Sonnet 4.6 Defaults Free/Pro — Near-Frontier on the Path Most People Actually Run
Anthropic launched Claude Sonnet 4.6 (Feb 17) as default Free/Pro on Claude.ai and Cowork; $3/$15 MTok same as 4.5; 1M context beta. Early Claude Code preference ~70% vs 4.5 and 59% vs Opus 4.5—vendor. Compresses gap between best model and default UX twelve days after Opus 4.6.
Twelve days after Opus 4.6, Anthropic’s proprietary GTM move: put near-frontier capability on the default Free/Pro path—not only Max/API power users. That compresses the gap between “best model” and “what most people actually run.”
Anthropic released Claude Sonnet 4.6 on February 17 as the most capable Sonnet yet and the default on Free and Pro in claude.ai and Claude Cowork. Pricing: $3/$15 per million tokens (same as Sonnet 4.5). Context: 1M token window in beta. Full upgrade pitch across coding, computer use, long-context reasoning, agent planning, knowledge work, design. API id claude-sonnet-4-6; adaptive/extended thinking; context compaction (beta); web search/fetch improvements; several tools GA. Free tier upgraded with file creation, connectors, skills, compaction. Safety: system card; researchers described broadly warm/prosocial character, strong safety behaviors; improved prompt-injection resistance vs 4.5, similar to Opus 4.6. Computer use: continued OSWorld gains; early users reported near human-level on complex spreadsheets/multi-step forms (still lagging top human operators). Early Claude Code testers preferred 4.6 over 4.5 ~70% and over Opus 4.5 59% (Anthropic)—less overengineering, better instruction following. Partners (Databricks, Replit, Cursor, GitHub, Cognition, others) highlighted near-Opus document/enterprise workloads at Sonnet cost.
Claims vs checks
Default status, pricing, and 1M beta are Anthropic primary. Preference percentages and partner quotes are vendor. Independent Elo/AA boards: verify when listed—do not treat 70%/59% as public Arena. OSWorld and computer-use “near human” claims remain lab/early-user.
Limits
- Free/Pro default ≠ Max rate limits or Opus ceiling.
- 1M context in beta: quality and pricing at long horizon may differ from marketed ideal.
- Preference tests are internal early-adopter samples.
Sources
- Anthropic: “Introducing Claude Sonnet 4.6” (February 17, 2026). Primary.
- Claude Sonnet 4.6 System Card (February 17, 2026).
- Snowflake Cortex same-day availability note (February 17, 2026).