There’s a new arrival in the Anthropic family. His name is Claude Sonnet 5, born June 30, 2026. As any AIDeskPro user knows, the Claude family has its characters: Fable the superhero, Opus the wise professor, Haiku the speedy little troublemaker. And then there was him - Sonnet 4.6, the reliable engineer, the one who never lets you down.
Well. Sonnet 4.6 is now holding a baby in his arms. And the baby is already smarter than him.
Why this matters to us personally
On AIDeskPro, we used Claude Sonnet 4.6 heavily. It was our go-to for agentic tasks: reliable, fast, great quality-to-price ratio. We know its limits well - and we know exactly what it feels like when a model “stalls halfway” through a complex task.
Sonnet 5 was built precisely to fix those moments.
What changes with Sonnet 5
Anthropic positions it as the most agentic Sonnet ever built: capable of planning, using tools like browsers and terminals, and completing complex tasks that just a few months ago required Opus.
The numbers speak for themselves. On SWE-bench Verified (the standard benchmark for real-world coding on GitHub), Sonnet 5 scores 77.2% against Sonnet 4.6’s 62.0%. On OSWorld-Verified, which measures real-world computer use skills, it reaches 80.4% against the predecessor’s 78.5%.
But the data point that struck us most isn’t a benchmark. It’s what early access partners reported:
“I asked Claude Sonnet 5 to investigate a bug. Unprompted, it wrote a reproducing test, implemented the fix, then stashed it to confirm the bug came back without the change. All in a single pass.”
That’s the point. It’s not just smarter - it’s more autonomous. It finishes things. It checks its own work. It doesn’t stop halfway.
The price: a launch offer worth noting
Through August 31, 2026, Sonnet 5 is available at $2 per million input tokens and $10 per million output tokens. After that date, it moves to $3/$15.
For a direct comparison with what we used before: Sonnet 4.6 was at $3/$15. Sonnet 5 at launch costs less and delivers more. That doesn’t happen often.
Opus 4.8, the flagship model, is at $5/$25. Sonnet 5 approaches its performance on many agentic tasks, at a fraction of the cost.
Safe by default
One aspect worth highlighting for those using AI in business contexts: Anthropic ran thorough safety evaluations before launch.
Sonnet 5 shows a lower rate of undesirable behaviors than Sonnet 4.6. It’s better at resisting prompt injection attacks and refusing malicious requests in agentic contexts. It hallucinates less. It’s less sycophantic - meaning it’s less likely to tell you what you want to hear instead of the truth.
For those running AI in automated pipelines or multi-step tasks with access to external systems, these aren’t footnotes. They’re the things that matter.
Available now on AIDeskPro
Claude Sonnet 5 is already available on AIDeskPro. You can select it from the model menu and start using it right away.
Our suggestion: try giving it a task that normally required multiple passes or tended to stall halfway. That’s where the difference is most felt.
Welcome, Sonnet 5. You’re already at home.
Source: Introducing Claude Sonnet 5 - Anthropic , June 30, 2026.