Companies Adopt Switching Between AI Models in 2025 | Keryc
Perplexity published 2025 data that shows something familiar: companies stopped using a single AI model as the standard and began treating models like a menu you switch from depending on the task. Why does that matter for your team or company? Because making access flexible can be more valuable than betting on the “best” model today.
What the numbers reveal
In 2025 the figures show clear fragmentation. At the start of the year two models concentrated more than 90% of enterprise usage; by year-end that picture changed: four models each exceeded 10% and the leader accounted for only about 23% of queries. What does that tell us? Preferences diversified as new, more specialized options appeared.
Perplexity, which claims to give access to models from multiple providers through a single interface, also reports that 92% of Fortune 500 companies use its platform. That gives it a privileged window into real behavior inside large organizations.
Key data
12.5% of enterprise seats are power users (active at least 12 out of every 28 days).
40% of those power users use 6 or more models; regular users, 20%.
The top 50 accounts use on average 30 models; typical accounts, 7.
43.6% of organizations used more than one model in 2025.
9.1% of users used multiple models in the same day.
Among those who pick specific models, 53% switched models within a single day at least once in 2025.
How models are being used by task
Not all models compete on the same ground. Some become preferred for specific tasks:
Programming: Claude models handled 38% of queries; at the organizational level 40% of companies use it by default, versus 22% that prefer GPT.
December 2025 - preferences by function:
Visual arts: Gemini Flash 40%
Financial analysis: Gemini 3.0 Pro Thinking 31%
Debugging: Claude Sonnet 4.5 30%
Software development: Claude Sonnet 4.5 29%
Legal / court cases: Claude Thinking 23%
Medical research: GPT-5.1 Thinking 13%
These percentages show that specialization matters: one model can be excellent for visual art and another for financial analysis. And that changes fast: what leads today might not lead next quarter.
Market dynamics during the year
At the start of 2025 Claude Sonnet 4 and GPT-4o made up 91.5% of enterprise use (47.5% and 44%). By year-end the pie was more split: Gemini 3.0 Pro Thinking 23.3%, Claude Sonnet 4.5 20.6%, Claude Sonnet 4.5 Thinking 10.7% and GPT-5 7.9%.
When new models are released they often spike to very high usage for a few days (over 50%) while teams experiment. That effect drops quickly: the following week they usually stabilize around 35% at most. In 2025 forty-six models were released and Perplexity made them available on its platform within 24 hours, which makes testing easy without managing separate contracts and access.
What this means for teams and companies
No tech-speak: if your team only allows one model, you’re probably leaving performance on the table. The practice of “model-switching” (changing models by task) has clear benefits:
Better fit of results by function: choose for accuracy, creativity, speed, or cost.
More experimentation: power users find combinations that optimize workflows.
Resilience to rapid changes: when a new model outperforms in something, you can adopt it without disrupting everything.
Practical tips if you manage AI in a company:
Give access to several options and control costs with limits and monitoring.
Define use cases and metrics by task (for example: best for legal summarization vs best for visual generation).
Encourage power users to document discoveries and create usage templates.
Monitor preference changes quarterly: in this market curves change fast.
A simple example
Imagine a finance team: for numeric analysis you use a model that prioritizes accuracy and traceability; for slide decks and visuals you pick another that produces flashier charts and narratives. If both are available without friction, the process speeds up and each task gets the right tool.
In the end, the lesson is practical: rather than betting everything on a single winner, it’s worth building processes that let you choose, test, and switch models based on what each task needs.
Summary: In 2025 companies moved from a single model to using multiple AI models by task. Perplexity’s data show fragmented usage, specialization by function, and an increase in power users who switch models within the same day.
Stay up to date!
Get AI news, tool launches, and innovative products straight to your inbox. Everything clear and useful.
Companies Adopt Switching Between AI Models in 2025
