Anthropic has launched Claude Sonnet 5, an upgraded version of its mid-tier model built specifically for agentic tasks. The model can browse the web, use terminals, make plans, and run autonomously through complex multi-step work. What makes this release notable is not just what it can do, but what it costs to do it.
Starting Tuesday, Sonnet 5 becomes the default model for free and Pro plans across all Anthropic subscriptions. It is priced at $2 per million input tokens and $10 per million output tokens through August 31, after which input pricing rises to $3 per million. That makes it cheaper than Opus 4.8, OpenAI’s GPT-5.5, and Google’s Gemini 3.1 Pro, though it is still more expensive than Google’s Gemini 3.5 Flash.
The practical upshot is that developers can now run agentic workloads at a meaningful cost reduction compared to what was previously required. Anthropic says Sonnet 5 delivers performance close to Opus 4.8, which is its most powerful model, without the matching price tag.
This release fits into a clear pattern across the AI industry right now. OpenAI launched GPT-5.6 Sol last week, pitching it as its most agentic model to date, capable of splitting work across subagents for longer autonomous tasks. Google’s Gemini 3.5 Flash, which launched in May, was positioned as a shift from conversational chatbot to an agentic tool that can plan, build, and iterate with minimal human input. Every major AI lab is now treating agentic capability as the new baseline, not a premium feature.
The real competition, then, is shifting. It is no longer about who can build a model that does agentic work. It is about who can do it most reliably and most cheaply without requiring a human watching over every step. Sonnet 5 is Anthropic’s clearest statement yet on where it sits in that race.
On benchmarks, Sonnet 5 scores 63.2% on agentic coding, compared to 58.1% for its predecessor Sonnet 4.6 and 69.2% for Opus 4.8. On knowledge work tasks, it actually edges out Opus 4.8 slightly, which is a meaningful result given that Opus has historically been the model of choice for deep research and subtle judgment calls. Anthropic describes the two models as complementary: Sonnet 5 for cost-efficient automation at scale, Opus 4.8 for the highest-accuracy work where getting it exactly right matters more than price.
Early testers reported that the model handles complex, multi-part tasks without stopping short, which has been a common frustration with previous versions. Daniel Shepard, a senior engineer at Zapier, said the model completed a two-step job involving Salesforce account updates and a customer email announcement from start to finish, something he said used to stall midway through. That kind of end-to-end reliability is what businesses need before they feel comfortable deploying AI agents in production workflows without a human in the loop.
Safety is part of the pitch too. Compared to Sonnet 4.6, the new model shows lower rates of:
- Cooperating with misuse attempts
- Deceptive behavior
- Hallucinations
- Sycophantic responses
- Falling for prompt-injection attacks
Fabian Hedin, co-founder of Lovable, said in a statement that Sonnet 5 “refuses unsafe requests cleanly and consistently,” adding that a model knowing when to say no is just as important as knowing how to build. That matters more as AI agents gain access to real tools and systems, not just text generation.
There are limits, though. Anthropic is clear that Sonnet 5 does not match Opus 4.8 or Claude Mythos Preview on safety evaluations related to misaligned behavior, and it has a lower ability to perform dangerous cybersecurity tasks than current Opus models. For most business use cases, that tradeoff is likely acceptable. But it is worth noting for anyone considering the model in higher-risk environments.
The broader significance here is what this release signals about where AI product development is heading. Price-competitive agentic models at the mid-tier level mean that autonomous AI workflows are becoming accessible to a much wider range of developers and businesses. That changes what is possible to build, and raises new questions about oversight, reliability, and accountability as these systems take on more real-world work with less human supervision.
