The AI race swaps giant models for smart and cheap systems

Running for ever-increasing artificial intelligence models is giving way to a new strategy: smart systems that choose the most suitable and inexpensive model for each task. Perplexity CEO Aravind Srinivas stated in an interview with CNBC that "the real product is the system that decides which model to use", while Benchmark's investor Peter Fenton predicts that open weight models will soon meet most business uses. This change pushes the economy of the largest providers and makes room for more accessible solutions.

Why does the AI race migrate to smart and cheap systems?

Build giant models like GPT-4 or Gemini Ultra costs hundreds of millions of dollars and consumes energy equivalent to a small town. The strategy of using a single model for everything is also inefficient: simple tasks such as summarizing an email do not need all the computational power of a border model. Perplexity, for example, operates a routing system that dynamically selects between internal models, external APIs and open weight models, combining performance and cost. Srinivas compares this to a "model orchestrator" — intelligence is on the router, not the engine.

How do open weight models create smart and cheap systems?

Models such as Llama 3 (Meta), Mistral (French) and open-weight community releases already compete closely with proprietary solutions in language benchmarks, reasoning and coding. Fenton believes that soon the opening and reduced cost of these models will make them the standard choice for most use cases. As Hugging CEO Face already pointed out, enterprises are exchanging expensive APIs for open source, reducing operating costs by up to 80%.

What do smart and cheap systems mean for Brazil?

For startups and medium-sized companies in Brazil, the trend is a breath of fresh air. Instead of relying on expensive APIs in dollars, it is possible to run local models in Portuguese with modest hardware, thanks to techniques such as efficient quantization and fine-tuning. Data sovereignty is also important: with open models, companies can train with sensitive data without sending them to servers in the US. The competition ceases to be "who has the best model" and becomes "who builds the best orchestration system" — a fairer arena for lean teams.

The savings of AI providers with smart and cheap systems?

The paradigm shift directly affects the business model of companies such as OpenAI and Anthropic. They profit from sold tokens. If routing systems start to prefer open weight models whenever possible, the revenue per query plummets. The answer has been to create premium subscription layers (such as ChatGPT Plus) and enterprise contracts with fixed price guarantees. But the pressure for cheapening is already visible: in the last year, the cost per token of models such as GPT-4 fell by about 60%, according to CNBC itself.

Frequently Asked Questions about Smart and Cheap AI Systems

What is a model routing system?

It is a software that analyzes the user's question and decides which AI model — among several available — should process it, balancing quality, speed and cost.

Are open weight models as good as owners?

In many benchmarks, models such as Llama 3 70B and Mistral Large already match or overcome previous GPT versions, especially in reasoning and code tasks. The difference for the largest proprietary models is decreasing rapidly.

How does this change affect those who already use paid APIs?

The trend is that providers reduce prices to compete with open models, and hybrid services arise that combine APIs with local models. For companies, it is time to evaluate costs and explore fine-tuning of open models as an alternative.

> Source: CNBC – The AI race is hiding from bigger models to cheater, smarter systems