Claude Sonnet 5 Lands, Closing In on Opus 4.8

Anthropic released Claude Sonnet 5 on Tuesday, June 30, calling it the most agentic Sonnet model it has built and pitching it as a markedly cheaper way to run autonomous AI agents. The model can make multi-step plans, drive tools like browsers and terminals, and run on its own "at a level that, just a few months ago, required larger and more expensive models," the company said — landing exactly the release that swirling rumors had teased for weeks.

The benchmarks back the framing. On agentic coding, measured by SWE-bench Pro, Sonnet 5 scores 63.2% — up from Sonnet 4.6's 58.1% and closing most of the gap to Claude Opus 4.8, which still leads at 69.2%. That puts Sonnet 5 ahead of OpenAI's GPT-5.5 (58.6%) and Google's Gemini 3.5 Flash (55.1%) on the same test, though GPT-5.5 still edges it on Terminal-Bench 2.1 (83.4% to 80.4%). On a knowledge-work benchmark, Anthropic says the mid-tier model even slightly outperforms Opus 4.8 — and developers can dial an adjustable "effort" level to trade cost against depth of reasoning.

Price is the real headline. Sonnet 5 launches with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, after which it settles at $3 and $15. That undercuts Opus 4.8, GPT-5.5 and Gemini 3.1 Pro, while remaining costlier than the bargain-tier Gemini 3.5 Flash. Anthropic's pitch is blunt about where the market has moved: agentic capability is now table stakes, and the contest is increasingly about how cheaply and reliably a model can do the work without a human watching over it.

Availability is broad from day one. Sonnet 5 is the new default model on the Free and Pro tiers of claude.ai and is available to Max, Team and Enterprise users. It is live in Claude Code and the Claude API as claude-sonnet-5, on AWS Bedrock, Google Cloud Vertex AI and Microsoft Foundry, and inside third-party developer tools including VS Code, GitHub Copilot, Cursor and OpenRouter. The model keeps the 1-million-token context window of the Sonnet line.

On safety, Anthropic reports that Sonnet 5 shows a lower overall rate of undesirable behaviors than Sonnet 4.6 — better at refusing malicious requests, more resistant to prompt-injection attacks, and with lower hallucination and sycophancy rates. Notably, it performs substantially worse than Opus 4.8 on dangerous cybersecurity tasks, a deliberate safety outcome: in tests it never produced a working Firefox exploit (a 0.0% success rate), and it ships with cyber safeguards enabled by default. Early testers said it finishes complex jobs where previous Sonnet models would stop short, and checks its own output without being asked.