AI competition shifts toward cheaper systems that pick the right model

The AI race is moving from a contest over the biggest model to a fight over which system can get work done at the lowest useful cost. For everyday investors, that shift could change the economics behind OpenAI, Anthropic, cloud providers and the companies spending heavily to put AI into their products.

CNBC reported that companies are starting to focus less on using the most powerful model for every task and more on matching each job with the right model, data and computing setup. In plain terms, that means an AI product may use a cheaper model for routine work, then send harder tasks to a more advanced model only when needed.

Perplexity CEO Aravind Srinivas told CNBC that “the model alone is no longer the product.” He said the product is becoming the “harness” or orchestration system around the model, meaning software that decides which model to use and connects it with outside tools and company data.

Srinivas said the right answer is to use the model that fits the task. A customer service request may not require a premium model. A difficult coding problem may need one. A routine internal process could run on an open model, while a more complex step could be passed to a stronger system.

Open models add pressure

The shift comes as companies watch AI spending more closely, according to CNBC. That creates a new challenge for OpenAI and Anthropic, which have built their businesses around selling access to advanced proprietary models.

Open-weight models are central to the change. These are AI models whose underlying parameters, or “weights,” can be downloaded, adjusted and run by companies themselves. CNBC reported that these models are improving and can cost less to operate than premium closed models from large AI labs.

Perplexity this week previewed a computer-use system built around GLM 5.2, an open model from China’s Z.ai, according to CNBC. The system is designed to let a lower-cost model handle more of the workload and call on a stronger model only when necessary.

Benchmark general partner Peter Fenton told CNBC that he expects open-weight models to generate more than 90% of tokens over the next 18 to 24 months, and possibly by the end of the year. Tokens are the pieces of text and data that AI models process and produce.

Fenton said inference margins at frontier model companies could face pressure. Inference means the process of running a trained AI model to produce an answer. If customers can run “good enough” open models without paying a provider’s markup, the business case for premium model access may become harder.

Fenton also told CNBC that cost is not the only reason companies are considering open models. Smaller models trained or adjusted for a specific job can sometimes respond faster and perform better than large general-purpose systems.

Control becomes part of the sale

Benchmark has invested in Ollama, a company that helps developers and businesses download, run and manage open models, CNBC reported. Ollama CEO Jeff Morgan told CNBC that companies care not only about where a model was built and trained, but also where it runs and how it runs.

Morgan said Ollama has been adopted by more than 85% of the Fortune 500, including companies in aviation, insurance and health care. He said many businesses begin with smaller models close to their own data, then expand to larger open models as they gain confidence.

The rise of open models also has a geopolitical angle. CNBC reported that several competitive open-weight models are coming from Chinese labs, including Z.ai and DeepSeek. Srinivas said the U.S. should support open models because they can make AI cheaper and more widely available to small businesses in America and allied countries.

The trend could also affect the data center buildout tied to AI demand. Srinivas told CNBC that some AI work may eventually run locally on consumer or business devices, while harder tasks still go to cloud-based models. That would point to a more mixed AI setup, rather than one that sends every task to large data centers packed with high-end chips.

For investors, the key issue is whether the largest AI labs can keep charging premium prices as open models improve and customers become more selective about which model handles each task.

This story draws on original reporting from CNBC.