# Sonnet 5 brings near-Opus agents at a fraction of the Opus price, and that is a deadline for machine-readable commerce > Claude Sonnet 5 matches or comes within two points of Opus 4.8 on most agentic benchmarks, at a fraction of the price. For commerce, the story is the cost curve. Canonical: https://brandonlazovic.dev/articles/sonnet-5-agent-cost-curve/ Author: Brandon Lazovic Published: 2026-07-01 ## The short version - Claude Sonnet 5 matches near-Opus agentic performance at a fraction of Opus's price, and that price drop is what pulls autonomous agents into economic reach. - Sonnet 5 scores 1618 on GDPval-AA knowledge work, edging Opus 4.8's 1615, and stays within about two points of Opus on most other agentic benchmarks. - Its introductory pricing is 2 dollars per million input tokens and 10 dollars per million output tokens through August 31, then rises to 3 and 15 dollars. - Because agents read a page's structure rather than its pixels, cheaper agents deployed at scale raise the urgency of clean semantic HTML, ARIA roles, and complete product feeds. Anthropic shipped Claude Sonnet 5 on June 30, and the headline most coverage led with is that a mid-tier model now performs close to a frontier one. That is true, but it is not the part that matters for anyone running an e-commerce site. The part that matters is the price: near-Opus agentic capability at a fraction of the Opus price pulls the economics of autonomous agents forward, and those agents read your pages the way a screen reader does, not the way a shopper does. > Sonnet 5 scores 1618 on GDPval-AA knowledge work, edging Opus 4.8's 1615, at a fraction of the price. [1] ## How close does Sonnet 5 actually get to Opus? Sonnet 5 matches or comes within about two points of Opus 4.8 on four of five agentic benchmarks. It edges Opus on GDPval-AA knowledge work (1618 to 1615), effectively ties it on Humanity's Last Exam with tools (57.4 percent to 57.9 percent), and stays within two points on computer use (81.2 to 83.4 percent on OSWorld-Verified) and terminal coding (80.4 to 82.7 percent on Terminal-Bench 2.1). [1] The one clear remaining gap is the hardest software engineering: SWE-bench Pro, where Opus still leads 69.2 to 63.2 percent. ![Grouped bars comparing Sonnet 5, Sonnet 4.6, and Opus 4.8 on four agentic benchmarks: SWE-bench Pro 63.2, 58.1, 69.2; Terminal-Bench 2.1 80.4, 67.0, 82.7; Humanity's Last Exam with tools 57.4, 46.8, 57.9; OSWorld-Verified 81.2, 78.5, 83.4 percent. Sonnet 5 trails Opus by about two points.](/figures/sonnet-5-agent-cost-curve/benchmarks.png) The jump from Sonnet 4.6 is the more useful comparison for anyone who was already using the mid-tier model. Terminal-Bench went from 67.0 to 80.4 percent, Humanity's Last Exam with tools from 46.8 to 57.4 percent, and SWE-bench Pro from 58.1 to 63.2 percent. [1] Anthropic frames the model as its most agentic Sonnet yet, built to plan, use browsers and terminals, and run autonomously at a level that recently required a larger model. [2] ## Why does a cheaper agent model matter more than a smarter one? Because an agent is not one model call. A single autonomous task, browsing a catalog, comparing products, filling a cart, checking out, spends many calls across many steps, and the cost compounds with every page the agent has to read and every decision it has to make. At frontier-model prices, running that loop at any real volume is hard to justify. That is the agent cost curve, and it is what decides whether autonomous commerce agents stay demos or become traffic. Sonnet 5's introductory pricing is 2 dollars per million input tokens and 10 dollars per million output through August 31. [1] Anthropic pitched it as a steep discount to the frontier tier while closing much of the capability gap with Opus 4.8. [5] It also became the default model in Claude Code and on the free and Pro plans, with a native one-million-token context window. [4] The direction is consistent: capability that was expensive is being pushed down the cost curve and made the default, which is precisely the condition under which agent workloads scale. ## What does that change for merchants? Claude Sonnet 5's pricing shift moves machine-readability from a theoretical concern to a near-term operational one. I have argued before that [agents read structure, not pixels](/articles/agents-read-accessibility-first/), leaning on the accessibility tree and semantic markup rather than a rendered screenshot. That was true before Sonnet 5; what changes now is the volume. Cheaper capable agents mean more agents actually deployed against real storefronts, and an agent that scores 81.2 percent on OSWorld-Verified is more likely to complete tasks on well-structured pages, and to stall on messy ones. [1] The practical read is that the same work that helps a screen reader, clean semantic HTML, accurate ARIA roles, server-rendered structured data, and a complete product feed, is the work that lets an agent complete a purchase on your site instead of a competitor's. That is not a new to-do list. It is the existing one, with a shorter fuse. ## What is the catch? Three things, and I would not model any of them loosely. First, the introductory price expires: after August 31 it rises to 3 dollars per million input and 15 dollars per million output. [3] Second, Sonnet 5 uses an updated tokenizer that maps the same text to roughly 1.0 to 1.35 times more tokens, so "a fifth of the price" is a headline, not a line item; Anthropic set the introductory rate so the move from 4.6 is roughly cost-neutral despite that change. [2] Run your own token math before you commit a budget. Third, the SWE-bench Pro gap is a reminder that the frontier model still leads on the hardest reasoning, so "close to Opus" is a statement about a benchmark set, not a blanket equivalence. The broader signal, and I label this as my inference rather than Anthropic's claim, is that the model providers are competing on cost-to-capability, not just capability, and are actively steering usage toward cheaper, faster responses. Sonnet 5 shipping as the default across Claude Code and the consumer plans is one tell. I have also been seeing a "quick answer" option surface alongside extended thinking in the Claude app, a nudge toward the cheaper response path when the full reasoning budget is not needed. More capable inference at a lower unit price is good for the people building agents and, not incidentally, good for the economics of the company selling the tokens. For commerce, none of that changes the instruction. It changes the timeline. The output of this model race that actually reaches your storefront is not a smarter chatbot; it is a cheaper autonomous agent, arriving sooner, reading your structure to decide whether it can act. The readability work has a date on it now. ## Sources 1. Introducing Claude Sonnet 5 (Anthropic) — https://www.anthropic.com/news/claude-sonnet-5 2. What's new in Claude Sonnet 5 (Claude Platform Docs) — https://platform.claude.com/docs/en/about-claude/models/whats-new-sonnet-5 3. Anthropic launches Claude Sonnet 5 as a cheaper way to run agents (TechCrunch) — https://techcrunch.com/2026/06/30/anthropic-launches-claude-sonnet-5-as-a-cheaper-way-to-run-agents/ 4. Claude Code model configuration (Sonnet 5 default, 1M context, v2.1.197) — https://code.claude.com/docs/en/model-config 5. Sonnet 5 closes the gap with Opus 4.8 (The New Stack) — https://thenewstack.io/claude-sonnet-5-launch/