Companies are training smaller AI models for in-house tasks at lower cost

In short: Some companies are getting strong results by training smaller, cheaper AI models on their own work, instead of relying only on the biggest AI systems.

What's going on

Big AI systems from companies like OpenAI and Anthropic are often used as general tools that can handle many kinds of questions. But a growing line of work suggests that smaller models can do just as well, or sometimes better, when they are trained for one specific job.

This process is often called “fine-tuning,” which is like giving a model extra lessons focused on a narrow subject. The Financial Times highlights a legal example where Harvey AI and Fireworks AI trained a lower-cost Chinese model, Kimi 2.6, to do better on legal tests. They filtered out the model’s wrong answers and used the good examples to retrain it. Harvey reported almost a 40% improvement, with performance similar to top models, at about one eleventh of the cost.

Another example came from Bridgewater Associates, the investment firm founded by Ray Dalio. Bridgewater partnered with Thinking Machines Lab to fine-tune a model using Bridgewater’s own internal records and the way its investment professionals make judgments. Bridgewater reported the tailored model reached 85% accuracy and made almost 30% fewer errors than “frontier” models (the biggest, most capable general models), at about one fourteenth of the cost. The results have been reported by the companies involved and were not independently verified.

What to watch

If more companies can build models that work like trained staff members who learned the company’s methods (rather than a generic helper), big AI providers could face pressure on pricing. It also raises practical questions, including whether employees will want to share hard-earned knowhow, and whether using Chinese base models could create political or business risks.

Source: Financial Times