What Is Test-Time Compute? Spending More Effort at Answer Time, Not Just Training Time
Instead of only making a model bigger, providers are increasingly letting it "think longer" on hard problems at the moment of answering — a different lever entirely.
A different lever than model size
Test-time compute refers to the amount of computation a model uses while actually generating a response — as opposed to the (separate, one-time) compute spent during training. Increasing test-time compute means letting the model work longer or more thoroughly on a specific problem at the moment it's being answered.
Why this is a genuinely different lever than making a model bigger
A larger, more capable model costs more on every single request, forever. Spending more test-time compute is more selective — a system can let a model "think longer" specifically on hard problems while answering simple ones quickly and cheaply, rather than paying the larger model's cost uniformly on everything.
Where this idea shows up
This is the core mechanism behind reasoning models — generating extended internal reasoning before answering — and reflects a broader industry shift toward treating "how much effort to spend answering this specific question" as its own tunable dial, separate from model size itself.
Frequently Asked Questions
What is test-time compute?
The amount of computation a model uses while generating a response, as opposed to the separate, one-time compute spent during training — increasing it means letting a model work longer or more thoroughly on a specific problem.
How is test-time compute different from just using a bigger model?
A bigger model costs more on every single request. Test-time compute is more selective — a system can spend extra effort specifically on hard problems while answering simple ones quickly, rather than paying a larger cost on everything.