How Do Hybrid Reasoning Models Work?

In February 2025, Anthropic released Claude 3.7 Sonnet - the first model they called "hybrid reasoning." The term caught my attention because it describes something genuinely different from what came before. Not just a better model. A model that can choose how hard to think.

I've been using it in production since launch. Here's how it actually works, and why it matters for anyone building with LLMs.

The two-speed brain

Every LLM you've used until now works the same way. You send a prompt, it generates tokens one by one, and out comes your answer. Same effort whether you asked for a haiku or a differential equation.

Hybrid reasoning changes that. Claude 3.7 Sonnet can operate in two modes:

Standard mode works like any other LLM. Fast responses, low cost, good enough for most tasks. Think of it as your brain on autopilot - recognizing patterns, producing familiar outputs without deep analysis.

Extended thinking mode is where things get interesting. Before answering, it produces a chain of internal reasoning tokens. It breaks down the problem, considers multiple approaches, checks its own logic, backtracks when something doesn't add up. Only then does it write the response you see.

This mirrors something cognitive scientists have studied for decades. Daniel Kahneman called them System 1 (fast, intuitive) and System 2 (slow, deliberate). This approach gives an LLM both systems in one package.

What happens in extended mode?

When you activate this mode, it starts by breaking your request into sub-problems. For a coding task, it identifies the algorithm needed, edge cases to handle, and the output format. Then instead of committing to the first approach, it considers alternatives. "I could solve this with recursion, but iteration would be clearer and avoid stack overflow on large inputs."

It also catches its own mistakes mid-process. I've seen it write "Wait, that's wrong because..." and then course-correct. This is visible in the thinking trace. After all that exploring and checking, it produces a final answer that reflects the full reasoning process.

The key detail: this internal thinking is billed. Each token costs money and takes time. Anthropic lets you set a budget (from a few hundred to 128,000 tokens) to control the quality-speed tradeoff.

When is it actually useful?

Not every task benefits from deeper analysis. Through my testing, I've found pretty clear patterns.

Where it shines:

Math and logic problems see the biggest improvement: tracking multiple constraints, checking intermediate steps, avoiding common traps. On SWE-bench (a coding benchmark), Claude 3.7 Sonnet in extended mode hit 70.3%, a significant jump.

Complex code generation is another clear win. When I'm asking for multi-file implementations or refactoring with subtle requirements, extended mode catches things that standard misses. It's the difference between a developer who codes from memory and one who thinks through the architecture first.

Structured analysis too: comparing options, evaluating tradeoffs, producing recommendations with reasoning. It takes time to weigh factors rather than jumping to a conclusion.

Where it's wasted:

Simple factual questions. "What's the capital of France?" doesn't need 10,000 tokens of deliberation.

Creative writing, surprisingly. Poems, stories, marketing copy - these tend to come out just as good (sometimes better) in standard mode. Over-thinking can make creative output feel stiff.

Translation and summarization. These are pattern-matching exercises where intuition is already strong. Extended mode adds latency without meaningful quality improvement.

The budget question

Here's what makes this practical rather than theoretical: you pick how much depth to allow.

At the low end (1,000-5,000 tokens), you get quick sanity checks. It pauses briefly, catches obvious errors, then responds. Good for everyday questions that might have subtle gotchas. The sweet spot for most technical work is around 10,000-30,000 tokens - enough room to explore alternatives and self-correct without burning through your credits. And for genuinely hard problems (multi-step math proofs, complex system design, debugging tricky code), you can go up to 128,000 tokens. I use that maybe 5% of the time.

The ability to control the budget matters because it puts the cost-quality tradeoff in your hands. You're not locked into paying for deep analysis on every request.

Why this changes how I build with LLMs

Before, I had two options when a problem needed careful analysis: a weaker model, fast and cheap, or a stronger one, slow and expensive. Both applied the same level of effort regardless of the prompt.

Now I can route different parts of a workflow through the same model with different depth levels. Quick classification? Standard mode. Complex decision that affects the rest of the pipeline? Extended mode with a generous budget.

This is particularly relevant for LLM integration projects where you're building systems that handle diverse inputs. Instead of overprovisioning everywhere "just in case," you can match the analysis depth to the actual difficulty of each request.

Where is the industry heading?

Claude 3.7 Sonnet shipped this approach in February 2025, and the industry followed fast. Extended thinking is now standard across the Claude 4.x family. Google added thinking capabilities to Gemini. OpenAI has chain-of-thought in o1 and o3. Qwen 3 brought hybrid reasoning to open-source models with a "thinking mode" toggle. The industry has fully converged on this pattern.

The practical takeaway is simple. If you're building systems that handle a mix of simple and complex questions, hybrid reasoning lets you stop choosing between fast-and-cheap and slow-and-good. Both in the same model, adjustable per request.

What's next?

Zach Galifianakis doing math

To see how these models fit into autonomous systems: what is an AI agent? And to deploy them on your own infrastructure, check out the self-hosting guide.