Reasoning Models

AI That Thinks Before It Answers

Reasoning models represent a qualitative shift in how AI approaches hard problems. Rather than generating a response immediately, reasoning models spend compute "thinking" — producing an explicit chain-of-thought scratchpad before committing to a final answer. OpenAI's o-series, Anthropic's extended thinking, and DeepSeek-R1 were the pioneering examples — RL-trained reasoning has since become a standard capability across every major vendor's lineup. The result is dramatically better performance on tasks requiring multi-step logic, mathematics, code debugging, and structured analysis.

• Standard LLMs: generate output tokens autoregressively, one token at a time, without deliberation
• Reasoning models: generate a "thinking" scratchpad first, then the final response — internal deliberation made visible
• Training approach: process reward models and reinforcement learning on reasoning traces, not just outcome correctness
• Strengths: mathematics, competitive coding, multi-step planning, legal analysis, complex debugging
• Limitations: 10–50× more tokens per response means meaningfully higher cost — not the right tool for simple tasks

When to Use Reasoning Models

Reasoning models are not universally better than standard LLMs — they are a specialised tool for tasks where deliberate multi-step thinking produces meaningfully better outcomes. The cost-to-benefit tradeoff depends entirely on the task complexity.

• Right for reasoning models: security vulnerability analysis, contract review, architectural planning, competitive maths, complex code audit
• Wrong for reasoning models: simple Q&A, summarisation, classification, routine drafting — standard models are cheaper and just as good
• The test: would a smart human benefit from thinking carefully for 30 seconds before answering? If yes, a reasoning model probably helps
• Cost signal: if you are spending 10× more tokens for a 5% quality improvement on a task, you have the wrong tool
• Hybrid pattern: use a fast cheap model for triage and routing, a reasoning model only for the subset of cases that genuinely need it

What Reasoning Models Change for Practitioners

Reasoning models shift the frontier of what can be automated. Tasks that previously required senior expert judgment — because they involved multi-step reasoning under ambiguity — can now be partially delegated to AI. This changes hiring decisions, workflow design, and the competitive baseline for knowledge work.

• The "junior analyst" uplift: reasoning models can do tasks that previously required experienced practitioners — the quality gap narrows significantly
• Security: complex threat analysis, attribution reasoning, and multi-step forensic investigation are all improved by reasoning models
• Legal and compliance: contract review, regulatory gap analysis, policy interpretation — areas where multi-step logic matters most
• Engineering: architecture review, security code auditing, debugging complex distributed system failures
• The new benchmark: if a reasoning model can handle 70% of a senior analyst's hard cases, the workflow changes — not just the tools

Try It Yourself

The "would a smart human benefit from thinking for 30 seconds" test only becomes a habit once you have run it on a problem of your own. Run it now, on something real.

I need help with a multi-step problem: [describe your situation].
Constraints: [what must be true for a solution to work]
A good answer would: [what would make it genuinely useful]
Walk me through your reasoning, not just the conclusion.