On June 30, 2026, Anthropic released Claude Sonnet 5, the next generation of its mid-tier model that promises to redefine cost-performance in AI agents. The model arrives with agentic capabilities that until recently were exclusive to larger, more expensive models — at a price point that makes scaling AI agents in production viable.
The agentic leap
The Sonnet line has an important history in the Claude ecosystem. Sonnet 3.5, 3.6, and 3.7 were the first models to demonstrate impressive coding and tool-use skills, marking the beginning of the AI agent era for many developers. But in recent months, the clearest gains in agentic capabilities had been concentrated in Opus-class models.
Sonnet 5 changes that. According to Anthropic, the model’s performance is close to Opus 4.8 — currently the company’s most capable model — at a significantly lower price. The improvement over its predecessor, Sonnet 4.6, is substantial across reasoning, tool use, coding, and knowledge work.
Internal evaluations show consistent gains across multiple benchmarks:
- BrowseComp (agentic search evaluation): outperforms Sonnet 4.6 at all effort levels, approaching Opus 4.8 at high effort settings
- OSWorld-Verified (computer use via agent): covers a wider cost-performance range than Opus 4.8
- Humanity’s Last Exam: benefits from the adjustable “effort” system, allowing users to balance cost and accuracy per task
Anthropic’s tests show Sonnet 5 narrowed the gap with Opus 4.8 so significantly that for many everyday tasks the practical difference is small — but the cost is nearly half.
Pricing and availability
Claude Sonnet 5 has been available since June 30 across all plans (Free, Pro, Max, Team, and Enterprise) as well as via Claude Code and the API.
| Parameter | Price (USD per million tokens) |
|---|---|
| Input (introductory until Aug 31) | $2 |
| Output (introductory until Aug 31) | $10 |
| Input (standard pricing) | $3 |
| Output (standard pricing) | $15 |
For comparison, Opus 4.8 costs $5/MTok input and $25/MTok output. The introductory price of $2/MTok was calibrated so that migrating from Sonnet 4.6 is approximately cost-neutral, considering Sonnet 5 uses an updated tokenizer that may increase token count by 1.0–1.35× depending on content.
Rate limits have also been increased across all tiers (Start, Build, Scale) to accommodate higher token consumption in high-effort configurations.
Safety: better than its predecessor
Anthropic reports that Sonnet 5 is safer than Sonnet 4.6 on agentic safety metrics:
- Better at refusing malicious requests
- More resistant to prompt injection hijack attempts
- Lower rates of hallucination and sycophancy
- Overall lower rate of misaligned behaviors in automated behavioral audits
The model was not deliberately trained on cybersecurity tasks. In evaluations using Firefox 147 (in partnership with Mozilla), Sonnet 5 never successfully developed a working exploit. It showed a slightly higher rate of partial success than Sonnet 4.6, attributed to general intelligence gains rather than specific training.
As a precaution, Anthropic launched Sonnet 5 with cyber safeguards enabled by default — the same ones present in Opus 4.7 and 4.8 (less restrictive than Fable 5’s safeguards).
What early adopters are saying
Early access partner reports paint a consistent picture: Sonnet 5 completes complex tasks where previous Sonnet models would stop halfway.
Highlights from Anthropic’s published feedback:
- An engineer at Lovable described the model as “getting more done with less — same output quality, fewer steps to get there”
- At Box, the model investigated a bug, wrote a reproducible test, implemented the fix, and verified the bug returned without the change — all in a single pass
- Pace (insurance) uses computer-use agents for insurance workflows, and Sonnet 5 “consistently takes the right action and does it quickly”
- Sentry noted Sonnet 5 is particularly good at “brownfield code — race conditions, hidden tests, the parts nobody wants to touch”
Implications for the AI agent market
Sonnet 5 arrives at a crucial moment. The AI agent market is maturing, and cost per operation is still the main barrier to scale adoption. By offering a model with Opus-level performance at Sonnet pricing, Anthropic is betting that cost-performance — not just raw capability — will be the competitive differentiator in the coming months.
Developers using agents for task automation, assisted coding, and multi-step workflows can benefit directly: Sonnet 5 allows setting the “effort” level per task, paying only for what you need. For simple tasks, the model works fast and cheap; for complex tasks, high effort delivers frontier-level performance.
Read also:
- Inside AI Brains: How Anthropic Decoded Claude’s Thinking Process
- Why I’m Breaking Up With Vibe Coding
- Claude Tag and the Risk to Companies’ Intellectual Property
Feel free to reach out at [email protected]