Anthropic drops Sonnet 4.6 with Opus-level coding

Anthropic just gave developers fewer reasons to pay for its most expensive model.

Two weeks after rolling out Opus 4.6, the company dropped Claude Sonnet 4.6, sticking to its brisk four-month release cycle. The new mid-tier powerhouse is now the default model on claude.ai for both Free and Pro subscribers, with API pricing unchanged at $3 per million input tokens and $15 per million output tokens.

The flagship killer inside

The headline numbers here aren't just synthetic benchmarks; they are human preferences. In Anthropic's Claude Code testing, developers favored Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time.

Even more telling: users picked Sonnet 4.6 over the flagship Opus 4.5 from November 59% of the time. Early testers reported far less overengineering and "laziness," praising the model for actually reading surrounding context before touching code and consolidating shared logic rather than duplicating it.

On the hardware and reasoning side, the beta context window expands to 1 million tokens, double the previous Sonnet limit. It notched a 60.4% on the ARC-AGI-2 benchmark. While it still trails apex systems like Opus 4.6, Gemini 3 Deep Think, and specialized variants of GPT 5.2, it comfortably leads its weight class.

In the business simulation benchmark Vending-Bench Arena, Sonnet 4.6 pulled off a fascinating long-horizon strategy: it aggressively outspent competitors on capacity during the first ten simulated months, before pivoting sharply to cash generation in the home stretch to secure a win.

Better mouse clicks and Excel superpowers

Sixteen months after Anthropic pioneered general-purpose computer use, Sonnet 4.6 shows tangible maturity in handling spreadsheets and multi-step web forms across tabs. Just as crucially, its resistance to prompt injection attacks now rivals Opus 4.6.

Platform features like adaptive and extended thinking are baked in, along with beta context compaction to summarize older dialogue automatically. Over in Claude for Excel, Pro and business tiers can now use MCP connectors to pull live data directly from S&P Global, PitchBook, and FactSet without leaving the workbook.

Why it matters

In our view, Sonnet 4.6 changes the calculus for practical AI adoption. You simply don't need to splurge on heavyweight frontier tiers for everyday software engineering or complex financial modeling anymore; getting Opus-grade execution at $3/$15 per million tokens makes real-world deployment economically viable for bootstrapped startups and individual developers right now.

It also ratchets up intense pressure on OpenAI and Google. By making its everyday workhorse fast, obedient, and largely immune to false completion claims, Anthropic is winning the developer goodwill game where daily friction matters more than theoretical ceiling performance.

The trade-off, of course, is that full-screen computer use remains inherently brittle compared to native API integrations when enterprise legacy software misbehaves. Anthropic still recommends Opus 4.6 for massive codebase overhauls and multi-agent systems, but for the vast majority of daily tasks, Sonnet 4.6 is now the smartest dollar-for-dollar pick on the market.