Interface
"OpenAI-compatible" appears in the documentation of 211 providers. Measured against what they publish, it means two operations.
| Identifier | openai-v1 |
|---|---|
| Originator | OpenAI |
| Source of truth |
openai/openai-openapi
MIT
Pinned at 38170fdddbb6a1813eae6c6587ee17cf2987185b
OpenAPI 3.1.0, info.version 2.3.0 — 215 paths, 338 operations. |
| How it was measured | 120 of the 211 claimants publish an OpenAPI harvested byte-identical from their own site, and those are the denominator. Grading is per method and path, not per path: a path-level count credits GET and POST equally, which put GET /chat/completions at the adoption rate of POST and DELETE /models/{id} at four times its real share. |
| Tier | What it covers |
|---|---|
| Core 2 operations |
POST /chat/completions at 87.5% and GET /models at 71.7%. Nothing else clears a third, so the boundary sits in a gap rather than at a threshold somebody picked. |
| Extended 19 operations |
5% to 50% — embeddings (32.5%), the legacy completions endpoint (30.0%), image generation (23.3%), the Responses API (21.7%), audio, files and batches. |
| Vendor 317 operations |
Under 5% — Assistants, ChatKit, Realtime, evals and the whole organization administration surface. Effectively OpenAI-only in practice. |
175 implementations recorded · as of 2026-09-13
| Evidence | Count | Meaning |
|---|---|---|
declared | 101 | Publishes their own OpenAPI, harvested byte-identical, declaring operations in this profile. 15 reach core and 52 reach extended; 34 declare part of the profile but not the whole core tier. |
prose | 74 | States compatibility in their own words and publishes no readable spec. |
tested | 0 | Not attempted. Unlike open data and blockchain, running the conformance flow here needs credentials and costs money. |
The full adopter registry · the adopter matrix — every entry carries the URL it was read from and the date it was read.
Artifacts
-
OpenAPI
— OpenAPI
Core and extended surface, self-contained, x-tier on every operation. -
Profile
— Profile
All 338 operations with their tier and the evidence for that grading. -
Spectral Ruleset
— SpectralRules
Lint any provider's OpenAPI and see which tier it reaches. -
Arazzo
— Arazzo
List models, then use one — the flow that proves a core claim. -
MCP Tools
— MCP
MCP tool definitions for the core tier.
The Anthropic Messages Dialect
The second language a growing share of model hosts speak. Measuring it required solving a problem the first profile did not have.
| Identifier | anthropic-messages-v1 |
|---|---|
| Originator | Anthropic |
| Source of truth |
anthropics/anthropic-sdk-python — scripts/mock-spec.json.gz
MIT
Pinned at eb21a4352015686c30f5759e8c2f02d70f5371e2
OpenAPI 3.1.0 — 165 paths, 244 operations, no declared info.version. Anthropic bundled this spec into their own repository on 2026-09-03, in the same commit that removed the Stainless CDN URL from .stats.yml. The in-repo file is first-party and pinnable by commit, which the CDN copy was not. |
| How it was measured | 29 of Anthropic's operations share a path shape with the OpenAI interface — /models, /files, /skills, /vaults, /agents. Graded against everyone who publishes a spec, GET /models came out at 86.3% here, and that number was measuring OpenAI. The cohort is therefore the 29 providers declaring at least one operation UNIQUE to Anthropic, and every shared operation is flagged rather than silently counted. |
| Tier | What it covers |
|---|---|
| Core 2 operations |
POST /v1/messages at 69.0% — the only operation that identifies the dialect — and GET /v1/models at 72.4%, which shares its shape with the OpenAI interface and so is evidence of a model listing rather than of this dialect. |
| Extended 11 operations |
count_tokens (13.8%), a single model read, files, skills and the legacy /v1/complete. |
| Vendor 231 operations |
Under 5% — agents, deployments, sessions, batches and the admin surface, plus the preview operations the upstream document expresses as a query string on the path key. |
20 implementations recorded · as of 2026-09-13
| Evidence | Count | Meaning |
|---|---|---|
declared | 20 | Every entry here publishes a spec — by construction, since a prose claim cannot be cross-checked when the paths are shared with another dialect. 6 reach core, 3 reach extended, 11 declare part of the profile but not the whole core tier. |
prose | 0 | Excluded by design for this profile. See the measurement note above. |
tested | 0 | Not attempted — needs credentials. |
The full adopter registry · the adopter matrix — every entry carries the URL it was read from and the date it was read.
Artifacts
-
OpenAPI
— OpenAPI
Core and extended surface, self-contained, x-tier on every operation. -
Profile
— Profile
All 244 operations with their tier, evidence, and shared-shape flags. -
Spectral Ruleset
— SpectralRules
Checks the two things that break a shim written against the wrong dialect — max_tokens required, and content blocks rather than choices. -
Arazzo
— Arazzo
List models, then send the smallest message with the first one. -
MCP Tools
— MCP
MCP tool definitions for the core tier.
Licensing
Both profiles are independently authored from openly licensed machine-readable sources; neither republishes a vendor's document. Both upstreams are MIT and carry their attribution. API Commons artifacts are CC BY-NC-SA 4.0 and code is Apache-2.0. "OpenAI" and "Anthropic" appear as factual references to the interfaces described, and nothing here implies endorsement.
Tags
ModelsLLMInferenceArtificial Intelligence
Further reading
- The adopters building block — The schema behind both registries, and the evidence grades it enforces.
- The adopter matrix — OpenAI — Operation by adopter, across everyone who publishes a spec.