The old idea that the biggest AI model always wins is fading fast. Companies now pick AI based on cost, control, and how well it fits specific tasks. Size alone isn’t the deciding factor anymore.
By the end of 2026, 40% of enterprise apps will use task-specific AI agents. Just a year earlier, less than 5% had them. This shows how fast businesses want AI that’s tuned for particular jobs.
Even though the cost per token has dropped sharply, company bills for AI have tripled. That’s because these new AI tools use many more tokens per task. It’s not just about cheaper tokens, but about how much AI is asked to do.
Nikesh Arora, CEO of Palo Alto Networks, says token prices need to fall by up to 90% for AI adoption to explode. Without that, many companies will hesitate before scaling their AI use.
Chinese AI models are now nearly matching U.S. labs but cost much less. This puts a hard cap on what companies can charge for decent AI output. The race isn’t just about who has the biggest model anymore.
Why Bigger Models Aren’t Always Better
Companies spent hundreds of billions on massive AI models, hoping they would stay far ahead. But buyers are choosing differently now. They want models that solve their problems efficiently, not just the largest ones on the market.
Using multiple AI models together might seem like a good way to reduce errors. But studies show that failure rates are often underestimated by more than double. One study of 67 models found a 5.2% co-failure rate, while estimates based on pairwise error correlation suggested only 2.3%.
When tasks change from multiple-choice to free-response, failure rates rise even more—to 12.7%. This means AI struggles more with open-ended questions, limiting its use in complex environments like open-ended math problems.
The Limits of Combining AI Models
Josef Chen, who studied these errors, explains that a “common-mode atom” causes many models to fail together. This is a set of queries where all models miss the mark, and no simple statistics can predict it.
Adding more models to a pool doesn’t help much. Chen says, “Adding a 20th model to your pool doesn’t buy tail coverage. The tail is shared.” The “tail” here means the hardest problems where AI fails.
He also warns that naive majority voting across different models often backfires. Weaker models can outvote the stronger one, leading to worse results. Combining models rarely beats using the single best model for tasks with clear right or wrong answers.
Teams can measure their own failure rates using simple tests. Chen says, “The measurement costs nothing, so any team can track its own co-failure rate across model generations and watch whether the tail is closing.” This helps companies decide if adding models improves performance.
In short, multi-model setups work least well where teams want them most: open-ended generation tasks. For now, focusing on the right model for the right job, at the right cost, is the smarter strategy.
Based on
- The AI race has quietly stopped being about who has the biggest model — thenextweb.com
- The real cost, security, and culture problems behind enterprise AI agents | VentureBeat — venturebeat.com
- Microsoft joins AI cost-cutting trend by relying more on its own models | TechCrunch — techcrunch.com
- Enterprises using multiple AI models are underestimating failure rates by 2.25x | VentureBeat — venturebeat.com
