Reasoning Models & Test-Time Compute

For most of AI's history, a model's capability was determined almost entirely at training time — the data it learned from, the scale of the run, the quality of the fine-tuning pass. In 2025, a second axis emerged: test-time compute. By spending more computation during inference — generating intermediate reasoning steps, verifying partial answers, and backtracking on failure — models can solve problems they would otherwise get wrong. This shift fundamentally changed both what AI systems can do and how organisations should budget for them.

What Is Test-Time Compute?

Standard language models generate a response in a single forward pass: input goes in, tokens come out. Reasoning models break that single pass into an extended chain-of-thought phase before producing a final answer. The model "thinks out loud," working through sub-problems, checking its logic, and correcting mistakes — all invisible to the end user but consuming tokens (and therefore compute) proportional to problem difficulty.

The key insight, validated across 2025 research, is that inference compute scales with accuracy in a similar way to pre-training compute — more compute generally yields better answers, up to a ceiling determined by the task type. This creates a new dimension for practitioners: for any given task, you can now choose between a cheaper fast model and a more expensive deep-thinking model, with the accuracy difference made concrete by published benchmarks.

Why It Changed AI Capabilities in 2025

Before reasoning models, hard mathematical proofs, complex multi-step coding problems, and adversarial logic puzzles were reliable failure modes for frontier LLMs. The 2025 generation of reasoning models changed this dramatically:

OpenAI o3 (April 2025) scored 88% on ARC-AGI — a benchmark designed to resist AI — compared to o1's 32%. It also supports multimodal reasoning, allowing it to analyse diagrams and sketches within its chain-of-thought phase.
OpenAI o3-mini (January 2025) proved that test-time gains are achievable at commodity inference costs, not just expensive research budgets.
Process supervision research (formalised throughout 2025 via PRM800K and derivatives) showed that training models to receive feedback on individual reasoning steps — rather than only final answers — produces more reliable and interpretable chains of thought.

Key Models

OpenAI o3 and o4-mini

Released April 16, 2025 alongside each other. o3 is optimised for accuracy on hard reasoning tasks; o4-mini delivers comparable performance at significantly lower cost. Both are the first o-series models with multimodal reasoning — they can incorporate images into their chain-of-thought.

DeepSeek R1

Released January 20, 2025 under the MIT License. A 671B-parameter Mixture-of-Experts model that directly competes with OpenAI o1 on math and coding benchmarks at a fraction of the compute cost. R1's technical report demonstrated that reasoning capabilities can emerge from reinforcement learning applied to a base model without requiring supervised chain-of-thought scaffolding — a paradigm shift in how the field thinks about training reasoning ability.

Qwen3

Released April 29, 2025 by Alibaba. The headline innovation is a unified thinking/non-thinking mode: developers can switch between deep reasoning and rapid response within the same model, configuring thinking-token budgets up to 38K. This simplifies cost management for applications that need both fast and slow reasoning paths within the same deployment.

Gemini 2.5 Pro Deep Think

Google's first explicit thinking-model tier, introduced at Google I/O in May 2025. Scored 84.0% on MMMU and led LiveCodeBench at time of release, validating test-time compute as a competitive differentiator outside of OpenAI.

Process Reward Models (PRMs)

Standard model training evaluates whether the final answer is correct (outcome supervision). Process reward models evaluate each intermediate reasoning step. Datasets like PRM800K (800,000 step-level labels) demonstrate that step-level supervision produces models that reason more reliably and hallucinate less in agentic deployments — because the model has learned that incorrect reasoning steps are penalised regardless of whether they produce the right endpoint.

What This Means for Practitioners

Budget reasoning compute per task type: Hard symbolic reasoning, complex code generation, and multi-step planning benefit greatly from reasoning models. Simple classification and content generation do not.
Benchmark numbers require context: A model scoring 92% on AIME 2025 at "high compute" may score 60% at standard compute. Always check which compute tier benchmarks were run at.
Hybrid deployments are emerging: Qwen3's unified mode and GPT-5's routing between fast and deep-thinking paths both suggest that production systems will increasingly route queries to the appropriate reasoning depth dynamically.

Architectures Overview — where reasoning models fit in the broader model taxonomy
Model Families — full catalogue of current reasoning and non-reasoning models
Training: Reasoning via RL — DeepSeek R1 methodology, GRPO, and process supervision
Optimizing — inference scaling laws and cost/accuracy trade-offs