Claude Sonnet 5: Opus-Level Agents at Sonnet Pricing

Quick answer (AEO): Claude Sonnet 5 launched June 30, 2026 as Anthropic’s most agentic Sonnet model. It supports a 1M-token context window, 128K max output tokens, adaptive thinking, and autonomous tool use (browsers, terminals, code execution) at a level previously requiring Opus-class models. Introductory pricing: $2/M input, $10/M output through August 31, 2026, then $3/$15. It succeeded Sonnet 4.6 (February 2026) and is now the default model for all free and pro Claude users.

What “most agentic Sonnet” means in practice

Sonnet 5 can:

• Make plans — decompose complex goals into multi-step execution strategies.
• Use tools autonomously — browsers, terminals, file systems, APIs — without step-by-step human approval.
• Run extended sessions — maintain coherent behavior across long workflows that previously only Opus handled reliably.
• Adaptive thinking — dynamically allocates reasoning depth based on problem complexity.

The key shift: tasks that required Opus 4.8 ($15/$75 per million tokens) now work reliably on Sonnet 5 at $2/$10 — a 7-8x cost reduction for agentic workloads.

Technical specifications

Where Sonnet 5 sits in the Claude hierarchy

Anthropic’s mid-2026 model lineup:

1. Claude Fable 5 — Mythos-class flagship. Highest capability. $10/$50. Subject to export controls.
2. Claude Opus 4.8 — #1 Intelligence Index (61). 88.6% SWE-bench. $15/$75. Premium workloads.
3. Claude Sonnet 5 — Near-Opus agentic capability. $2-3/$10-15. Default for most users.
4. Claude Haiku — Fast, cheap. High-throughput extraction and simple tasks.

Sonnet 5 is positioned as the workhorse model — good enough for most agentic tasks, cheap enough for production scale, with 1M context for long sessions.

The pricing strategy

The introductory pricing ($2/$10 through August) is clearly designed to capture market share during the GPT-5.6 Sol launch window. At $2/M input, Sonnet 5 is:

• Cheaper than Muse Spark 1.1 ($1.25 input but $4.25 output vs. Sonnet’s $2/$10)
• Dramatically cheaper than GPT-5.6 Terra for comparable capability
• 8x cheaper than Opus 4.8 for tasks where Sonnet 5 quality is sufficient

For teams building multi-agent systems, this pricing makes high-throughput agent loops financially viable where they weren’t before.

Practical implications for builders

Migration from Opus

If you’re using Opus 4.8 for agent workloads, test Sonnet 5 on your eval suite. For many tasks — code generation, tool use, planning, research — Sonnet 5 performs at near-Opus quality. Reserve Opus for:

• Maximum SWE-bench performance (critical bug fixes).
• Tasks requiring the absolute highest reasoning depth.
• Safety-critical decisions where you need peak judgment.

Multi-agent cost optimization

The classic pattern: use a frontier model (Opus/Sol) as the supervisor and cheaper models as workers. Sonnet 5 changes this — it’s good enough to be both supervisor and worker for many workflows, simplifying architecture:

# Before Sonnet 5: complex multi-tier routing
supervisor = "claude-opus-4.8"      # $15/$75
worker = "claude-haiku"              # $0.25/$1.25

# After Sonnet 5: single-tier for most workflows
agent = "claude-sonnet-5"           # $2/$10
# Only escalate to Opus for genuinely hard reasoning
escalation = "claude-opus-4.8"      # Still available when needed

1M context window

The full 1M-token context (same as Opus) means Sonnet 5 handles:

• Large codebase analysis without chunking.
• Long conversation histories without summarization.
• Full document processing for RAG-heavy workflows.

What’s not included

• Priority Tier — not available on Sonnet 5. High-traffic enterprise users may hit rate limits.
• Fable-class capability — for the absolute frontier (whatever Fable 5 offers beyond Opus), Sonnet 5 is not a substitute.
• Guaranteed availability — the export control situation showed that even Claude models can go offline. Build fallback chains.

The competitive impact

Sonnet 5’s launch at $2/$10 creates pricing pressure across the entire market:

• GPT-5.6 Terra must justify its price premium over Sonnet 5.
• Muse Spark 1.1 ($1.25/$4.25) is cheaper on input but comparable on output — the gap narrows.
• Open-weight models (DeepSeek V4 Pro, Qwen) still win on self-hosting and privacy, but Sonnet 5 reduces the cost argument for going open.

For most engineering teams building AI products in July 2026, Claude Sonnet 5 is the default choice — good enough quality, affordable pricing, 1M context, and Anthropic’s safety track record.

Related reading: Frontier model landscape mid-2026, MCP ecosystem, and context engineering.