Intercom just demonstrated something that should worry OpenAI and Anthropic: a 15-year-old customer service company built an AI model that outperforms their best work at its specific task.
Fin Apex 1.0, announced on March 26, resolves customer support issues at a 73.1% rate. That beats GPT-5.4 (71.1%), Claude Opus 4.5 (71.1%), and Claude Sonnet 4.6 (69.6%). The model also responds in 3.7 seconds—0.6 seconds faster than any competitor—and hallucinates 65% less than Sonnet 4.6.
“This is objectively the highest performing, fastest, and cheapest model for customer service,” wrote Intercom CEO Eoghan McCabe. “Pre-training is kind of a commodity now. The frontier is actually in post-training.”
How a Customer Service Company Beat the Labs
Intercom built Apex by taking an open-weights base model (they won’t say which one) and post-training it on their proprietary data: billions of customer service interactions from their Fin AI agent, which already handles over two million conversations weekly.
The key insight: general-purpose models from OpenAI and Anthropic are massively over-serving customer service use cases. They can write poetry, debate philosophy, and analyze scientific papers. But for the narrow task of “resolve this customer’s problem,” all that extra capability is wasted computation.
Apex strips away the generality and optimizes for what matters: resolution rate, speed, cost, and accuracy on support-specific queries.
One of Intercom’s largest gaming customers saw their resolution rate jump from 68% to 75% overnight—a 22% reduction in unresolved conversations. That’s not incremental. That’s transformative for a support operation.
The Economics Change Everything
Apex runs at roughly one-fifth the cost of using frontier models directly. It’s included in Intercom’s existing per-outcome pricing, meaning customers don’t pay more for the upgrade.
For Intercom, the math is simple: if they can deliver better results at lower cost, they capture more of the value they create. If they stayed dependent on OpenAI or Anthropic, those companies could raise API prices and squeeze Intercom’s margins at any time.
This is the vertical model playbook: own your stack, own your destiny.
The “Speciation” of AI
Former Tesla AI director Andrej Karpathy has been predicting this for months. In a recent podcast, he described the coming “speciation” of AI models—a proliferation of specialized systems optimized for narrow tasks rather than general intelligence.
The logic is compelling. A 22-billion parameter model fine-tuned on customer service data can outperform a 2-trillion parameter general model at customer service. You don’t need to solve all of intelligence to solve your specific problem.
This creates an opening for every company with proprietary domain data. Medical companies with clinical records. Legal firms with case files. Financial institutions with transaction histories. Each could theoretically build vertical models that beat frontier labs at their specific use cases.
What This Means for the Frontier Labs
OpenAI and Anthropic are racing to build AGI—artificial general intelligence that can do everything. But Intercom just proved that for many practical applications, “general” is a bug, not a feature.
The frontier labs still have advantages: vast compute resources, the best researchers, and models that can handle novel tasks. If you need an AI to do something no one has done before, you still need GPT-5 or Claude Opus.
But most business use cases aren’t novel. They’re repetitive, well-defined, and drowning in historical data. These are exactly the conditions where vertical models thrive.
The next few years may see the AI market split into two tiers:
- Frontier labs competing on general capability and novel reasoning
- Vertical specialists competing on domain-specific performance and cost
The Open-Weights Advantage
Intercom’s approach depends on one crucial factor: access to capable open-weights base models. Without Meta’s Llama, Mistral’s releases, or similar open foundations, building vertical models would require training from scratch—economically impossible for most companies.
This validates the open-source AI ecosystem. Every company that releases open weights enables countless vertical applications. The value isn’t just in the model itself; it’s in what others can build on top of it.
What You Can Do
If you’re running customer service operations, the immediate takeaway is straightforward: evaluate Fin. A 2-percentage-point improvement in resolution rate might sound small, but at scale, it means thousands of fewer escalations to human agents.
If you’re building AI products, the strategic lesson is more profound. Ask yourself: do you have proprietary data that could make a vertical model outperform general-purpose alternatives? If yes, you may have more leverage against the frontier labs than you realized.
The age of vertical models has begun. The question is no longer whether specialized AI can beat general AI at specific tasks. It’s how many tasks will fall to specialists before the generalists can respond.