Claude 3.7 Sonnet and Claude Code: Hybrid Reasoning for Real Work

Most people use one AI setting for everything. Fast and fine, all day.

That works until you hand it a task that needs real thought, like a pricing model, a messy plan or a script that has to run without breaking. Then you want a model that slows down and works the problem. And you do not want to pay that price for every email.

On February 24, 2025, Anthropic shipped a model built around that exact tension. The useful idea is a dial for thinking. Here is what it is, what it costs and who should care.

What did Anthropic release on February 24, 2025?

Two things: a model and a preview of a tool.

Claude 3.7 Sonnet

Anthropic announced Claude 3.7 Sonnet as the "first hybrid reasoning model on the market." It offers standard mode and extended thinking mode. In the API, users can set a thinking budget, capping how many tokens the model spends thinking, up to the 128K output limit. It is available on all Claude plans, though extended thinking is not on the free tier, along with the Claude developer platform, Amazon Bedrock and Google Cloud's Vertex AI.¹

On coding, Anthropic reports 70.3% on SWE-bench Verified with high compute and 63.7% without the extra scaffold, on a subset of 489 problems that were solvable on its infrastructure.¹ Those are Anthropic's numbers on a software benchmark, so treat them as a signal about coding ability, not a promise about your business tasks.

Claude Code

Anthropic also introduced Claude Code, an agentic coding tool that runs in the command line, as a limited research preview. It lets developers delegate engineering tasks from the terminal. Anthropic reports that in early testing it completed some tasks in a single pass that would normally take 45 minutes or more of manual work.¹

What does hybrid reasoning mean, and when should you turn thinking on?

It means one model can behave two ways, and you choose. That is a real change for anyone who has had to pick between a fast model and a thoughtful one, the split I first covered in OpenAI o1 and reasoning models.

Standard mode is for most of your day

Drafting an email, rewording a headline, summarizing a call, answering a quick question. These do not need extended thinking, and paying for it is wasteful.

Extended thinking is for tasks with steps

Planning a launch with dependencies, finding the hole in a pricing model, debugging a stubborn spreadsheet formula, comparing several offers against several constraints. The model works through the problem before it answers, and you can set a budget for how much.

What does it cost, and how do you keep it in check?

Anthropic lists the price at $3 per million input tokens and $15 per million output tokens, and says that includes thinking tokens.¹ That means the longer the model thinks, the more output tokens you are billed for.

Cap the thinking budget

In the API you can limit tokens spent thinking. If you use it through a chat plan, the cost is the subscription, but the same principle holds: thinking takes time, and you should reserve it for tasks that earn it.

Measure per task, not per token

Token prices are abstract, and they keep falling, as I wrote when DeepSeek R1 landed. What matters is what a task costs and what it saves. If a long think replaces an hour of your time on a pricing review, it is cheap. If it replaces thirty seconds of work, it is not.

What can you ask a reasoning model to do for a small business?

Think of the work where you would normally want a smart second pair of eyes.

• Stress test an offer. Paste your promise, price and audience, and ask where a skeptical buyer would push back.
• Audit an email sequence. Ask it to find gaps, repeated points and missing next steps.
• Draft a process document. Describe how you do something and ask for a first-pass checklist you can correct.
• Check a spreadsheet's logic. Ask it to walk through your formulas and flag assumptions.

In every case, you supply the facts and the judgment. The model supplies speed and a fresh angle.

Is Claude Code for you?

Maybe not yet, and that is fine.

Who it is built for

It is a command-line tool for developers, in a limited research preview.¹ If you do not write or manage code, you will not use it directly. If you work with a developer who builds your internal tools, it is worth knowing it exists.

Who should be curious anyway

Operators who build small internal tools, like a script that cleans an export or a dashboard that pulls numbers together, may see this as a sign of where tooling is going, a thread that started for me with Claude computer use. But a preview is a preview. Anthropic calls it research, and a research preview is where things change and break.

How do people actually use Claude?

Anthropic's Economic Index, published February 10, 2025, analyzed about one million anonymized conversations on Claude.ai. Software development was the largest category at 37.2%, and writing and editing in the arts, design and media group came in at 10.3%.²

The part that matters for operators

Most use was augmentation, where a person works with the model, not automation, where the model does the task alone. Anthropic notes the data comes from Free and Pro conversations and reflects share of conversations, not all AI use.² Still, it matches how I think about this. The strongest results come from augmentation, not autopilot (the same line I draw in agents vs workflows): a person who knows the work directs the model and checks the output.

How should you adopt it without a mess?

Any tool that changes what one person can produce gets my attention, because one person is my whole staff, which shapes the lean systems I build for clients. But the winners here will not be the people who chased the newest model. They will be the ones who know their work well enough to tell when the output is good.

Keep building.Ekai