Sonnet Is the New Opus: Why Mid-Tier Models Keep Eating the Premium Tier

Anthropic shipped Claude Sonnet 4.6 yesterday. Within hours it hit 1,200+ points on Hacker News and became the default model on claude.ai. That alone isn't the story. The story is this line from the announcement:

Developers with early access prefer Sonnet 4.6 to its predecessor by a wide margin. They often even prefer it to our smartest model from November 2025, Claude Opus 4.5.

Read that again. The $3/million-token model is beating the $15/million-token model in developer preference. Not on some narrow benchmark. In daily, real-world use.

This has happened before. It keeps happening. And if you're building on top of LLMs, it should change how you architect your systems.

The pattern nobody plans for

Every major AI lab follows the same playbook. Ship a frontier model at premium pricing. Six months later, ship a mid-tier model that captures 80-90% of the capability at 20-30% of the cost. Repeat.

GPT-4 gave way to GPT-4 Turbo. GPT-4 Turbo gave way to GPT-4o. Claude 3 Opus gave way to Claude 3.5 Sonnet. Now Opus 4.5 is giving way to Sonnet 4.6.

The gap between tiers keeps shrinking. Not because the premium models are getting worse. They're still improving. But the mid-tier models are improving faster. They benefit from everything learned building the bigger model, plus months of additional optimization, distillation, and infrastructure work.

If you hardcode model selections in your production systems, you're leaving money on the table every quarter.

What actually changed in Sonnet 4.6

Three things matter here beyond the usual benchmark improvements.

1M token context window (beta). This is the quiet big deal. Sonnet 4.5 topped out at 200K. Going to 1M at the Sonnet price point means entire codebases, full document sets, and multi-hour conversation histories fit in a single call. The use cases this unlocks at $3/M input tokens are fundamentally different from what's viable at $15/M.

Computer use that actually works. Anthropic's been shipping computer use since October 2024. Back then it was a demo. Now it's scoring high enough on OSWorld that early users report "human-level capability" on tasks like navigating complex spreadsheets and filling out multi-step web forms. The jump from Sonnet 4.5 to 4.6 on this benchmark is significant. More importantly, prompt injection resistance improved substantially, which was the main reason you couldn't deploy computer use in production before.

Coding consistency. This is the one developers actually feel. Sonnet 4.5 was good at coding but inconsistent. It'd nail a complex refactor and then fumble a simple function signature. Sonnet 4.6 reportedly smooths that out. Instruction following got tighter. The model does what you ask more reliably, which matters more than raw capability when you're using it as a coding assistant eight hours a day.

The infrastructure lesson

Here's what I keep telling teams I work with: don't pick a model. Pick a routing strategy.

The smart play in February 2026 looks like this:

• Default tier: Sonnet 4.6 for 90% of your traffic. Coding, summarization, classification, chat, document analysis. It handles all of it well enough that most users won't notice the difference from Opus.
• Escalation tier: Opus 4.6 for the remaining 10%. Complex multi-step reasoning, novel problem-solving, tasks where you've measured a meaningful quality gap.
• Speed tier: Haiku for latency-sensitive paths. Autocomplete, inline suggestions, real-time classification.

The routing logic doesn't have to be fancy. Start with task type. Measure quality. Adjust thresholds. The savings compound fast. A team doing 100M tokens/month saves $1.2M annually by moving from all-Opus to a routed setup. That's not a rounding error.

What this means for the next 12 months

If you're building on LLMs, plan for the mid-tier to be your default. Plan for the premium tier to be your exception. And plan for both to get cheaper and better every quarter.

The developers who win aren't the ones using the most powerful model. They're the ones who built their systems to swap models without rewriting their applications. Abstractions matter. Evals matter. Routing matters.

Sonnet 4.6 isn't just a model upgrade. It's further evidence that the mid-tier is where the action is. It has been for a while now.

The frontier models break new ground. The mid-tier models are where that ground gets paved into a road you can actually drive on.