Reasoning models are AI models with extended thinking capabilities. Rather than producing an answer immediately, they work through a problem step by step — exploring approaches, checking their reasoning, and refining their thinking before producing a final response. The result is dramatically better accuracy on complex problems.
What makes reasoning models different
Standard models generate responses token by token in a single forward pass. Reasoning models have an additional “thinking” phase: an extended internal reasoning process that runs before the final answer is produced. This thinking process is often visible as a collapsed section in ZeroTwo — you can expand it to see how the model worked through the problem.
The tradeoff: reasoning models are slower and typically use more of your premium quota, but they’re significantly more accurate on hard problems where standard models stumble.
Available reasoning models in ZeroTwo
| Model | Provider | Notes |
|---|
| o3 | OpenAI | Advanced reasoning, among the most capable for math and logic |
| o4-mini | OpenAI | Faster and more efficient reasoning; strong balance of speed vs. depth |
| DeepSeek Reasoner | DeepSeek | Excellent for math, logical proofs, and structured analysis |
| Claude Sonnet / Opus with extended thinking | Anthropic | Configurable thinking depth via thinking level slider |
| Gemini 2.5 Pro | Google | Built-in reasoning; also multimodal |
Thinking levels
Some models — especially Claude — expose a thinking level slider in the ZeroTwo prompt bar when selected. This controls how much reasoning the model does before responding:
| Thinking level | Behavior | Best for |
|---|
| Low | Light reasoning, quick response | Problems that need some structured thinking but not exhaustive analysis |
| Medium | Balanced depth and speed | Most complex tasks — good default |
| High | Deep, extended reasoning; significantly slower | The hardest problems where accuracy is paramount |
Not all reasoning models expose a thinking level slider. For models like o3 and o4-mini, the reasoning depth is handled internally by the model. For Claude, the slider gives you explicit control.
When to use reasoning models
Reasoning models provide the most benefit for tasks that require careful, multi-step thinking:
Use a reasoning model for:
- Complex math problems: multi-step calculations, proofs, statistics, algebra
- Hard coding challenges: algorithm design, complex debugging, system architecture decisions, performance optimization
- Logical reasoning: puzzles, deductive inference, formal logic
- Research synthesis: pulling together insights from multiple sources or a long document into a coherent structured analysis
- Detailed analysis: when you want the model to carefully consider multiple angles before reaching a conclusion
- Planning tasks: anything where the right answer requires working through dependencies and ordering carefully
Examples:
"Prove that the sum of the first n odd numbers equals n².""Debug this recursive algorithm that's producing incorrect output for edge cases: [code]""Analyze this 50-page business strategy document and identify the three most significant risks, with supporting evidence from the text.""Design a database schema for a multi-tenant SaaS application with these requirements: [requirements]"
When NOT to use reasoning models
Reasoning models are overkill — and slower — for tasks that don’t require deep analysis:
Use a standard model for:
- Simple questions with straightforward answers
- Quick drafts and brainstorming
- Conversational exchanges
- Reformatting or transforming text
- Simple code snippets
- Summarizing short content
Standard models are faster, just as good for easy tasks, and don’t consume as much of your premium quota.
Tradeoffs at a glance
| Aspect | Standard models | Reasoning models |
|---|
| Response time | Seconds | Seconds to minutes |
| Quota cost | 1 premium message | 1+ premium messages |
| Simple tasks | Excellent | Overkill |
| Complex reasoning | Good | Excellent |
| Math and logic | Adequate | Significantly better |
| Multi-step coding | Good | Significantly better |
Start with a standard model. Switch if needed. For a new task, try a standard model like Claude Sonnet 4.6 or GPT-4o first. If the answer seems off, shallow, or the problem is genuinely complex, switch to a reasoning model and try again. This strategy saves quota without sacrificing quality when it matters.
Viewing the thinking process
When a reasoning model is working on a problem, ZeroTwo displays a “Thinking…” indicator. Once complete, you can expand a Thinking section in the response to see the model’s internal reasoning process — the steps it took, the approaches it considered, and how it arrived at its answer.
This is useful for:
- Verifying the model understood the problem correctly
- Learning from the model’s problem-solving approach
- Catching reasoning errors before acting on the conclusion
