Artificial intelligence has fundamentally shifted its approach to getting smarter: instead of making models larger during training, AI labs are now giving models more time to "think" when answering questions. This strategy, called test-time compute or inference-time reasoning, allows models to generate longer chains of thought, explore multiple solution paths, and verify their own work before responding. The results are dramatic. Gemini Deep Think reached gold-medal standard at the 2025 International Mathematical Olympiad, solving 5 of 6 problems, while OpenAI's Deep Research scored 26.6% on Humanity's Last Exam, roughly triple the performance of earlier reasoning models.

What Is Test-Time Compute and Why Does It Matter?

For most of the deep-learning era, AI capability came from training scale: bigger models, more data, more computing power during the initial training phase. Test-time compute represents a different axis entirely. Instead of investing all the computational effort upfront, models now spend additional computing resources at the moment of answering a question. This approach is powered by reinforcement learning, which trains the model to use that extra thinking time effectively.

The practical payoff is substantial in domains where mistakes are expensive. Across 2025 and 2026, frontier AI models gained approximately 30 percentage points on Humanity's Last Exam, a benchmark specifically designed to be difficult for artificial intelligence. Software engineering benchmarks show even more dramatic gains: coding solve rates on SWE-bench jumped from 4.4% to 71.7% in a single year, while the widely used MMLU knowledge benchmark is now saturated above 92%.

Which AI Models Are Using This Approach?

Several major AI labs have built reasoning models based on test-time compute principles. OpenAI's o-series models, Google DeepMind's Gemini Deep Think, DeepSeek-R1, and Anthropic's Claude with extended thinking all operate on this inference-scaling model. These systems represent a fundamental rethinking of how to allocate computational resources in AI systems.

The benchmark results speak to the effectiveness of this approach. Gemini Deep Think achieved an official gold medal at the 2025 International Mathematical Olympiad with a score of 35 out of 42 points, solving 5 of 6 problems, marking the first time an AI system reached gold-medal standard at this prestigious mathematics competition. OpenAI's Deep Research, which combines test-time compute with agentic multi-step reasoning and web tools, scored 26.6% on Humanity's Last Exam in February 2025, compared to roughly 9% for earlier-generation reasoning models like o1 and DeepSeek-R1.

What Are the Trade-Offs of Reasoning Models?

The gains in accuracy come with significant costs. Reasoning models consume far more tokens per query than standard models, which increases both latency (response time) and computational expense. This trade-off is particularly important for enterprises deciding which tasks justify paying for extended reasoning.

The cost picture is complex. Per-token inference costs fell roughly 280-fold between late 2022 and late 2024, a dramatic decline driven by hardware improvements and optimization. However, because reasoning models generate many more tokens per answer, the per-query cost of solving a difficult problem can actually rise even as the unit price per token drops. This means organizations must carefully evaluate which high-stakes tasks warrant the additional compute investment.

How to Evaluate Whether Test-Time Compute Is Right for Your Use Case

• Task Complexity: Test-time compute delivers the largest gains in math, code, and science problems where mistakes are expensive and solutions can be verified. Simpler tasks may not justify the added latency and cost.
• Accuracy Requirements: If your application requires near-perfect accuracy in high-stakes domains like medicine, law, or engineering, the performance jump from 9% to 26.6% on difficult benchmarks may be worth the additional compute cost.
• Latency Tolerance: Reasoning models take longer to respond because they spend more time thinking. Applications requiring instant responses may not be suitable, while research tools or complex analysis tasks can tolerate longer wait times.
• Cost-Benefit Analysis: Calculate whether the accuracy improvement justifies the higher per-query cost for your specific use case, considering that per-token prices continue to decline even as reasoning models consume more tokens.

Why Is This Shift Significant for AI Development?

The move toward test-time compute represents a fundamental reorientation of AI research and development. For years, the industry focused on scaling up models during training, leading to increasingly large language models (LLMs) with billions or trillions of parameters. Test-time compute offers a different path to capability: instead of making models permanently larger, systems can allocate more computing resources dynamically based on the difficulty of each specific question.

This shift has practical implications for data centers and energy consumption. Even as per-token costs decline, total inference compute and energy demand are climbing because reasoning models consume significantly more tokens per query. This creates a new economic equation for AI infrastructure, where the focus shifts from training efficiency to inference optimization and the strategic allocation of compute at query time.

The evidence suggests that test-time compute will remain a central strategy for advancing AI capabilities in domains where accuracy matters most. The benchmark jumps in mathematics, coding, and complex reasoning tasks demonstrate that allowing models to think longer can clear performance thresholds that seemed out of reach just months earlier.