Claude Sonnet 5 narrows the gap with Opus

Running real, autonomous agent workflows used to feel like burning cash. If you wanted an AI model that could use a terminal, browse the web, and not get hopelessly stuck halfway through a task, you had to pony up for top-tier flagship models.

Anthropic wants to change that math. On June 30, the AI lab launched Claude Sonnet 5, branding it as its most agentic Sonnet model to date. The new release narrows the gap with Anthropic's flagship Opus 4.8, matching much of its reasoning, coding, and tool-use capabilities at a fraction of the cost.

Sonnet 5 is rolling out everywhere immediately. It is now the default model for Free and Pro users, available on Max, Team, and Enterprise tiers, integrated into Claude Code, and accessible via the API as well as Amazon Bedrock. Anthropic originally launched it at an introductory rate of $2 per million input tokens and $10 per million output tokens, and later made that pricing permanent instead of hiking it to $3/$15.

Agents that finish the job

Compared to Sonnet 4.6, released in February, Sonnet 5 is built specifically for autonomous execution. It handles multi-step tool interactions, navigates real-world codebases, and checks its own work without prompting.

On benchmarks, Sonnet 5 hit 63.2% on agentic coding, up from 4.6's 58.1% and closing in on Opus 4.8's 69.2%. On knowledge work evaluations, it even edged past Opus 4.8. Anthropic also bumped rate limits across Chat, Cowork, and Claude Code to account for the higher token burn that autonomous agent loops demand.

Early enterprise testers noticed the jump in stamina. Daniel Shepard, a senior engineer at Zapier, noted that when handed a two-part task—updating Salesforce account tiers and firing off launch announcements to enterprise contacts—Sonnet 5 completed it end to end. Previous models routinely stalled out halfway.

Built-in boundaries

Anthropic also took a deliberate stance on safety. Because autonomous tools can interact with operating systems and networks, Sonnet 5 was launched with real-time cybersecurity safeguards enabled by default.

Anthropic noted it did not deliberately train the model on cybersecurity tasks. When tested against Firefox browser exploits, Sonnet 5 never managed to build a fully working exploit, scoring significantly lower on dangerous offensive tasks than Opus 4.8 or Mythos 5.

It also showed fewer hallucinations, less sycophancy, and better resistance to prompt injection attacks than Sonnet 4.6. Lovable co-founder Fabian Hedin praised the model's discipline, stating that knowing when to say no is just as vital as knowing how to build.

Why it matters

For developers building automated pipelines and coders looking for reliable assistants, Sonnet 5 is an immediate upgrade. You get near-flagship autonomy without the Opus-sized invoice, making long-running background tasks economically viable for the first time.

Across the industry, agentic competence has officially become the baseline. With OpenAI testing GPT-5.6 Sol and Google pushing Gemini 3.5 Flash, the battleground is no longer just raw reasoning benchmarks. The fight has shifted to autonomous reliability per dollar.

We think making the $2/$10 introductory pricing permanent was a masterstroke to squeeze competing mid-tier models. That said, autonomous execution eats tokens at a staggering rate. While individual tokens are cheap, multi-step loops that retry failed actions can still run up hefty bills if you don't keep strict guardrails on your agents.