It has been almost two years since OpenAI released o1, a model that popularized the idea of LLM-based reasoning models.[reference:95] DeepSeek-R1 followed about four months later, together with details of a reinforcement learning with verifiable rewards recipe to train such reasoning models.[reference:96] Last week, OpenAI released the GPT-5.6 model family. It comes in three sizes, each with roughly five or six reasoning-effort settings.[reference:97]
Reasoning models are here to stay. They have become a standard part of modern model releases.[reference:98]
What Are Reasoning Models?
In the context of AI and LLM research, a "reasoning model" means a model that outputs an intermediate reasoning trace, which is like an intermediate response that works through a question or task step by step.[reference:99] This is not literal human reasoning, but rather a technical approach to improving performance on complex tasks.
Training vs. Inference Scaling
There are essentially two ways to improve reasoning task performance: training scaling and inference scaling.[reference:100]
Training scaling involves training models to reason better through reinforcement learning and other methods. Inference scaling involves giving models more time and tokens to think during inference.
Multiple Effort Modes
The key insight is that not all questions warrant the same amount of reasoning. Some queries are simple and require minimal thought, while others demand extensive deliberation. By training models to operate at multiple effort levels, developers can trade off accuracy for cost and speed.
A smaller model at a higher reasoning effort can sometimes reach a similar score as a larger model at a lower reasoning effort.[reference:101] This counterintuitive finding suggests that "parameter size" as the core metric of AI capability is losing its dominance.[reference:102]
Implementation Approaches
Researchers have proposed various methods for controlling reasoning effort:
- Adaptive Effort Control (AEC): A self-adaptive reinforcement learning method that trains models to use a user-specified fraction of tokens relative to the current average chain-of-thought length for each query.[reference:103]
- ThinkDial: An open recipe enabling seamless switching between three distinct reasoning regimes: High mode (full reasoning capability), Medium mode (50% token reduction with less than 10% performance degradation), and Low mode (75% token reduction with less than 15% performance degradation).[reference:104]
Why This Matters
Increasing the thinking budget of AI models can significantly improve accuracy, but not all questions warrant the same amount of reasoning.[reference:105] The ability to control reasoning effort gives developers a powerful lever to manage costs, latency, and quality.
For inference serving, the ability to dynamically adjust reasoning effort means more efficient use of compute resources. Users can choose to pay more for harder problems and less for simpler ones.
The Future
As reasoning models become more sophisticated, we can expect even finer-grained control over reasoning effort. The GPT-5.6 family's five to six effort settings is just the beginning. The standardization of reasoning effort controls will likely become a key differentiator among model providers, and a critical consideration for developers building AI applications.