Claude Sonnet 5 Released: Near-Opus 4.8 Performance, Major Agent Capability Gains

📖 Glossary

AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

Abstract: On July 1, 2026, Anthropic released Claude Sonnet 5, positioned as "the most agentic Sonnet model yet." It significantly surpasses the previous Sonnet 4.6 across reasoning, tool use, coding, and knowledge work, with multiple metrics approaching the flagship Opus 4.8 — at roughly one-third the cost. Now set as the default Claude platform model, available to all free and paid users.

What It Is: The Most Agentic Sonnet Yet

Anthropic's positioning of Sonnet 5 is crystal clear: the strongest agent-capable model in the Sonnet lineup.

What does "agentic capability" mean? It means AI no longer just answers questions — it can independently plan, invoke tools like browsers and terminals, and complete multi-step tasks without human intervention. Tell it to "research competitors and write an analysis report," and it can search, organize data, and write the report autonomously.

This happens to be the most competitive direction in AI right now — not who answers more accurately, but who can actually get work done. Sonnet 5 is Anthropic's answer for the mid-tier market.

Claude Sonnet 5 performance comparison and agent capability upgrade diagram

Core Upgrades: Four Dimensions of Improvement

Reasoning

Sonnet 5 shows significant improvement in complex reasoning tasks. Anthropic states its overall reasoning ability approaches Opus 4.8 — their current flagship. This means you get near-flagship reasoning at mid-tier pricing.

Tool Use

This is Sonnet 5's biggest selling point. In agentic search evaluation BrowseComp and computer use evaluation OSWorld-Verified, it shows marked improvement over Sonnet 4.6, with some tasks approaching Opus 4.8 levels. It can autonomously invoke browsers to search, operate computer interfaces, and execute terminal commands.

Coding

On the SWE-bench Pro agentic coding test, Sonnet 5 scores 63.2%, up from Sonnet 4.6's 58.1%, approaching Opus 4.8's 69.2%. Cursor's co-founder reports that Sonnet 5 can "follow plans, adhere to specifications, and complete multi-step changes cost-efficiently."

Knowledge Work

On the GDPval-AA v2 knowledge work benchmark, Sonnet 5 scores 1618, directly surpassing Opus 4.8's 1615. This is one of the rare cases where a mid-tier model beats the flagship on an individual test.

Pricing: Flagship Performance at One-Third Cost

Sonnet 5 launches with introductory API pricing of $2 per million input tokens and $10 per million output tokens, valid through August 31. After the promotional period, pricing adjusts to $3 and $15.

For comparison, the flagship Opus 4.8 is priced at $5 per million input tokens and $25 per million output tokens. Sonnet 5's regular price is roughly 60% of Opus 4.8, and during the promotional period, as low as 40%.

This is Anthropic's strategy: near-flagship performance at significantly reduced prices to capture enterprise market share. Particularly noteworthy as Anthropic pursues its IPO, needing to demonstrate large-scale API revenue capability to public markets.

Sonnet 5 agent workflow and tool invocation diagram

Safety: Lower Hallucination Rates but Cybersecurity Gap Remains

Sonnet 5 shows lower hallucination and sycophancy rates than its predecessor, with stronger refusal of malicious requests. However, in the Firefox vulnerability assessment conducted with Mozilla, its partial success rate is 13.2% — higher than Sonnet 4.6's 8.8% but far behind Opus 4.8's 68.8%.

This indicates Sonnet 5 has improved in general safety but still lags the flagship in deep cybersecurity scenarios. Anthropic has enabled real-time cybersecurity protection by default to partially address this gap.

What It Means for Developers

Zapier's senior engineer reports: multi-part automation tasks that previously "stalled halfway" with older models can now be completed end-to-end by Sonnet 5. This reliability is exactly what enterprises need to move AI from pilot to production deployment.

If you're running OpenClaw on Kaihe AIBOX, Sonnet 5 is a cloud model worth integrating: local agents handle task scheduling and data processing, while Sonnet 5 API handles heavy reasoning and complex coding — at a fraction of Opus 4.8's cost with minimal performance gap.

A Cool-Headed Look

First, the promotional pricing ends August 31, after which costs increase 50%. Try it now while it's cheap.

Second, high benchmark scores don't guarantee real-world performance. BrowseComp and OSWorld are standardized tests; real business scenarios are more complex. Recommend small-scale validation before large-scale deployment.

Third, the cybersecurity capability gap with the flagship is significant. If your business involves security auditing, you still need Opus 4.8.

Fourth, this launch comes against the backdrop of Anthropic's IPO push — pricing strategy may shift. Don't treat promotional prices as long-term rates.

Claude Sonnet 5 Released: Near-Opus 4.8 Performance, Major Agent Capability Gains