Claude Sonnet 4.6: Opus-Level AI at Sonnet Price

TL;DR

Anthropic released Claude Sonnet 4.6 on February 17, 2026. The key takeaway:

• 79.6% SWE-bench — near-identical to Opus 4.6 (80.8%) on real-world coding
• 72.5% OSWorld — essentially tied with Opus 4.6 (72.7%) on computer use, nearly double GPT-5.2 (38.2%)
• $3/$15 per million tokens — unchanged from Sonnet 4.5, 5x cheaper than Opus
• 1M token context window (beta) — up from 200K
• Now the default model for all Free and Pro Claude users

What Anthropic Announced

Claude Sonnet 4.6 is Anthropic's second major model release in under two weeks (following Opus 4.6 on February 6). In their blog post, Anthropic describes it as "a full upgrade of the model's skills across coding, computer use, long-context reasoning, agent planning, knowledge work, and design."

The core claim: "Performance that would have previously required reaching for an Opus-class model — including on real world, economically valuable office tasks — is now available with Sonnet 4.6."

This is a significant statement. Anthropic is effectively saying: for most production workloads, you no longer need to pay for Opus.

Full Benchmark Breakdown

Where Sonnet 4.6 Matches or Beats Opus

Sonnet 4.6 actually leads on office tasks and financial analysis — two economically significant categories.

Where Opus 4.6 Retains the Lead

The pattern is clear: Opus wins on tasks that require the deepest, most novel reasoning — codebase-scale refactoring, multi-step research, and problems the model hasn't seen before. Sonnet wins on speed-sensitive, production-ready tasks.

Computer Use: The Standout Improvement

The computer use numbers deserve special attention:

In 16 months, Sonnet went from 14.9% to 72.5% on computer use — a 4.9x improvement. Jamie Cuffe, CEO of Pace (an insurance tech company), reported that Sonnet 4.6 hit 94% on their internal computer use benchmark: "It reasons through failures and self-corrects in ways we haven't seen before."

What's New vs. Sonnet 4.5

1. 1M Token Context Window (Beta)

The context window expands from 200K to 1 million tokens. This means entire codebases, lengthy legal documents, or hours of conversation history fit within a single prompt.

A new context compaction feature (also in beta) auto-summarizes older conversation segments, effectively extending usable context even further.

2. Better Instruction Following, Fewer Hallucinations

This is what developers noticed first. In Claude Code testing:

• 70% preferred Sonnet 4.6 over Sonnet 4.5
• 59% preferred it even over Opus 4.5 (the November 2025 frontier model)

• Reads existing code before modifying it (instead of guessing)
• Consolidates logic instead of duplicating it
• Fewer false claims of success ("I've fixed the bug" when it hasn't)
• Less over-engineering — doesn't add unnecessary abstractions
• Better follow-through on multi-step tasks

3. Computer Use Goes Production-Ready

The jump from 61.4% to 72.5% on OSWorld crosses a threshold. Users describe "human-level capability in tasks like navigating complex spreadsheets or filling out multi-step web forms."

Sonnet 4.6 also improved significantly on prompt injection resistance for computer use — performing at Opus 4.6 levels. This is critical for any agent that browses the web or processes untrusted input.

4. Extended Thinking + Adaptive Thinking

Both are supported, letting the model allocate more computation to harder problems. But notably, Sonnet 4.6 performs strongly even without extended thinking enabled — the base model is fundamentally better.

5. Free Tier Upgrade

Free Claude users now get Sonnet 4.6 as default, plus:

• File creation capabilities


• Connectors (integrations with external data)


• Skills (reusable instructions)


• Context compaction

6. MCP Connectors in Excel

Claude in Excel now supports connectors for S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet — pulling live financial data directly into spreadsheets.

Pricing

No price change from Sonnet 4.5:

For comparison:

• Opus 4.6 API: $15/$75 per million tokens (5x more)


• GPT-5.2 API: $5/$15 per million tokens (1.7x more input)


• Gemini 3 Pro API: $7/$21 per million tokens (2.3x more input)

Cost Per Claude Code Session

For a typical coding session (100K input + 20K output tokens):

A team running 100 agent sessions/day saves ~$240/day by using Sonnet 4.6 instead of Opus.

How to Access

claude.ai

Claude Code

claude  # Sonnet 4.6 is now the default
claude --model claude-sonnet-4-6-20250217  # explicit selection

API

import anthropic

client = anthropic.Anthropic()
message = client.messages.create(
    model="claude-sonnet-4-6-20250217",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Your prompt here"}]
)

Cloud Platforms

Industry Context

Sonnet 4.6 is Anthropic's second major release in 11 days (after Opus 4.6 on February 6). CNBC described the pace as "continuing breakneck speed of AI model releases." VentureBeat called it "a seismic repricing event for the AI industry."

The broader trend: the performance floor is rising. What required a $15/$75 flagship model six months ago now ships at $3/$15. For AI product builders, this means:

• AI features cost 5x less to run
• Computer use agents are economically viable at scale
• The model is no longer the bottleneck — shipping is

• Anthropic: Introducing Claude Sonnet 4.6
• CNBC: Anthropic releases Claude Sonnet 4.6
• VentureBeat: Sonnet 4.6 matches flagship at one-fifth the cost
• 9to5Mac: Claude Sonnet 4.6 improved coding skills
• MacRumors: Claude Sonnet 4.6 improved coding, computer use
• MarkTechPost: Claude 4.6 Sonnet with 1M token context
• OfficeChai: Claude Sonnet 4.6 Benchmarks
• SiliconANGLE: Anthropic debuts Sonnet 4.6