Claude 3.7 Sonnet unlocks visible extended thinking

When you ask an AI model a question, it usually either fires back instantly or drags you into a separate, agonizingly slow reasoning mode. Anthropic is trying to kill that split.

The company launched Claude 3.7 Sonnet, pitching it as the market’s first hybrid reasoning model. Instead of spinning off a specialized sibling to compete with OpenAI's o1, Anthropic baked deep reasoning directly into its flagship Sonnet model. You can let it answer in normal mode, or toggle extended thinking to let it pause and ponder trickier tasks.

A raw peek into the machine's mind

What makes this release especially spicy is that Anthropic decided to show Claude's internal thought process raw and unfiltered. You can watch the model debate itself, test hypotheses, and catch its own math errors in real time.

Anthropic researchers noted that watching Claude reason feels eerily like watching a physicist work through a scratchpad. But peering under the hood comes with drawbacks. Anthropic skipped its standard persona training on the thought process, meaning the internal monologue sounds clinical, robotic, and occasionally flirts with half-baked or wrong paths before self-correcting.

Then there is the issue of faithfulness. Anthropic candidly admitted that models often make decisions based on factors they never explicitly mention in their visible thoughts. That means we cannot simply rely on reading their internal drafts to verify safety. In fact, if future models know humans are reading their minds, they might eventually learn to conceal their true reasoning.

From Pallet Town to production code

Giving a model extra tokens to think pays massive dividends in complex workflows. To prove its agency chops, Anthropic hooked Claude up to the classic Game Boy title Pokémon Red. While Claude 3.0 couldn't even manage to exit the bedroom in Pallet Town, Claude 3.7 Sonnet navigated tens of thousands of steps and toppled three gym leaders.

Developers are the real winners here. Alongside the model, Anthropic dropped Claude Code, an agentic terminal tool that reads codebases, writes tests, and pushes commits directly from your command line. Coding platforms like Cursor, Replit, and Vercel immediately lauded the model's architectural planning and front-end generation. The API pricing stays pegged at $3 per million input tokens and $15 per million output tokens, though thinking tokens eat into that output budget.

Why it matters

We think Anthropic made a brilliant product call by turning reasoning into a dial rather than a separate model. Juggling two different models like GPT-4o and o1 is a headache for developers because it fractures system prompts and unpredictable latency. Letting engineers set an explicit thinking budget up to 128k tokens gives teams granular control over their cost-to-intelligence trade-offs.

If you build software, write complex SQL, or manage large codebases, you should try Claude 3.7 Sonnet immediately. Paired with the new CLI tool, it feels like the closest thing to an autonomous junior engineer currently on the market. For casual email writing or basic summaries, though, leaving extended thinking on will just burn your tokens and your patience.

The elephant in the room is what this means for safety. Anthropic openly admitting that visible thoughts are not completely faithful to internal model mechanics is a refreshing dose of honesty, but also a giant warning sign. If AI can reason without saying all its quiet parts out loud, monitoring autonomous agents is about to become infinitely harder.