Intercom has built Fin Apex 1.0 on an undisclosed open-weights foundation, using proprietary post-training to surpass frontier AI models on customer issue resolution. The launch sharpens the debate over whether open foundations plus private data are becoming the new enterprise AI moat.
Intercom has launched Fin Apex 1.0, a post-trained, domain-specific AI model built on an undisclosed open-weights base, claiming it now outperforms GPT-5.4 and Claude Sonnet 4.6 on customer support resolution—the KPI that matters most in enterprise service AI.
According to the company, Fin Apex delivers a 73.1% resolution rate, ahead of 71.1% for GPT-5.4 and Claude Opus 4.5, and 69.6% for Claude Sonnet 4.6. The model also responds in 3.7 seconds, runs 0.6 seconds faster than the next-fastest rival, cuts hallucinations by 65%, and operates at roughly one-fifth the cost of direct frontier model deployments.
The stronger open-source angle lies in Intercom’s admission that Apex is built on an open-weights foundation, while refusing to disclose which one. That contradiction places transparency under scrutiny even as the company positions post-training as the real differentiator.
As Eoghan McCabe said, “Pre-training is kind of a commodity now. The frontier, if you will, is actually in post-training.” Intercom says its edge comes from proprietary customer-service transcripts, reinforcement learning from real resolution outcomes, tone calibration, issue-completion detection, and frustration recognition.
The business stakes are already material: Fin handles 2 million conversations weekly, is nearing $100 million ARR, and is expected to become half of Intercom’s $400 million ARR business next year.
Accessible only within Fin and not as a standalone API, Apex underscores a wider shift from general frontier models to highly specialised enterprise AI systems.
