Claude Sonnet 4.6: The Coding Machine

Let that sink in. A Sonnet-class model — priced at $3/$15 per million tokens — is now the third-best coding model on the planet. It's beating GPT-5.2 High by 53 Elo points. It's beating Gemini 3.1 Pro Preview by 63 points. It costs roughly one-fifth what Opus charges.

Anthropic released Sonnet 4.6 on February 17th as the new default for Free and Pro users. Four days later, the arena data is clear: this isn't a marginal upgrade. Sonnet 4.6 doesn't just beat its predecessor — it outperforms every model that isn't named Opus 4.6 or Opus 4.5.

The Sonnet-Opus gap is shrinking

This is the real story. Historically, Anthropic's tier gap was massive. Claude Sonnet 4 (May 2025) scored 1390 on text while Opus 4 hit 1413 — a 23-point gap. Now Sonnet 4.6 (1524 on code) sits just 37 points behind Opus 4.6 (1561). The mid-tier is absorbing capabilities that were exclusive to the flagship three months ago.

For developers, this is the important signal. You probably don't need Opus. Sonnet 4.6 handles complex coding tasks — multi-file refactors, debugging gnarly race conditions, writing test suites — at a level that would have been state-of-the-art six months ago.

What makes it good at code?

Anthropic highlights three improvements in Sonnet 4.6:

1M token context window. This is the first Sonnet to match Opus's context length. For coding, this means feeding an entire codebase into a single prompt — not just the file you're editing, but the imports, tests, configs, and documentation. The model can reason about cross-cutting concerns that shorter contexts miss entirely.

Instruction following. Anthropic says developers with early access “often even prefer it to Opus 4.5.” That tracks with the arena data. When a model follows complex, multi-step coding instructions precisely, it wins blind comparisons — even against models with higher raw reasoning capacity.

Computer use. Sonnet 4.6 comes with improved computer use capabilities. This matters for AI coding agents that need to run tests, check build output, and iterate autonomously. Better tool use means fewer human interventions in agentic coding workflows.

Anthropic owns the code leaderboard

With Sonnet 4.6's debut, Anthropic now holds the top 4 positions on the Code leaderboard. Positions 1 through 4 are all Claude models. GPT-5.2 High sits at #5. Google's best code entry, Gemini 3.1 Pro Preview, is #7.

This is a remarkable position for a company that, a year ago, wasn't considered a top contender in coding. Anthropic's investment in Claude Code, their terminal-based coding agent, is clearly driving model development in the right direction. When your business depends on developers choosing your model, you optimize for developers.

What about Text?

Sonnet 4.6 also debuted at #13 on the Text leaderboard with 1457 Elo. Solid, but less dramatic than its code showing. The text leaderboard is more crowded and competitive — models from Anthropic, Google, Baidu, and Alibaba are all clustered within 10 Elo points in the 1449-1457 range.

The pattern is clear: Sonnet 4.6's relative strength is coding. And with Claude Code adoption growing and Anthropic positioning themselves as the developer-first AI company, that's exactly where it matters most.

The value proposition is insane

Let's talk pricing. Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens. Opus 4.6 costs $15/$75. GPT-5.2 High is in a similar premium tier. Gemini 3 Pro is comparable.

So for one-fifth the cost of Opus, you get a model that's only 37 Elo points behind on code. That translates to roughly a 45% win rate heads-up against Opus 4.6 — not parity, but close enough that for most coding tasks, the 5x savings makes Sonnet the rational choice.

The free tier upgrade is equally significant. Every Claude user who opens claude.ai now gets access to a model that outperforms GPT-5.2 High on code. For free. That's a powerful acquisition play.