A 4x smaller AI model, when given an effective way to reason through a problem, matched or exceeded the results of a model up to 14x larger on hard reasoning tasks. The result points developers towards another route to better performance: allocating more computing work to a question after it reaches the system.
This is known as test-time compute. It means a model is allowed to keep working after a prompt arrives, rather than relying solely on extra scale during training. Before returning a final answer, one reasoning request can produce 16,000–128,000 internal thinking tokens.
OpenAI’s o1 offered a clear example in maths competitions, where accuracy rose roughly log-linearly as the number of thinking tokens increased. Current reasoning requests use 10–100x more tokens than older one-shot models required.
Several techniques can turn that added processing into better answers. On GSM8K grade-school maths, self-consistency moved PaLM 540B from 56.5% to 74.4%, a 17.9-point improvement produced by voting alone.
Process reward models beat checks that assess only the final outcome by 10–20% on maths tasks. Claude Opus 5.5, GPT-5.6 Sol, Grok 4.7, DeepSeek-V4-Pro, Qwen3.8 and GLM-5.3 use a 'think first, then answer' approach by default.
For developers, the choice is increasingly where to spend computing power: on a larger model during training, on longer reasoning once a question arrives, or on both. A smaller system given enough room to work through a problem may produce results comparable with much larger models.
Comments (0)