Anthropic launches Claude Sonnet 5, bringing agentic capabilities near Opus 4.8 at a steep discount

Launch details

Sonnet 5 became the default model for Claude Free and Pro tiers on June 30, and is also available to Max, Team, and Enterprise customers. Anthropic calls it “the most agentic Sonnet model yet,” capable of planning, using tools like browsers and terminals, and running autonomously on tasks that previously required larger models.

“It can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models.”

Performance benchmarks close the gap with Opus

On SWE-bench Pro, an agentic coding test, Sonnet 5 scores 63.2%, up from Sonnet 4.6’s 58.1% and approaching Opus 4.8’s 69.2%. On Terminal-Bench 2.1, it reaches 80.4% versus Opus’s 82.7%. In a knowledge-work benchmark (GDPval-AA v2), Sonnet 5 scored 1,618, marginally above Opus’s 1,615. Across agentic tasks, the new model overlaps substantially with the flagship while running at lower cost.

Sonnet 4.658.1%

Sonnet 563.2%

Opus 4.869.2%

Sonnet 4.6

58.1 %

Sonnet 5

63.2 %

Opus 4.8

69.2 %

Data from the article text

Pricing and market strategy

Introductory API pricing is set at $2 per million input tokens and $10 per million output tokens through August 31. After that, prices rise to $3/$15 according to VentureBeat, though TechCrunch reports the output price will stay at $10. Either figure keeps Sonnet 5 well below Opus 4.8’s $5 input and $25 output rates, and below GPT-5.5 and Gemini 3.1 Pro. The company wants to democratize access to agentic AI while building a broad developer base ahead of its planned IPO.

Regulatory backdrop and security

Anthropic says Sonnet 5 was not deliberately trained on cybersecurity tasks and poses “much lower ability to perform dangerous cybersecurity tasks” than Opus models. The release comes amid ongoing talks with the Trump administration about model access. The government had previously asked Anthropic to remove its most powerful models (Mythos and Fable) over security concerns. Mythos is now available on a limited basis, and Fable 5 is expected to return soon. The administration also requested OpenAI stagger the release of its GPT-5.6 class of models.

Competitive landscape

The launch mirrors similar agentic pushes from OpenAI (GPT-5.6 Sol previewed last week) and Google (Gemini 3.5 Flash in May). TechCrunch notes that agentic capability is now a baseline expectation, with differentiation shifting to cost and reliability. Axios reports that enterprise users are moving away from chat toward delegating autonomous tasks to AI agents. Anthropic’s move also comes amid rising sensitivity to AI costs, which has driven some developers toward cheaper Chinese models.