AI NEWS · MODELS

Anthropic Launches Claude Sonnet 5

Anthropic's most agentic Sonnet model to date arrives with near-Opus performance, same-day availability across every major platform, and launch pricing that undercuts its predecessor.

30 JUNE 2026 · MODELS

Read the Story

A Same-Day Launch Across Every Major Platform

Anthropic released Claude Sonnet 5 on 30 June 2026, making it available simultaneously across the Claude consumer apps, Claude Code, the Claude Developer Platform API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. That breadth of same-day availability is itself notable: rather than staging a model through a narrow developer preview before wider release, Anthropic pushed Sonnet 5 to essentially every channel where organisations already consume Claude at once.

Sonnet 5 also became the new default model for Claude's Free and Pro plans immediately, meaning the bulk of Anthropic's consumer user base was upgraded automatically rather than needing to opt in.

Anthropic's Most Agentic Sonnet Model to Date

Anthropic describes Sonnet 5 as its most agentic Sonnet model yet, built for autonomous multi-step tool use, working across browsers, terminals, and file systems without a human directing each individual step. That framing matters because it positions Sonnet, historically the mid-tier model in Anthropic's lineup, as a genuine option for the kind of long-running agentic work that was previously the preserve of the flagship Opus model.

Anthropic states that Sonnet 5's overall performance now sits close to Opus 4.8 on a range of benchmarks, at a fraction of the cost, a positioning that echoes the industry-wide trend toward making near-frontier capability available through cheaper, faster mid-tier models rather than reserving it exclusively for the most expensive tier.

How It Benchmarks Against GPT-5.6 and Gemini

On SWE-bench Pro, a benchmark for real-world software engineering tasks, Sonnet 5 scored 63.2% against GPT-5.6 Sol's 64.6%, a result close enough to be considered essentially tied, particularly given that OpenAI's own 8 July 2026 audit found roughly 30% of SWE-bench Pro tasks to be flawed in some way, which introduces meaningful uncertainty into any comparison built on that benchmark alone.

On OSWorld, a benchmark for computer-use tasks, Sonnet 5 scored 81.2% against GPT-5.5's 78.7%. In blind human evaluations, developers preferred Sonnet 5's code output 47% of the time, compared with 29% for GPT-5.4 and 24% for Gemini's comparable model, though it's worth noting that comparison used GPT-5.4 rather than the newer GPT-5.6, so a like-for-like comparison against OpenAI's latest release was not yet available at launch.

Google's Gemini 3.5 Flash, for its part, leads on the MCP Atlas benchmark at 83.6%, and runs roughly four times faster (289 tokens per second) at around half the cost of Sonnet 5, a reminder that raw capability, speed, and cost remain three separate axes on which no single model currently leads across the board.

Launch Pricing, Then a Return to Standard Rates

Anthropic set launch pricing for Sonnet 5 at $2 per million input tokens and $10 per million output tokens, running through 31 August 2026, after which pricing reverts to Anthropic's standard rate of $3 per million input tokens and $15 per million output tokens.

That two-month discount window gives organisations a concrete incentive to trial Sonnet 5 in production workloads early, while Anthropic gathers real-world usage data ahead of the model settling into its standard, higher price point.

What This Means for Organisations Evaluating Model Choice

For organisations already standardised on Claude, Sonnet 5's near-Opus performance at Sonnet-tier pricing is a straightforward reason to re-evaluate which tasks still require the more expensive Opus model versus which can now be handled adequately, and considerably more cheaply, by Sonnet 5.

For organisations evaluating providers more broadly, the picture that emerges from this week's cluster of releases is one of rough parity at the frontier: GPT-5.6, Claude Sonnet 5, and Gemini 3.5 Flash each lead on different benchmarks and different cost/performance trade-offs, which reinforces the case for evaluating models against an organisation's own representative tasks rather than relying on any single published leaderboard.