Salesforce and OpenAI are the two providers the Artificial Intelligence tag page ranks first, at 91.8 and 86.0 on the Kin Score, both exemplar, 5.8 points apart. It is the biggest tag in the catalog, 3,655 providers and 580 APIs, and one of the weakest: 3,642 scored members average 21.6, barely above the 20.8 for everyone else, so both of these sit more than 60 points above their own category. The Kin Score measures the public API surface, not the product and not the company. That clause matters here more than usual, because on the six facets that make up that surface OpenAI does not lose a single one, and Salesforce still ranks first.
The business. They reach the tag from opposite directions. Salesforce is a CRM platform that sells AI as a layer on customer data: Einstein predictions, Agentforce agents and models APIs, and a Data Cloud metered by credits, with Sales Cloud editions from $25 to $550 per user per month and API allowances that scale with the edition, 100,000 daily calls on Pro up to 5 million on Unlimited. OpenAI sells the models themselves, from GPT-5.4 Nano at $0.20 per million input tokens to GPT-5.5 Pro at $30, with a Batch API at half price and six usage tiers that advance automatically on spend. Both are hosted, paid, and self-serve to start. Salesforce documents 36 OAuth scopes. OpenAI authenticates with a project-scoped bearer key, and its OAuth at auth.openai.com grants user identity only, not API authority. A buyer is rarely choosing one instead of the other. What a comparison can show is how two very different companies expose themselves to a developer, and the answer is: almost identically.
The APIs. Salesforce publishes 185 API pages from 266 contracts, none derived by us, including the Salesforce REST API, the Salesforce Pub/Sub API over gRPC, the Salesforce GraphQL API, and the Salesforce Einstein Prediction Service API. OpenAI publishes one unversioned REST API, an OpenAPI 3.1.0 document with 242 operations that OpenAI maintains itself under an MIT license, which the catalog carves into 40 pages from 38 contracts, among them the OpenAI Chat API, the OpenAI Batch API, and the OpenAI Audit Logs API. The disclosures run in one direction. Salesforce’s agent skills and its agentic access profile of 367 operations were generated by the catalog, and its 23 Arazzo workflows carry no first-party marker and are written against the catalog’s own split contracts, so none of them is counted as Salesforce’s work. OpenAI’s 24 agent skills and its described event surface are first-party. Both run an MCP server first-party, but OpenAI’s is a documentation server that searches and fetches the platform docs, not one that operates the API.
The experience, facet by facet. OpenAI takes three and ties three. It wins operational transparency 94.7 to 86.8, developer ergonomics 96.4 to 94.6, and contract quality by 0.2, 73.0 to 72.8. The two are exactly level on discoverability at 75.0, contract governance at 45.5, and access clarity at 100.0. On the six-facet base OpenAI leads, 82.8 to 81.4. Neither side’s regulatory facet separates them: both are scored under the same horizontal regime at 41.8. The whole difference is the conditional upsert layer. Salesforce scores 72.2 on it, because Bulk API 2.0 can upsert on an external ID and the contract says so. OpenAI scores 0.0. That layer moved Salesforce’s composite up 10.4 points and OpenAI’s up 3.2, and the 7.2-point swing is larger than the 5.8-point lead.
Is that fair? It is a real capability honestly measured. Upsert is the operation that makes a sync idempotent at the record level, and Salesforce ships it as a first-class operation. It is also a strange thing to decide a comparison on when one side sells inference, where there is little to upsert. The rubric applies the facet to OpenAI because its contract has create and update operations, and scores it at zero because none of them is an upsert. Read the ranking as a statement about data platforms: on the surface both companies expose, OpenAI is marginally the better API, and Salesforce is ahead for doing a thing only a record system needs to do.
Agent readiness is close and splits on identity. Salesforce scores 51.8 and OpenAI 48.1, both agent-ready, neither agent-native, because idempotency is unlit on both. Salesforce lights protected resource metadata and dynamic client registration, so an agent can discover its authorization server and register itself. OpenAI lights neither, which follows from a key-based API whose OAuth carries no API authority. Both document reversibility, a rate limit signal, and delegated identity, and both have error semantics verified. Neither has an agent card, a dry-run mode, or a well-known catalog, and against a tag where 3.2% light idempotency and 0.6% dry-run mode, neither is behind its category.
The verdict. If you are building on customer records, Salesforce is the platform and the upsert layer is exactly why. If you are building on models, OpenAI’s surface is the tighter of the two, first-party skills included. What would move OpenAI’s number is the identity layer: protected resource metadata and a client registration path would let an agent find and enroll against the API rather than be handed a pasted key, and an idempotency key on create would clear the gate both of them are stuck behind. What would move Salesforce’s is the same gate, plus first-party versions of the workflow and agent artifacts the catalog currently writes on its behalf.