Anthropic shipped Claude Sonnet 5 on June 30, scoring 63.2% on its agentic coding benchmark against Sonnet 4.6's 58.1% and Opus 4.8's 69.2%. That's a five-point lift over the prior Sonnet and within six points of the flagship, at an introductory API price of $2 per million input tokens and $10 per million output, well under what Opus 4.8 costs and under OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, though still above Gemini 3.5 Flash.

The framing Anthropic wants is captured cleanly in its own launch post: "Between Sonnet 5 and Opus 4.8, users can adjust the effort level to find the right balance of cost and performance." Opus is still recommended for "subtle judgment calls and deep research." Everything else, the company is plainly steering toward Sonnet. It's now the default for Free and Pro plans, available to Max, Team, and Enterprise tiers, and live in Claude Code and on the Claude Platform. On an unnamed knowledge-work benchmark Anthropic cites, Sonnet 5 slightly beats Opus 4.8 outright.

The cost story is more complicated than the sticker. Sonnet 5 uses the same updated tokenizer as Opus 4.7, and Anthropic concedes "the same input can map to more tokens: roughly 1.0–1.35× depending on the content type." The introductory pricing, the company says, is calibrated to make the transition "roughly cost-neutral" against Sonnet 4.6. Translation: the headline cut absorbs the tokenizer inflation. On August 31 the price moves to $3/$15, and the cost-neutral claim no longer applies.

Field reports track the benchmark. Daniel Shepard, senior engineer at Zapier, told TechCrunch a two-part Salesforce-plus-outreach workflow that prior models "stalled halfway" on, Sonnet 5 "finished end to end." That's the agentic story Anthropic is selling, and at least one customer is buying it on the record.

Safety posture is where the release gets interesting. The system card classifies Sonnet 5 as posing "very low alignment risk, though higher than for previous Sonnet models," and notes it doesn't cross Anthropic's CB-2 or automated AI R&D thresholds. The model wasn't deliberately trained on cybersecurity tasks, and Axios reports it has "much lower ability" to perform dangerous cyber activities than current Opus models. Real-time cyber-misuse safeguards are on by default, though "less strict than those launched with Fable 5." Mythos is back on a limited basis; Fable 5 is "on track to return soon" after the Trump administration asked Anthropic to take both models down over security concerns.

Yahoo Finance lands the financial frame: the launch arrives weeks after Anthropic's June 1 confidential SEC filing. The pricing of Sonnet 5 reads, in that light, less like a generational upgrade than a margin instrument timed to a road show. Cheaper agents are easier to sell to enterprises. Cheaper agents at introductory pricing, with a tokenizer that quietly raises the effective rate, are easier to sell twice.

Reviewed: Claude Sonnet 5 closes the Sonnet-to-Opus gap at $2/$10 introductory

Sources