Claude Sonnet 5 Becomes Anthropic's New Default Model

Anthropic has made Claude Sonnet 5 the default model for every Free and Pro user, cementing a launch that reframes how much autonomous AI work should cost. The company released the model on June 30, 2026 and rolled it out as the standard option across its consumer plans, with paid availability on Max, Team, and Enterprise tiers.

The pitch is unusually direct. Anthropic says Sonnet 5 is its most agentic mid-tier model, delivering near-flagship performance at a fraction of the price of its top model, Opus 4.8.

What Anthropic Shipped

Sonnet 5 is built to act rather than only answer. Anthropic said the model can make plans, use tools such as browsers and terminals, and run autonomously at a level that a few months ago required larger and more expensive systems, as reported by TechCrunch.

That framing mirrors what rivals have said about their own releases. TechCrunch noted that OpenAI's GPT-5.6 Sol and Google's Gemini 3.5 Flash carried similar agentic pitches, which makes autonomous capability the new baseline at every tier.

The release also lands as Anthropic prepares for a public listing. VentureBeat framed the launch as a bid to give cost-conscious enterprise developers powerful agentic tools as the lab races toward a blockbuster IPO.

How the Pricing Works

Price sits at the center of this launch. Introductory API pricing is two dollars per million input tokens and ten dollars per million output tokens through August 31, 2026, after which it rises to three and fifteen dollars, per PYMNTS.

That undercuts Anthropic's own Opus 4.8 by roughly 40 percent at standard rates and about 60 percent during the introductory window, according to VentureBeat. It also prices Sonnet 5 below OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, though it remains costlier than Gemini 3.5 Flash.

There is a catch that buyers should model carefully. Sonnet 5 uses an updated tokenizer, so the same text can map to as much as 1.35 times more tokens, and The Next Web reported that Anthropic set the introductory price to keep the switch roughly cost neutral.

Where Sonnet 5 Lands on Benchmarks

The quality gains are real but stop just short of the flagship. On the SWE-bench Pro agentic coding test, Sonnet 5 scored 63.2 percent, sitting between Sonnet 4.6 at 58.1 percent and Opus 4.8 at 69.2 percent, according to figures published by VentureBeat.

On a knowledge-work benchmark called GDPval-AA v2, the same VentureBeat reporting put Sonnet 5 at 1,618, narrowly ahead of Opus 4.8 at 1,615. The pattern is consistent, namely a large step up from Sonnet 4.6 that closes much of the gap to a model costing far more to run.

Early testers described a model that finishes work. A senior engineer at Zapier, Daniel Shepard, said Sonnet 5 completed a two-part Salesforce job end to end that used to stall halfway, per TechCrunch. Cursor co-founder Sualeh Asif told VentureBeat that agents built on the model stay on plan and ship clean multi-step changes at an efficient cost.

Why Cheaper Agents Matter Now

The timing is the story. Quartz reported that heavy AI token consumption had grown costly even for large firms, with Meta, Amazon, and Uber curbing how much their systems spend on token processing.

That squeeze is exactly what a capable mid-tier model is meant to relieve. For teams weighing autonomous workflows, the economics of running large AI models have often been the blocker rather than raw capability, a tension explored in earlier coverage of the economics of running large AI models.

Safety Shipped by Default

Anthropic paired the launch with tighter controls. The company said Sonnet 5 hallucinates less than its predecessor, flatters users less, and resists prompt injection better, and TechTimes reported that it ships with real-time cyber safeguards enabled by default.

Anthropic did not train the model for cybersecurity tasks, and in a Firefox exploit evaluation built with Mozilla it produced no working exploit, TechTimes noted. The company still turned the guardrails on, a sign that shipping safely has become a condition of shipping at all.

What Operators Should Do Next

Treat Sonnet 5 as a reason to re-run your own numbers rather than trust the headline rate. Because of the new tokenizer, teams should benchmark real workloads inside their own stack before assuming a straight per-token saving.

The model is now the default in Claude's apps, in Claude Code, and through the API. Anyone who has debated which Claude models to standardize on, a question visible in earlier shifts inside GitHub Copilot's Claude lineup, now has a cheaper agentic baseline to test, and the practical value of tools like Claude increasingly turns on cost per completed task rather than benchmark bragging rights.