Anthropic has given its mid-tier lineup a serious upgrade. Claude Sonnet 5, launched June 30, 2026, is the company's most capable mid-range model to date and is explicitly designed to run autonomous agentic workflows without breaking the budget. According to TechCrunch, Anthropic is positioning the model as a more affordable alternative to its own Opus models, OpenAI's GPT-5.5, and Google's Gemini Pro.

What's new in Sonnet 5?

The most significant change centers on agentic capability — the ability to plan and execute multi-step processes independently. Claude Sonnet 5 can use tools such as browsers and terminals, handle complex workflows, and act autonomously at a level that, just months ago, required far more expensive models, according to InfoWorld's coverage of the launch.

Anthropic itself describes the performance as "close to Opus 4.8, but at lower prices."

63.2%
SWE-bench Pro (agentic coding)
1,618
GDPval-AA v2 (knowledge work)
Claude Sonnet 5 cuts prices and takes on GPT-5.5 on agentic tasks - Bilde 1

Benchmarks: closing in on the top

Research data from the launch shows a clear step up from predecessor Sonnet 4.6 across key evaluations:

  • SWE-bench Pro (agentic coding): Sonnet 5 achieves 63.2%, versus 58.1% for Sonnet 4.6 and 69.2% for Opus 4.8
  • OSWorld-Verified (computer control): 81.2% versus 78.5% for Sonnet 4.6
  • Terminal-Bench 2.1 (code evaluation): Sonnet 5 jumps from 67.0% to 80.4%, approaching Opus 4.8's 82.7%
  • Humanity's Last Exam (multidisciplinary reasoning with tools): 57.4% — nearly identical to Opus 4.8's 57.9%
  • GDPval-AA v2 (knowledge work): Sonnet 5 scores 1,618, actually beating Opus 4.8's 1,615

On the Artificial Analysis Intelligence Index, the model earns 53 points and reaches the same level as GPT-5.5 with high reasoning at maximum effort — something MarkTechPost describes as "Anthropic's most agentic mid-tier model, closing much of the gap to Opus 4.8."

On knowledge work, Sonnet 5 actually surpasses Opus 4.8 — a rare instance of a mid-range model beating the flagship

Pricing and the tokenizer caveat

The introductory price is set at $2 per million input tokens and $10 per million output tokens, available via API through August 31, 2026. After that, prices adjust to $3 and $15 respectively — still notably lower than Opus 4.8's $5 and $25.

One important caveat: Anthropic notes that an updated tokenizer may produce 1.0 to 1.35 times more tokens for the same text. This means real-world cost savings could be somewhat lower than the raw numbers suggest, and users should calculate their actual token volumes for their specific use case before drawing conclusions.

Safety and limitations

Anthropic's own safety reports indicate that Sonnet 5 generally behaves better than Sonnet 4.6 in agentic scenarios — with greater resistance to prompt injection attacks, lower rates of hallucination, and reduced sycophancy. The model is also more effective at refusing harmful requests.

At the same time, it is worth noting that Sonnet 5 exhibits, according to Anthropic, "a slightly higher degree of misaligned behavior" compared to Opus 4.8 and Claude Mythos Preview. In addition, cybersecurity capabilities are deliberately constrained: the model performs markedly worse than Opus 4.8 and Mythos 5 on tasks related to developing software exploits.

Who is this model for?

In practice, Claude Sonnet 5 is aimed at developers and enterprises that want to run autonomous agentic pipelines in production but cannot justify Opus pricing for every task. The performance gap to Opus 4.8 remains on coding and general intelligence, but for knowledge work and tool-based reasoning, Sonnet 5 is now genuinely competitive — including against GPT-5.5.

Sources emphasize that this is a fast-moving segment, and that direct comparisons with the latest versions of competing models are still being developed.