OpenAI Details the Test-Time Compute Scaling Law: Shifting from Pre-Training Parameters to Inference-Time Search

The Three-Question Framework

What Happened?

OpenAI published formal empirical research detailing test-time compute scaling laws demonstrated by its o1 series reasoning models. Unlike traditional autoregressive LLMs where answer quality is bounded by pre-training compute and model parameter size, o1 utilizes large-scale reinforcement learning to generate private internal 'thinking tokens'. This search-like exploration explores multiple problem-solving hypotheses, corrects intermediate mistakes, and verifies mathematical steps prior to outputting a solution.

Why It Matters

This marks a profound paradigm shift in AI development. For the first time, researchers can improve model performance on hard problems simply by allowing the model to think longer at test time, following a smooth power law. On the qualifying examination for the International Mathematical Olympiad (AIME 2024), o1 solved 83% of problems compared to just 13% for GPT-4o.

What Could Happen Next

Enterprises are redesigning LLM infrastructure from fixed per-token billing to dynamic latency budgets, allocating 10,000+ reasoning tokens to mission-critical code verification while using lightweight models for routine queries.

Executive Key Takeaways

Test-time compute scaling: reasoning accuracy scales predictably with the number of internal thinking tokens generated.

AIME 2024 performance surged from 13.4% (GPT-4o) to 83.3% with o1, and reached 93% with test-time majority voting.

Reinforcement learning trains the model to recognize and backtrack from dead-ends in its internal scratchpad.

Inference economics: shifts spend from billion-dollar cluster pre-training runs toward per-query inference compute allocation.

On September 6, 2026, OpenAI released empirical data formalizing the 'test-time compute scaling law,' demonstrating that foundation models can achieve orders-of-magnitude reasoning gains without expanding parameter counts.

Historically, improving AI performance required scaling dataset sizes and parameter counts during pre-training, governed by the Chinchilla scaling laws. However, as frontier models approached hundreds of billions of parameters, pre-training returns began encountering diminishing returns and high capital expenditures. The o1 architecture bypasses this constraint by scaling compute during inference.

Trained with novel reinforcement learning algorithms, o1 learns to produce an extensive chain of thought before providing a final response. Crucially, the model's thinking tokens are not simply longer prompts; they represent active heuristic search. The model learns to break down multi-step proofs, construct counterexamples, catch its own arithmetic errors, and backtrack when a hypothesis fails.

Empirical results published across competitive benchmarks showed steep vertical gains. On GPQA Diamond—a benchmark of PhD-level physics, chemistry, and biology questions—o1 achieved 78.0%, surpassing human expert baselines. On SWE-bench Verified, it solved 48.9% of real-world GitHub issues.

The findings establish a dual-scaling era: pre-training builds the underlying world knowledge, while inference-time search unlocks rigorous formal reasoning.