Anthropic is pushing its midsize model into a more agentic role with Claude Sonnet 5, a release aimed at developers and users who want autonomous work without paying for the most expensive model tier.
The company says the model can plan, use tools such as browsers and terminals, and operate independently at a level that recently required larger and more expensive systems. That puts Claude Sonnet 5 directly in the center of a broader shift in AI: agentic capability is no longer being treated as a premium-only feature.
Agentic AI moves into the default tier
Claude Sonnet 5 arrives as foundation model companies increasingly present agentic behavior as a standard expectation. The source article points to OpenAI’s GPT-5.6 Sol, launched in preview last week, and Google’s Gemini 3.5 Flash, launched in May, as recent examples of the same direction.
In each case, the message is similar: models are being positioned less as chatbots and more as systems that can plan, use tools, split work, build, iterate and carry out longer tasks with less human input.
Anthropic’s framing for Sonnet 5 follows that pattern. The model is not just being sold as smarter in conversation. Its value is tied to whether it can complete real tasks across reasoning, coding, tool use and knowledge work.
That changes the competitive question. If agentic ability is becoming table stakes, the next pressure point is not simply which model can act most independently. It is which model can do that work cheaply enough, reliably enough and safely enough for regular use.
Lower pricing is central to the Sonnet 5 pitch
Anthropic says Sonnet 5 offers performance close to Opus 4.8 while costing much less. Starting Tuesday, Claude Sonnet 5 becomes the default model for free and Pro plans, and it is available for every subscription.
At launch, the model is priced at $2 per million input tokens and $10 per million output tokens through August 31. After that, the price rises to $3 per million input tokens and $15 per million output tokens.
That makes Sonnet 5 cheaper than Opus 4.8, OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro, according to the source article. It remains more expensive than Gemini 3.5 Flash.
For developers, the practical point is choice. Anthropic is presenting Sonnet 5 and Opus 4.8 as models that can sit at different effort levels, letting users decide how much performance they need for a given job and how much they are willing to spend.
"Opus 4.8 is still the model of choice for higher accuracy on these tasks, but Sonnet 5 provides developers with lower-priced options that are of much higher quality than what was previously available," Anthropic says. "Between Sonnet 5 and Opus 4.8, users can adjust the effort level to find the right balance of cost and performance."
Benchmarks show gains over Sonnet 4.6
Anthropic says Claude Sonnet 5 improves significantly over Sonnet 4.6, which was released in February. The gains are described across agentic performance, including reasoning, tool use, software coding and knowledge work.
On one agentic coding benchmark, Sonnet 5 scores 63.2%. That compares with 69.2% for Opus 4.8 and 58.1% for Sonnet 4.6.
The source article also says Sonnet 5 slightly outperforms Opus 4.8 on a knowledge work benchmark. That matters because Opus 4.8 is described as the model known for solving the hardest problems, including subtle judgment calls and deep research.
The message is not that Sonnet 5 replaces Opus 4.8 for every use. Rather, Anthropic is arguing that its midsize model now handles a larger share of demanding work at a lower price point.
Task completion is part of the upgrade
Testers cited in Anthropic’s blog post said Sonnet 5 is better at finishing complex tasks where earlier versions might stop before completion. The model is also described as checking its own output without needing to be explicitly asked.
That matters for agentic workflows because partial completion can limit the usefulness of automation. A model that can start a multi-step task but cannot carry it through still leaves the user managing the handoff.
Daniel Shepard, a senior engineer at Zapier, described one example involving Salesforce account tiers and a launch announcement to enterprise contacts.
"We handed Claude Sonnet 5 a two-part job — update Salesforce account tiers, send a launch announcement to enterprise contacts — and it finished end to end," Daniel Shepard, a senior engineer at Zapier, said in a statement. "That used to stall halfway. For day-to-day automation, it’s a no-brainer. "
The example illustrates the kind of routine business automation Anthropic is targeting: tasks that combine planning, tool use and execution across more than one step.
Safety claims focus on agentic risks
Anthropic also says Sonnet 5 shows a lower rate of undesirable behaviors than Sonnet 4.6. The source article lists cooperation with misuse and deception as examples, and says the model is better at refusing malicious requests and avoiding hijack attempts in prompt-injection attacks.
The model is also described as hallucinating less and engaging in sycophantic behavior at a lower rate than Sonnet 4.6. For agentic systems, those issues matter because the model may be acting through tools rather than only generating text.
Still, Anthropic does not present Sonnet 5 as matching its strongest models on every safety dimension. The source article says it is not on the same level as Opus 4.8 and Claude Mythos Preview for misaligned behavior.
"Evaluations also show that it has a much lower ability to perform dangerous cybersecurity tasks than our current Opus models," reads the blog post.
Lovable co-founder Fabian Hedin also emphasized refusal behavior, saying Claude Sonnet 5 "refuses unsafe requests cleanly and consistently."
"At Lovable, we’re putting powerful tools in the hands of millions of builders," Hedin said. "A model that knows when to say no is just as important as one that knows how to build."
That sums up the release’s broader position. Claude Sonnet 5 is not only about making agentic AI more capable. It is also about making that capability cheaper, more widely available and controlled enough for everyday automation.