Salesforce Koa vs Claude vs GPT-5.5 for CRM Agents (2026)
The short answer
Koa is a CRM specialist, not a frontier replacement. On Salesforce’s CRM Bench it scores 0.86, within a point of Claude Opus 4.8 (0.87) and four points behind GPT-5.5 (0.90), while Salesforce claims three times fewer errors on CRM actions and a lower token bill. Choose Koa for high-volume, well-defined Agentforce workflows where data must stay inside Salesforce; keep a frontier model for open-ended reasoning.
| Salesforce Koa | Claude Opus 4.8 / Opus 5 | GPT-5.5 → GPT-5.6 Sol | |
|---|---|---|---|
| Announced | September 15, 2026 (Dreamforce) | Opus 4.8: spring 2026; Opus 5: July 24, 2026 | GPT-5.5 superseded by GPT-5.6 line (Sol/Terra/Luna) |
| Base | NVIDIA Nemotron 3 Super (open weights), post-trained by Salesforce | Anthropic proprietary | OpenAI proprietary |
| CRM Bench (Salesforce technical report) | 0.86 | 0.87 (Opus 4.8) | 0.90 (GPT-5.5) |
| Error rate on CRM actions | ”3x fewer errors” (Salesforce claim) | Baseline | Baseline |
| Where it runs | Inside Salesforce’s trust boundary; weights controlled by Salesforce | Anthropic API / cloud partners via Agentforce gateway | OpenAI API via Agentforce gateway |
| List price (per MTok) | Not published; billed through Agentforce | $5 / $25 | Sol $4 / $20 (Sep 2026) |
| Availability | Pilots now; US GA winter 2026 | GA | GA |
| Best for | Opportunity updates, case routing, follow-ups, lead qualification | Default Agentforce Coworker/Slackbot reasoning; drafting; long-horizon tasks | Broad reasoning, coding, multimodal |
What Koa actually is
Koa is the first time Salesforce has trained its own reasoning model rather than a small task model. Jayesh Govindarajan, EVP of Salesforce AI, told TechCrunch that reasoning had always been outsourced to frontier providers because there was no “sovereign American pre-trained model” that was state of the art with clear data provenance — until Nemotron. Salesforce and NVIDIA post-trained Nemotron 3 Super with SFT plus GRPO reinforcement learning using NeMo RL, NeMo Gym and NeMo AutoModel.
The training corpus is entirely synthetic: simulated personas (an irate caller into a service center, a rep trying to close a deal) paired with tasks and the exact sequence of tool calls needed to finish them, across more than 14 industries including manufacturing, financial services, healthcare and travel. Salesforce’s technical report describes turning Agent Script workflow rules into simulated tasks and rewarding successful tool actions. No customer data was used, and Salesforce says none crosses the trust boundary at inference.
How the three compare on CRM work
Task success. Salesforce’s CRM Bench (opportunity updates, case routing, follow-up scheduling and similar) puts GPT-5.5 first at 0.90, Opus 4.8 at 0.87 and Koa at 0.86. That is a narrow spread; the marketing line “matches or exceeds leading model performance” is true against Claude and a stretch against GPT.
Errors. The “three times fewer errors” claim is the more interesting number for an agent that writes to production CRM records, because a wrong stage change or mis-routed case costs more than a slow one. The claim is Salesforce’s own and has not been independently reproduced.
Token efficiency and cost. NVIDIA’s Kari Ann Briski framed Koa’s advantage as “sovereign AI, time to first token, efficient reasoning.” Salesforce has not published Koa pricing; as a hosted Agentforce option its cost is embedded in consumption. Frontier list prices as of September 2026 are $5/$25 per million tokens for Opus 4.8 and Opus 5 and $4/$20 for GPT-5.6 Sol. Remember that GPT-5.5, the model in Salesforce’s benchmark, is no longer OpenAI’s current line; a fresh CRM Bench run against GPT-5.6 Sol and Claude Opus 5 would be more informative.
Data residency. This is Koa’s clearest win. Salesforce controls the weights and runs inference in its own infrastructure. The same Nemotron stack is going into Missionforce for government and regulated buyers, including private-cloud and air-gapped deployments.
Salesforce is customer, partner and competitor at once
Koa does not mean Salesforce is leaving Anthropic or OpenAI. On August 26, 2026 Salesforce and Anthropic announced Claudeforce, and Claude powers Agentforce Coworker and Slackbot by default. On September 15 Salesforce also launched AIforce, opening its data, permissions and actions to outside AI tools, and the first Claudeforce experience — Salesforce in Claude with 37 prebuilt sales skills — went into beta the same day. Koa gives Salesforce a model it controls for recurring, well-specified reasoning, which also strengthens its hand in negotiating frontier model costs.
Decision guide
- High-volume, structured CRM actions with strict data residency → Koa (pilot now, GA winter 2026).
- Reps working conversationally across Salesforce, Slack and email → Claude via Salesforce in Claude or Agentforce Coworker.
- Open-ended reasoning, document generation, multimodal input → frontier model through the Agentforce gateway (Claude Opus 5 or GPT-5.6 Sol).
- Government / air-gapped → Missionforce with Nemotron-based models.
Use per-agent model selection to run Koa on one workflow, measure error rate against your current model for a month, then expand.