STARTUPS-VCMarch 26, 20261 min read17 views

Intercom's new post-trained Fin Apex 1.0 beats GPT-5.4 and Claude Sonnet 4.6 at customer service resolutions

Intercom has announced Fin Apex 1.0, a purpose-built AI model that outperforms GPT-5.4 and Claude Sonnet 4.6 in customer support metrics. The model achieves a 73.1% resolution rate while running at one-fifth the cost of frontier models.

Intercom is taking an unusual gamble for a legacy software company: building its own AI model. The 15-year-old, Dublin-based platform announced Fin Apex 1.0, a small, purpose-built AI model that outperforms leading models from OpenAI and Anthropic in customer support metrics. According to benchmarks, Fin Apex 1.0 achieves a 73.1% resolution rate, compared to 71.1% for GPT-5.4 and 69.6% for Claude Sonnet 4.6. The model is also faster, delivering responses in 3.7 seconds, and reduces hallucinations by 65% compared to Claude. Most importantly for enterprises, it runs at one-fifth the cost of frontier models. Intercom declined to name the specific open-weights base model used, confirming only that it has hundreds of millions of parameters. CEO Eoghan McCabe argues that the future of AI lies in "post-training" using proprietary domain data rather than generic pre-training. This pivot is paying off, with Intercom's AI segment approaching $100 million in ARR and driving significant company growth.

Source: VentureBeat

More in this category

startups-vc

The cleanup trap: Stop asking RAG to fix bad data

Many generative AI initiatives fail not because of model limitations, but due to weak enterprise data foundations. Relying on RAG to clean up bad data at the retrieval layer is a costly mistake.

NOW LET US Related – Agents think in milliseconds, legacy infrastructure doesn't. LinkedIn, Walmart and Zendesk shared how they closed the gap at VB Transform 2026

startups-vc

Agents think in milliseconds, legacy infrastructure doesn't. LinkedIn, Walmart and Zendesk shared how they closed the gap at VB Transform 2026

Legacy infrastructure, not the models themselves, is what's actually slowing AI agents down. That was the shared conclusion of three infrastructure leaders — from LinkedIn, Walmart, and Zendesk — at VB Transform 2026.

startups-vc

China’s Moonshot AI releases Kimi K3, the largest open-source model ever, rivaling top U.S. systems

Beijing-based Moonshot AI has launched Kimi K3, a massive 2.8-trillion-parameter model that claims the title of the world's largest open-source AI. Benchmarks show the model performing on par with leading proprietary systems from OpenAI and Anthropic, marking a major milestone for open-source AI.

startups-vc

The desktop infrastructure problem that Kubernetes finally solves

While most enterprise workloads have migrated to cloud-native models, virtual desktop infrastructure (VDI) has remained isolated. Bringing containerized workspaces to Kubernetes unifies operations, reduces overhead, and enhances security.

startups-vc

DeepSeek cut prices 75%. The 100x problem remains

DeepSeek's recent 75% price cut on its V4-Pro model should have been great news, but developers are finding that cheaper models don't guarantee better margins. The culprit is 'token amplification' in AI agents, which consumes resources faster than prices fall.

startups-vc

Forget typosquatting; slopsquatting is the software supply chain threat created by AI coding tools

Slopsquatting represents an emerging supply chain threat made possible by AI hallucinations. As developers increasingly rely on AI coding assistants, they unknowingly grant cybercriminals access to their software from day one.