Salesforce debuts Koa, a homegrown agent model built on Nemotron 3 Super

Salesforce's September 15 announcements place it on three sides of the same table. It sells AI software, it distributes rival models through its own products — Claude powers Agentforce Coworker and Slackbot by default — and now it fields its own reasoning model in Koa. The August 26 partnership announcement with Anthropic described the companies as strategic customers of each other, so the competitive framing coexists with a commercial one.

Koa is the piece that changes the arithmetic. Because Salesforce controls its version of the model and runs requests on its own systems, recurring sales and support reasoning can stay on infrastructure it governs, and its technical report benchmarks Koa at 0.86 on CRM Bench against 0.87 for Claude Opus 4.8 and 0.90 for GPT-5.5. AIforce pushes in the opposite direction, opening Salesforce's data, business logic, permissions, and actions to outside AI tools so a salesperson can work through Claude while Salesforce governs access and carries out changes to CRM records.

The near-tie on CRM Bench suggests the contest is less about raw capability than about cost and control. Whether Salesforce shifts more agent reasoning onto Koa likely hinges on how it performs on live deal updates, ticket routing, and handoffs, and on the total serving cost and permission-failure rates that selected customers record during the pilots — a question the winter 2026 general availability timeline leaves open.