Claude Sonnet 5 Is Anthropic's Most Agentic Mid-Tier Model Yet, Nearly Matching Opus at a Third of the Cost

On June 30, Anthropic released Claude Sonnet 5, a model that closes 90 percent of the gap to Opus 4.8 while costing roughly a third as much per token. For founders building on AI, this is the kind of inflection point that quietly reshapes the math behind whole product categories.

Sonnet-class models have been Anthropic's bread-and-butter for developer adoption ever since Sonnet 3.5 proved that mid-tier pricing could deliver near-frontier results. But the last few releases saw the biggest agentic gains concentrated in the expensive Opus line. Sonnet 5 flips that script: it is the first Sonnet that can competently plan, browse the web, use terminals, and execute multi-step workflows autonomously. Those capabilities were effectively locked behind the Opus price tier just three months ago.

What Changed Between Sonnet 4.6 and Sonnet 5

The benchmark table Anthropic published tells the story plainly. On SWE-bench Verified, a coding benchmark, Sonnet 5 scores 72.3 percent compared to Sonnet 4.6's 60.4 percent and Opus 4.8's 73.8 percent. On TAU-bench (tool use), it hits 94.1 percent versus 83.0 percent for Sonnet 4.6 and 93.0 percent for Opus 4.8. On the agentic search benchmark BrowseComp, Sonnet 5 at high effort matches Opus 4.8 outright. On OSWorld-Verified (computer use), it improves from Sonnet 4.6's 26.0 percent to 37.0 percent, though Opus 4.8 still leads at 51.8 percent.

The practical implication: for the majority of coding and tool-use workflows, Sonnet 5 is functionally interchangeable with Opus 4.8. A developer who was paying $15 per million output tokens for Opus can now get essentially the same results at $10 per million. At the introductory rate of $10 until August 31, the gap widens further. This also means that startups previously priced out of using Opus-level intelligence for agentic workflows can now afford Sonnet 5, which expands the addressable market for AI agent products significantly.

The Effort Mechanism Changes the Cost Calculus

Sonnet 5 introduces adjustable effort levels, letting developers dial between speed and depth. At medium effort, BrowseComp scores are within spitting distance of Opus 4.8. At low effort, costs drop sharply for tasks that do not need deep reasoning. This is significant because it changes the unit economics of AI-powered products. A startup that wraps Claude for customer support can now route simple queries through low-effort Sonnet 5 calls at a fraction of the per-ticket cost, reserving deeper reasoning for escalated cases. The same model does both, which removes the complexity of managing multiple model endpoints for different query tiers.

Anthropic's effort-level chart shows Sonnet 5 covering a much wider cost-performance envelope than Opus 4.8. Opus is a one-speed model: you pay full price every time. Sonnet 5 gives you a sliding scale. For founders running AI startups where inference cost is a first-class input to gross margin, this is the difference between building a sustainable business and burning through runway on API calls. The effort mechanism also opens up new product design possibilities, such as offering users a turbo mode for quick answers and a deep-dive mode for complex analysis, all powered by the same underlying model.

Safety in Agentic Contexts

The system card for Sonnet 5 reports a lower overall rate of undesirable behaviors than Sonnet 4.6 across safety evaluations, particularly in agentic contexts where models have more freedom to act. However, Anthropic also notes that Sonnet 5 has a much lower ability to perform cybersecurity tasks than their current Opus models. For founders deploying AI agents in production, this is a double-edged signal: the model is safer by design in some dimensions, but if your use case requires robust cybersecurity capabilities, you still need Opus. The safety improvements are particularly relevant for startups building in regulated industries where model behavior is under regulatory scrutiny.

What This Means for AI Founders

Three implications stand out for startup builders. First, the cost of agentic AI just dropped by roughly 60 percent overnight. Any startup that had dismissed Sonnet-class models as insufficient for autonomous workflows should re-run their cost projections with Sonnet 5. Second, the effort mechanism creates a new vector for product differentiation. Startups that build their own routing logic between low-effort and high-effort inference can offer variable pricing to users or optimize internal margin in ways that were not possible before. Third, the narrowing gap between Sonnet and Opus puts pressure on competitors like Google and OpenAI to deliver similar cost compression in their mid-tier models, which benefits every founder consuming AI APIs regardless of which provider they choose.

Sonnet 5 is available now as the default model across all Claude plans and can be accessed via the Claude API using the model name claude-sonnet-5. The introductory pricing of $2 per million input tokens and $10 per million output tokens runs through August 31, 2026.