Claude Sonnet 4.6 Just Dethroned Anthropic's Own Flagship Model
Here's the most shocking thing about Claude Sonnet 4.6: developers with early access prefer it over Opus 4.5, Anthropic's supposed flagship model that costs significantly more.
That's not supposed to happen. Sonnet models are the middle child—capable but not premium. Opus is where you go when you need the absolute best. Yet here we are, watching a $3 per million token model embarrass its expensive sibling.
The Million Token Flex
Sonnet 4.6 ships with a 1 million token context window in beta. That's entire codebases. Dozens of research papers. Your company's entire documentation set—all in a single conversation.
<> "Users can process entire codebases or dozens of research papers in a single request"/>
The context compaction feature automatically summarizes older parts of your conversation as you approach limits. It's like having an AI assistant with perfect memory management.
But here's what makes this interesting: they kept the same pricing as Sonnet 4.5. $3 for input, $15 for output per million tokens. While competitors are jacking up prices for advanced features, Anthropic is playing the long game.
Computer Use Gets Scary Good
The computer use improvements are where things get genuinely impressive. On OSWorld benchmark—which tests AI's ability to navigate real software like Chrome and VS Code—Sonnet 4.6 shows human-level capability for complex tasks.
We're talking about:
- Navigating complex spreadsheets
- Completing multi-step web forms
- Managing real software interfaces
Sure, it's not better than skilled humans yet. But "human-level" for routine computer tasks? That's the sound of jobs reshuffling.
What Nobody Is Talking About
The real story isn't the features—it's the developer preference data. When people with access to both models consistently choose the cheaper option, that reveals something profound about model architecture efficiency.
Either Anthropic accidentally made their mid-tier model better than their flagship, or they're deliberately cannibalizing Opus sales to grab market share. Both scenarios are fascinating.
The timing tells a story too:
- September 2025: Sonnet 4.5 released
- November 2025: Opus 4.5 released
- February 5, 2026: Opus 4.6 released
- February 17, 2026: Sonnet 4.6 released
Twelve days between releases? That's not coincidence—that's competitive maneuvering.
The Free Tier Trojan Horse
Anthropic made Sonnet 4.6 the default for free users, complete with file creation, connectors, and skills. This isn't generosity—it's strategy.
Get users hooked on advanced capabilities for free, then convert them when they hit limits. Meanwhile, paid users get a model that outperforms the previous flagship at the same price point.
Developer Platform Reality Check
The technical improvements are solid but not revolutionary:
- Web search tools now auto-generate code to filter results
- Adaptive thinking and extended reasoning in beta
- Better prompt injection resistance
What matters more is the 820 points and 714 comments this generated on Hacker News. That's serious developer attention.
The Anthropic Gambit
This release strategy feels deliberate. Make your middle-tier model so good that it threatens your premium offering. Force competitors to respond on both price and performance simultaneously.
It's working. When developers consistently choose Sonnet over Opus, you've fundamentally shifted the value equation in AI tooling.
The question isn't whether Sonnet 4.6 is good—it clearly is. The question is whether Anthropic just accidentally revealed that their model tiers were always more about pricing psychology than actual capability differences.
