Anthropic has introduced the Claude Sonnet 5, its most agent mid-range model yet. It performs close to the flagship Opus 4.8 in many functions, but costs less than half.
anthropic He said on June 30, 2026 Sonnet 5 is available today in all plans. The company built it to act, not just to respond. He can make plans, manage browsers and terminals, and work alone for long periods of time. This kind of work required larger, more expensive models just a few months ago.
The pitch is simple. The Sonnet 5 offers near-flagship performance at a mid-range price. It is close to Anthropic’s most capable model, Opus 4.8, in reasoning, tooling, coding, and knowledge work. It is clearly superior to its predecessor Sonnet 4.6. And it costs a lot less than Opus to run.
Cheaper agents, on purpose
Price is at the heart of this presentation. Sonnet starts at $2 for 5 million entry tokens and $10 for a million exit tokens. This introductory rate lasts until August 31, 2026. After that, it goes to $3 and $15. Opus 4.8, on the other hand, costs $5 and $25. TechCrunch framed the model as a cheaper way to manage agents, and that’s the point.
Time is important. Companies rushed to deploy AI agents, then backed out of the invoices. Agents make rounds, calling tools and burning tokens at a rapid pace. A model approaching Opus quality for a fraction of the cost speaks directly to this pain. It also mentions market hunting savings after enterprise AI invoices ballooned.
There is a catch in the small print. Sonnet 5 uses a new tokenizer, so the same text can be tokenized up to 1.35 times more than before. Anthropic has set the entry price so that the switch remains roughly cost neutral. The cap rate seems low, but the number of tokens may increase.
How good is it?
By Anthropic’s own criteria, the Sonnet 5 is a clear step up from the 4.6 without quite catching the Opus. In the agent encoding test, it scored 63.2 percent versus 69.2 percent for Opus 4.8 and 58.1 percent for Sonnet 4.6. according to an early report. In one knowledge work benchmark, it outperformed Opus. Anthropic also offers “effort” dialing, which allows developers to swap values between two models for accuracy.
Early testers told Anthropic that the model completes complex tasks that older Sonnets refused to do, and that it checks its output without being prompted. These claims come from the company’s source material, so they deserve the usual caution. Independent testing will tell the real story.
Safer with cyber alert
Anthropic says that Sonnet 5 behaves better than 4.6 in terms of security. It rejects malicious requests more often and resists operational injection attacks where hidden instructions try to hijack the agent. It also hallucinates and is less flattering. In an automated audit of maladaptive behavior, it scored worse than Opus 4.8 and the Mythos preview, but safer than 4.6.
Cyber security is a sharper point. Anthropic has not trained the Sonnet 5 for cyber missions, and it does poorly at creating software exploits. In testing with Mozilla on the Firefox browser, the model never produced a working exploit. Nevertheless, Anthropic shipped it with the same real-time cyber security measures used in Opus 4.7 and 4.8. These guardrails stay lighter than their surroundings Tale 5its locked public model.
A discount with a strategy behind it
A low price is not a charity. Anthropic competes with competitors for developers, and a competent, cost-effective agent model is how you win them over. The company also writes most of it own code With Claude, therefore, the better, cheaper Sonnet also helps its engineers. It is also moving toward a planned public listing, where revenue growth and the developer have both been reached.
The broader context is cost. 24/7 agents can be rack tearful accountsand Anthropic set off ambitious revenue targets to finance his modeling work. Sonnet 5 is the answer to both. Push the capability down the price curve, keep the developers in the ecosystem, and let the effort do the rest.
Claude Sonnet 5 is now live in Claude apps, Claude Code and API, with higher rate limits across the board. For most developers, the question is no longer whether the model is smart enough. Whether it’s cheap enough to run all day. Anthropic is betting that the answer is ultimately yes.
