Claude Sonnet 5 Lands Near Opus-Class Performance at Roughly Half the Price, as Anthropic Eyes an IPO

Anthropic released Claude Sonnet 5 on June 30, pricing its new mid-tier model at less than half of what its flagship Opus 4.8 costs, while closing much of the performance gap between the two. The launch lands weeks after Anthropic confidentially filed paperwork with the U.S. Securities and Exchange Commission for an initial public offering — and the timing is doing a lot of the talking.

Sonnet 5 is now the default model for Claude’s Free and Pro users, and it’s also available to Max, Team, and Enterprise customers, through Claude Code, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry on Azure. Anthropic is calling it “the most agentic Sonnet model yet” — one built to plan multi-step work, operate browsers and terminals, and run with less human supervision than its predecessor required.

The pricing is the headline, not the ceiling

For most Claude launches, the pitch has been a new capability ceiling. This one is built around the cost floor. Anthropic set Sonnet 5’s introductory pricing at $2 per million input tokens and $10 per million output tokens, a window that holds through August 31. After that, standard pricing takes over at $3 and $15 — which, notably, is exactly what Sonnet 4.6 cost, meaning the model this one replaces gets a straightforward capability upgrade at no added price once the promotion ends.

There’s a wrinkle worth flagging for anyone doing the math on migration savings. Sonnet 5 runs on an updated tokenizer — the same kind of change Anthropic introduced with Opus 4.7 — that can turn the same input into roughly 1.0 to 1.35 times more tokens than older Claude models produced. Anthropic says it calibrated the introductory pricing to make the switch roughly cost-neutral, but code and non-English text tend to feel the tokenizer shift more than plain English prose does. Teams migrating from Sonnet 4.6 should re-baseline actual usage on representative prompts rather than assume the lower sticker price maps directly onto a lower bill.

How close does it actually get to Opus 4.8?

Closer than any previous Sonnet model, but not all the way there. On Anthropic’s agentic coding evaluation, Sonnet 5 scored 63.2%, a clear step up from Sonnet 4.6’s 58.1% — though still behind Opus 4.8’s 69.2%. Sonnet 5 reportedly edges ahead of Opus 4.8 on at least one knowledge-work benchmark. Anthropic’s own framing is that the two models now sit on the same cost-performance curve at different points: push Sonnet 5’s effort level up, and it can match Opus 4.8 on some tasks, while Opus still leads on the hardest reasoning and coding work.

“With Claude Sonnet 5, agents stay on plan, follow our conventions, and ship clean multi-step changes, all at an efficient cost.”
— Sualeh Asif, co-founder, Cursor

Anthropic frames the release around a broader shift: agentic behavior — planning, calling tools, running without a human checking every step — used to be the exclusive territory of the largest, most expensive frontier models. The company now argues that capability has become table stakes across price tiers, which pushes competition among labs toward cost and reliability instead of raw intelligence alone. Zapier engineer Daniel Shepard described handing the model a two-part Salesforce and outreach task that finished end-to-end without stalling, calling it a “no-brainer” for day-to-day automation.

Real-time cyber safeguards, enabled by default

Anthropic says it did not deliberately train Sonnet 5 on cybersecurity tasks, and its safety evaluations back that up. On an exploit-development test built with Mozilla against Firefox 147, neither Sonnet 5 nor Sonnet 4.6 produced a working exploit — both scored 0.0% — though Sonnet 5 showed a modestly higher partial-success rate (13.2% versus 8.8%). Both trail well behind Opus 4.8 (68.8% working exploits) and the restricted Mythos 5 model (88.4%).

That small uptick was enough for Anthropic to ship Sonnet 5 with real-time cyber safeguards switched on from day one — the same detection system used on Opus 4.7 and 4.8, though less restrictive than the controls applied to Fable 5. On Anthropic’s broader behavioral audit, which checks for issues like deception and cooperation with misuse, Sonnet 5 scored safer than Sonnet 4.6 overall, but still showed somewhat higher rates of misaligned behavior than Opus 4.8 and the internal Mythos Preview model.

The IPO backdrop

None of this is happening in a vacuum. Anthropic confidentially filed IPO paperwork with the SEC on June 1, and reporting on the offering has described it as one of the most closely watched public listings in the AI sector. A day before Sonnet 5 shipped, California Governor Gavin Newsom announced a deal giving state agencies — plus the state’s cities and counties — access to Claude at a 50% discount alongside free workforce training, an arrangement Anthropic’s Head of Americas, Kate Jensen, described as putting Claude to work for the people who run the state.

Pieced together, the pattern points toward a specific story Anthropic wants to tell public-market investors: that it can win on price as well as capability, that its agentic differentiation is real, and that its user base is broad enough to make an upgraded, cheaper model the default experience for millions of free and paid users rather than a niche offering reserved for enterprise budgets. Whether institutional and retail investors reward that story the way private markets have is the open question the IPO will actually answer.

Where it fits in Anthropic’s lineup

• Haiku — fastest, cheapest tier for lightweight tasks.
• Sonnet 5 — new mid-tier default; agentic, near-Opus performance, lowest cost-to-capability ratio in the lineup.
• Opus 4.8 — Anthropic’s public flagship; still the recommended choice for the highest-accuracy coding and reasoning work, and for cybersecurity tasks requiring fewer guardrails.
• Fable 5 and Mythos 5 — Anthropic’s most capable, most restricted models, available in limited form through the Project Glasswing program; Fable 5 briefly went offline in mid-June under U.S. export controls before access was restored July 1.

Frequently Asked Questions