AI Trends — 2 July 2026

Your briefing on the most important AI developments from the last 48 hours.

🔥 Top Story: Claude Sonnet 5 Launches as Default for Every Free and Pro User

Anthropic launched Claude Sonnet 5 on June 30 and made it the default model for every Free and Pro user worldwide starting July 1 — its biggest mass-market model launch of the year. Anthropic calls it "the most agentic Sonnet ever built," performing close to flagship Opus 4.8 on many tasks at introductory pricing of $2/$10 per million tokens (input/output) through August 31 — cheaper than Sonnet 4.6. Benchmark highlights include 63.2% on agentic coding (vs. Opus 4.8's 69.2%), 81.2% on OSWorld-Verified desktop automation, and an eye-catching 80.4% on Terminal-Bench 2.1 — a 20.7-point jump over Sonnet 4.6. Notably, adaptive thinking is now always on, sampling parameters like temperature have been removed, and the new tokenizer produces 1.0–1.35x more tokens than before, so existing integrations need migration work before switching production traffic.

🏛️ California Signs Landmark 50%-Off Anthropic Deal for All State Agencies

Governor Gavin Newsom announced on June 29 the largest US state government AI deployment in history: every California state agency and any opted-in city or county can now access Claude at a 50% discount through the new Statewide IT Shared Services (SITeS) portal, plus free workforce training and Anthropic technical consultation. The deal builds on existing programs including Poppy, a state-worker-built AI assistant piloted across 2,800 employees in 67 departments and headed for full statewide rollout this month, along with active Claude deployments at the DMV, Department of Healthcare Services, and CalOES.

⚠️ Five Eyes Alliance: AI-Powered Cyberattacks Are "Months Away, Not Years"

The Five Eyes intelligence alliance (US, UK, Australia, Canada, New Zealand) issued a rare joint statement warning that frontier AI models are expected to fundamentally transform offensive and defensive cyber capabilities imminently. Reuters and CyberScoop tied the warning to Claude Mythos and OpenAI's GPT-5.5-Cyber; a related report claims an Anthropic AI agent penetrated nearly all classified NSA and US Cyber Command test systems within hours during an assessment. CISA has already cut the mandatory federal patch deadline to three days, citing AI-driven threats directly.

🐛 AI Finds a 29-Year-Old Bug: Squidbleed (CVE-2026-47729)

Claude Mythos 5, working under Anthropic's Project Glasswing defensive research program, discovered Squidbleed — a 29-year-old memory-leak vulnerability in the widely deployed Squid proxy server that exposes user HTTP credentials to network-adjacent attackers. Glasswing has now surfaced 23,019 vulnerabilities across 1,000 open-source projects in its first month, with a 90.6% confirmation rate on independent sampling — a strong early signal for AI-assisted vulnerability research at scale.

🖼️ Google Ships Two New Gemini Image Models

Google released Gemini 3.1 Flash Image ($0.50/$3.00 per million tokens) and Gemini 3 Pro Image ($2.00/$12.00) on June 30, both live now in Google AI Studio and the Gemini API. The releases fill out Google's model lineup while the long-awaited Gemini 3.5 Pro continues to slip — now expected sometime in July after Google cited enterprise feedback about excessive token consumption in extended agentic tasks.

🔓 Export Controls Lifted on Claude Fable 5

The US Department of Commerce lifted export controls affecting Claude Fable 5, with global access restoring on July 1. Fable 5 is available at 50% of weekly usage limits through July 7 for Pro, Max, Team, and some Enterprise plans, after which it moves to a paid usage-credits system.

🚨 Security Watch: Critical LiteLLM RCE Added to CISA's Known Exploited Vulnerabilities

CISA added CVE-2026-42271, an unauthenticated remote-code-execution chain in LiteLLM's AI Gateway, to its Known Exploited Vulnerabilities catalog on June 27. The flaw exploits MCP endpoints to expose all API keys configured on the gateway. Organizations running LiteLLM should patch immediately (fix released June 20), rotate all API keys, and audit access logs predating the patch.

📊 Also This Week

Claude Science: Anthropic launched a dedicated app for drug discovery, protein analysis, and computational biology, building on its Coefficient Bio acquisition and the hire of AlphaFold lead John Jumper.
Claude on Azure GA: Claude is now generally available on Microsoft Azure AI Foundry, running on NVIDIA GB300 Blackwell Ultra GPUs.
Grok 4.3 on Amazon Bedrock: xAI's mid-tier model landed on Bedrock at $1.25/$2.50 per million tokens — the cheapest major-lab model on the platform.
GPT-5.6 Sol on Cerebras: OpenAI confirmed plans to deploy GPT-5.6 Sol on Cerebras wafer-scale hardware in July, targeting ~750 tokens/second — roughly 15x current GPU serving speeds.

What to Watch

Gemini 3.5 Pro: Still no confirmed July date after another delay — watch for GA announcements.
Sonnet 5 migration fallout: Will breaking changes (no temperature control, new tokenizer) cause friction for enterprise developers before the introductory pricing window closes August 31?
AI-driven cyber incidents: Following the Five Eyes warning, watch for the first publicly disclosed AI-orchestrated attack.
Triple IPO season: SpaceX is already public; OpenAI targets a September listing and Anthropic October — both companies are racing to strengthen their numbers beforehand.

Sources: Build Fast with AI · TechCrunch · VentureBeat · CBS Sacramento · Cybersecurity Dive · Computer Weekly · llm-stats.com