Anthropic's Claude Sonnet 5 Brings Opus-Level Agentic Power at Half the Price

Anthropic on June 30 released Claude Sonnet 5, the newest iteration of its mid-tier model family, delivering what the company calls near-Opus performance at a fraction of the cost. The release marks a significant shift in the economics of agentic AI: tasks that previously demanded the company’s flagship Opus 4.8 — and the price tag that came with it — can now be handled by a model that costs roughly one-third as much per token.

Closing the Gap with the Flagship

The headline benchmark numbers tell a precise story. On agentic coding — measured via the SWE-bench Verified suite that asks models to fix real GitHub issues autonomously — Sonnet 5 scores 63.2%. Opus 4.8 scores 69.2%. The previous Sonnet generation, 4.6, sat at 58.1%. In four months, Anthropic narrowed the gap between its mid-tier and flagship by roughly 80%.

On knowledge-work benchmarks, the gap closes further: Sonnet 5 actually edges out Opus 4.8 on reasoning-heavy tasks that don’t require raw coding throughput. For most enterprise automation workflows — the drafting, synthesizing, researching, and multi-step decision-making that makes up the bulk of real-world agent deployments — Sonnet 5 is the more efficient choice by any measure.

The model’s most-cited improvement over its predecessor is end-to-end task completion. Earlier versions of Sonnet routinely abandoned complex multi-part assignments midway, requiring human re-prompting. Sonnet 5 was specifically trained on such sequences and “checks its own output without explicitly being asked,” according to Anthropic. That behavioral change matters enormously in autonomous pipelines where a model recovering from its own mistakes eliminates a whole class of failure modes.

Pricing That Reshapes Agent Economics

At launch, Sonnet 5 is priced at $2 per million input tokens and $10 per million output tokens — an introductory rate that holds through August 31. After that, prices step up to $3 and $15 respectively, still below Opus 4.8 and well below GPT-5.5. By contrast, Gemini 3.5 Flash is cheaper, but Anthropic is betting that Sonnet 5’s reliability premium justifies the spread.

For developers building agentic systems, where a single user session can consume tens of millions of tokens across dozens of tool calls, the difference between Opus and Sonnet pricing is not a rounding error — it’s the difference between a viable product margin and an unsustainable cost structure. The Zapier team put it plainly: “For day-to-day automation, it’s a no-brainer.”

Sonnet 5 is immediately available across all Anthropic plans: it replaces the previous default model for Free and Pro subscribers, and is accessible to Max, Team, and Enterprise users. It’s rolling out on AWS Bedrock and is being integrated into GitHub Copilot and third-party platforms.

Safety as a Product Feature

Anthropic made safety improvements central to the Sonnet 5 announcement, framing them not as compliance overhead but as production reliability metrics. The model scores lower rates of undesirable behaviors compared to Sonnet 4.6, including better refusal of malicious requests and stronger resistance to prompt-injection attacks — a critical capability as agents are increasingly deployed in environments where adversarial inputs arrive through external websites, emails, or user-provided documents.

Hallucination and sycophancy rates also dropped. The latter is particularly important for enterprise use cases where models agreeing with users rather than correcting them can introduce silent errors into automated workflows. “A model that knows when to say no is just as important as one that knows how to build,” Lovable’s co-founder noted when commenting on the launch.

The model is not yet at Opus 4.8’s level on preventing the most dangerous task executions, a category Anthropic defines as assisting with actions that could cause mass harm. But for the overwhelming majority of production deployments, Sonnet 5’s safety profile represents a meaningful step forward.

The Timing Is Strategic

The release comes at a pivotal moment for Anthropic. The company is on track for a blockbuster IPO and recently crossed $965 billion in private valuation following a $65 billion Series H — overtaking OpenAI as the most valuable standalone AI lab. Revenue is accelerating, with enterprise customers in particular deepening their API commitments.

Launching Sonnet 5 as the default model for millions of free users accomplishes two things simultaneously: it puts a best-in-class agentic experience in front of consumers at no cost, building the usage data and habit formation that sustains long-term market share; and it gives the developer community a price point that makes building production-grade agents economically feasible without negotiating enterprise contracts.

The timing also keeps pressure on OpenAI, which markets GPT-5.5 at significantly higher rates, and on Google, whose Gemini 3.5 Pro has been delayed to July and remains unreleased at time of publication. In the race to be the default reasoning layer for autonomous software, Anthropic is combining performance, price, and ubiquity into a single product moment.

What’s Next

Anthropic has signaled continued investment in the Sonnet tier as the volume workhorse of its lineup. While Opus will remain the choice for the most demanding tasks — particularly safety-critical deployments where cost is secondary to reliability — the gap is narrowing with each release cycle.

For the developer ecosystem, Sonnet 5 arrives at the moment when agentic frameworks — from LangGraph and AutoGen to custom enterprise scaffolding — are maturing into production-ready infrastructure. Having a model that matches Opus-grade reasoning on real agentic tasks, but at mid-tier pricing, could accelerate the shift from experimental pilots to large-scale deployments across industries from legal to healthcare to financial services.

The autonomous agent era doesn’t need a better flagship model right now. It needs a reliable, affordable mid-tier. Anthropic just shipped it.