The niche model beats the giant

When Salesforce announced it was buying the support-agent company Fin for $3.6 billion, most of the coverage went to the number. The more interesting detail is what the agent runs on. Not GPT, not Claude, not Gemini — Apex, Fin's own model, purpose-built for customer support, which Salesforce says outperforms the top frontier models at resolving tickets. It resolves 76% of support volume end to end.

Sit with that. A model you've never heard of, smaller than the household-name giants, beats them at the one job it was built for. That's not a fluke — it's how specialization works, and it argues against the reflex almost everyone has when they start an AI project: reach for the biggest, smartest model and assume it wins. For a narrow job, it usually doesn't. Let me explain.

Biggest is a general-purpose answer to a specific question

The frontier models are extraordinary because they do everything passably — write code, plan trips, explain tax law, draft emails. That breadth is exactly why they're not optimized for any single one of those things. A giant general model is a brilliant generalist: wide knowledge, no specialty.

Most real products don't need a generalist. They need one job done extremely well. Resolving support tickets isn't "know everything" — it's "understand this company's products, follow its policies, handle these few hundred recurring situations correctly, and know when to hand off to a human." A model trained on exactly that, over years of real tickets, will beat a bigger model that's splitting its attention across the entire universe of human questions. Depth on the task beats breadth across all tasks.

Specialized usually means smaller, cheaper, and faster too

The part that makes this more than a curiosity: the specialized model doesn't just win on quality for its task — it usually wins on cost and speed too. A model purpose-built for support can be smaller, because it doesn't carry the weight of knowing everything else. Smaller means cheaper to run and faster to respond. So you're not trading quality for economy. On the narrow job, the specialist can be better and cheaper and faster than the giant, all at once.

That flips the usual mental model. We tend to assume there's a ladder — small-and-cheap at the bottom, big-and-best at the top — and you pick how far up you can afford to climb. For a defined task, that ladder is the wrong picture. The question isn't "how big a model can I afford," it's "is there a model built for exactly this," because if there is, it likely beats the giant on every axis you care about.

What to do with this

Next time you start an AI feature, resist the instinct to default to the biggest model. Ask a different question first:

• Is my task narrow and well-defined? If it's one job with clear rules — support, document extraction, classification, a specific kind of drafting — it's a candidate for a specialist, not a generalist.
• Does a purpose-built model already exist for it? A model fine-tuned for your domain may beat the frontier out of the box, for less money.
• Could I specialize a smaller model myself? Fine-tuning or tightly prompting a small model on your actual task often beats renting a giant to do the same thing generically.

Reach for the giant when the job really is open-ended. For everything narrow, look for the specialist first.

The bottom line

Salesforce paid $3.6 billion for an agent, and the engine inside it isn't the biggest model in the world — it's a smaller one built for the job. That's the whole lesson, sitting in plain sight inside the headline.

For a narrow, well-defined task, a model purpose-built for it usually beats the general giant — on quality, cost, and speed at once. The biggest model is the right default far less often than people assume. Match the model to the job, and the job is usually narrower than the giant.