Microsoft researchers have found a way to make a cheaper AI model outperform its own expensive reasoning mode, and the trick costs about as much as a latte.

A paper published on August 11, titled “Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills,” introduces a passive skill-distillation method that extracts compact rule sets from a small number of task examples. Those rules, written in plain markdown, are injected into a model’s system prompt and effectively teach it shortcuts that bypass the need for costly chain-of-thought reasoning. The target model in this case is GPT-5.4-mini, which launched back in March.

The result matters beyond one benchmark win. Reasoning modes drive up inference costs because they generate long chains of intermediate tokens before answering, and token output is a primary cost driver for AI providers at scale. Cutting output tokens by multiples—while holding or improving accuracy—directly changes the economics of running agentic workloads in production, where the same task may be executed thousands or millions of times.

How It Works

The process is deceptively simple. A coding agent reviews between 35 and 50 task trajectories, essentially recordings of how a model approached—and sometimes failed at—specific tasks. From those trajectories, the agent distills a natural-language “skill” document ranging from 40 to 130 lines of markdown. These are not abstract embeddings or fine-tuned weight adjustments. They are readable, auditable rules derived from what went wrong and what went right.

That readability is a meaningful departure from the two dominant ways domain knowledge is usually baked into models: fine-tuning, which requires training runs and specialized hardware, and opaque retrieval or embedding-based approaches. Because the distilled knowledge sits in plain text, developers can inspect, edit, and version-control a skill document the same way they would any other prompt asset—no MLOps pipeline required.

Once that skill document is folded into the system prompt of a non-reasoning model, something interesting happens. The model recovers between 55% and over 100% of the performance gap that normally separates reasoning and non-reasoning modes. In plainer terms, the cheap mode starts performing like the expensive mode—and sometimes better.

The compute cost for generating each domain’s skill document runs between $1 and $3, with no model retraining required.

Benchmark Results

The research team, led by Agamdeep Singh, Srishti Gautam, Priyanshu Gupta, Nikita Mehrotra, Tanmay Bakshi, and Sumit Gulwani, evaluated the approach across four agentic benchmarks, including ALFWorld and SpreadsheetBench-Verified.

On ALFWorld, the skill-enhanced GPT-5.4-mini achieved a score of 0.787, while the full reasoning mode scored 0.713. The cheaper, faster version did not just close the gap—it surpassed it.

Across benchmarks, the skill-augmented model used 2.7 to 6 times fewer output tokens than the reasoning mode. Fewer tokens also means faster responses, a practical advantage for interactive agents where latency shapes user experience.

Building on SkillOpt

The paper builds directly on Microsoft’s SkillOpt framework, released in June, which reframed agent skills as trainable parameters rather than static model weights. SkillOpt demonstrated a 23.5-point average gain for GPT-5.5 across six benchmarks, establishing the theoretical foundation that skills could be optimized independently of the underlying model.

The new distillation method takes that concept and makes it radically more accessible. Because skill documents are just markdown files injected via system prompts, they work with existing rollout strategies—no custom inference infrastructure, no specialized hardware, and no retraining runs required.

For practitioners, the open questions are how durable distilled skills remain as underlying models are updated, and how the approach generalizes beyond the agentic domains tested—questions the benchmark results alone do not yet answer.