Claude Sonnet 5 anchors a wave of model and security moves

Anthropic made Claude Sonnet 5 the default model for Claude Free and Pro users starting July 1, 2026, pitching it as a lower-cost agentic model with near-Opus performance. The launch came with introductory $2/$10 pricing through August 31 and benchmark claims including 63.2% on agentic coding, 81.2% on OSWorld-Verified and 80.4% on Terminal-Bench 2.1. Developers face migration work because adaptive thinking is always on, sampling parameters were removed, and the new tokenizer can produce 1.0 to 1.35 times more tokens from the same text.

California signed a broad Claude deployment deal giving state agencies and participating local governments access at a 50% discount through the SITeS portal, alongside training and technical support. Existing uses include Poppy, Engaged California, DMV customer service, Medicaid case worker support, and cybersecurity workflows, with Poppy piloted across more than 2,800 employees and aimed at the state’s 300,000 workers.

Security dominated the wider agenda as the Five Eyes alliance warned that AI-powered cyber risks are months away, while Claude Mythos 5 was credited with finding Squidbleed, CVE-2026-47729, a 29-year-old Squid proxy memory leak. Other developments included GPT-5.6 Sol targeting up to 750 tokens per second on Cerebras, Claude becoming generally available on Azure AI Foundry, Claude Science launching for life sciences workflows, Gemini image models arriving in Google AI Studio, a LiteLLM RCE entering CISA’s KEV catalog, Grok 4.3 reaching Amazon Bedrock, and mounting IPO and wage-pressure narratives around frontier AI companies.