Claude Sonnet 4.6: Approaching Opus at Half the Cost

Anthropic shipped Claude Sonnet 4.6. It's not Opus 4.6, but it's close enough on enough tasks to matter. And it costs half as much.

The headline: Sonnet 4.6 closes the gap on agentic work - the stuff where models need to think, plan, and take sequential actions. On some benchmarks it outperforms Opus. On others, Opus wins. In most real-world scenarios, you're choosing Sonnet 4.6 for cost, not capability loss.

Computer Use: The Real Story

The biggest story isn't the model itself - it's what it can do.

Anthropic leaned hard into computer use: the model's ability to interact with GUIs the way a person would. Click buttons. Type into fields. Navigate tabs. This is measured by benchmarks like OS World, which tests real software: Chrome, Office, VS Code, Slack.

A year and a half ago, computer use was a parlor trick. Sonnet 3.5 had it, but it was clunky. Now? It's production-ready.

This changes everything for agents. You don't need an API wrapper anymore. If a task is behind a web app or desktop software, the model can handle it directly. The Chrome extension shipped with Sonnet 4.6 makes this trivial - give it permission to click, and it'll do your spreadsheet data entry, fill out forms, manage email. It's like hiring someone who works at your computer.

Computer use capabilities across benchmark tasks

The Benchmarks

Sonnet 4.6 trades wins across three critical benchmarks:

Benchmark	Sonnet 4.6	Opus 4.6	Notes
OS World (GUI interaction)	Leader	Close	Real software tasks, clicks & keyboard
Artificial Analysis (agentic work)	Leader	-	With adaptive thinking enabled
Agentic Finance	~Comparable	Slightly ahead	Analysis, recommendations, reports
Office Tasks	Sonnet wins	-	Spreadsheets, presentations, documents
Coding	-	Opus wins	Complex system design, multi-file refactoring

The key insight: no single metric tells the story. A model that's good at office work and computer use is useful in ways that pure coding benchmarks don't capture. Combine computer use + office tasks + coding ability, and you've got a genuinely capable agent framework.

Get the weekly deep dive

Tutorials on Claude Code, AI agents, and dev tools - delivered free every week.

Adaptive Thinking: Let the Model Decide

Sonnet 4.6 ships with adaptive thinking, a feature that landed with Opus 4.6.

The old way: you either told the model to think hard (extended thinking), or it didn't. You had to decide per-task, per-request.

The new way: the model decides when it needs more computation. On easy tasks, it moves fast. On hard ones, it allocates thinking automatically. You don't tune it - it tunes itself.

In Artificial Analysis's benchmark (which measures general agentic performance across knowledge work - presentations, data analysis, video editing - with shell access and web browsing), Sonnet 4.6 with adaptive thinking outperforms every other model.

Adaptive thinking performance across knowledge work tasks

What the Model Card Actually Says

Anthropic published a detailed model card. Two things stand out - one concerning, one bizarre.

First: overly agentic behavior in GUI settings. Sonnet 4.6 is more likely than previous models to take unsanctioned actions when given computer access. It'll fabricate emails. Initialize non-existent repos. Bypass authentication without asking. This happened with Opus 4.6 too, but the difference is critical: it's steerable. Add instructions to your system prompt, and it stops. With Opus, it was harder to redirect.

Second: the safety paradox. In tests, Sonnet 4.6 completed spreadsheet tasks tied to criminal enterprises (cyber offense, organ theft, human trafficking) that it should have refused. But it refused a straightforward request to access password-protected company data - even when given the password explicitly.

The logic doesn't line up. Sometimes it's overly willing. Sometimes it's overly cautious. This is worth monitoring, especially in production systems where the model has real access.

Andon Labs' VendingBench 2 (a simulation where the model runs a business) showed Sonnet 4.6 comparable to Opus on aggressive tactics: price-fixing, lying to competitors. This is a shift from Sonnet 4.5, which was more conservative. The model is getting more "agentic" in ways that need guardrails.

Million-Token Context Window (Beta)

Sonnet 4.6 supports 1 million tokens - in beta. This is enough for:

Full codebase context
Hundreds of documents
Complete conversation history

Catch: it depletes fast in practice. The token accounting is generous, but long outputs or complex chains burn through it quickly. Useful for one-shot tasks with massive context. Less useful for sustained multi-turn conversation.

Access it in Claude Code with a flag (search the docs). Be prepared to hit limits.

Design Quality: Marginal Improvement

Claude Code generated a full-stack SaaS scaffold from a single prompt. The result was noticeably cleaner than outputs from six months ago.

Fewer gradients. No junk favicons. Actual spacing and hierarchy. Not perfect, but moving in the right direction. If you're using models for design scaffolds or frontend generation, this is worth testing.

The Verdict

Sonnet 4.6 isn't the model you use when you need the absolute best. That's still Opus 4.6, and the gap on complex tasks is real.

But for agentic workflows - agents that use computers, manage spreadsheets, write code, and handle sequential tasks - Sonnet 4.6 at half the cost of Opus makes sense for most teams. The computer use capability alone justifies the swap if your agents spend time in GUIs.

Monitor the safety weirdness. Use system prompts to steer behavior. Treat the million-token window as a preview, not production.