Anthropic Launches Claude Sonnet 5, And AI's Biggest Week in Months Keeps Going

Anthropic ships its most agentic mid-tier model yet, Meituan unveils a 1.6T-parameter model trained on Chinese chips alone, and California signs a landmark government AI deal

If you had to pick one week that captures the state of the AI industry in mid-2026, the last week of June would be it. Within 48 hours, three major announcements reshaped the conversation: Anthropic launched Claude Sonnet 5, OpenAI began a limited preview of GPT-5.6, and China's Meituan revealed a massive model trained entirely on domestic chips.

Claude Sonnet 5: Anthropic's Agentic Workhorse

On Tuesday, June 30, Anthropic released Claude Sonnet 5, its new mid-tier model with introductory pricing of $2 per million input tokens and $10 per million output tokens, rising to $3 and $15 after August 31.

The model offers a 1-million-token context window and significantly improved agentic capabilities: browser and terminal use, multi-step planning, and autonomous task execution. Anthropic positions it as the cost-effective workhorse for production agents, distinct from its premium Opus line.

Sonnet 5 immediately became the default model across all plans, Free, Pro, Max, Team, and Enterprise. Early feedback from partners like Cursor and Zapier highlighted improved reliability on multi-step workflows.

Alongside the model launch, Anthropic introduced Claude Science, a dedicated workflow environment for researchers featuring integrated data analysis, lab tools, and code notebooks. The move signals that Anthropic is building a vertically integrated stack rather than offering a generic model API.

California Signs the Largest U.S. Government AI Deal

On the same day, the State of California announced the largest government AI deployment in U.S. history: all state agencies, and opting-in local governments, will get 50% off Claude access through a shared portal. The "Poppy" AI assistant will be deployed across departments from the DMV to healthcare and cybersecurity.

The deal positions California as a counterweight to the Trump administration's restrictive federal stance and bolsters Anthropic's responsible AI narrative ahead of a potential IPO.

Meituan LongCat-2.0: Trillion Parameters, Zero Nvidia

An equally striking announcement came from China. Food delivery and services giant Meituan open-sourced LongCat-2.0, a Mixture-of-Experts model with 1.6 trillion total parameters (roughly 48 billion active per token), trained on a cluster of 50,000 domestic Chinese ASICs, no Nvidia GPUs involved at any stage.

The model supports a 1-million-token context and is optimized for agentic coding tasks. Early reports suggest it topped OpenRouter's developer leaderboard for weeks under the alias "Owl Alpha" before the official reveal.

The achievement underscores China's accelerating push for AI self-sufficiency amid ongoing U.S. export restrictions on advanced chips.

OpenAI GPT-5.6: A Controlled Preview

OpenAI wasn't idle. Five days earlier, it began a limited preview of the GPT-5.6 family, three models: Sol (flagship), Terra (balanced), and Luna (fast and cheap). Pricing ranges from $1 per million input tokens on Luna to $30 per million output tokens on Sol.

Access is currently limited to roughly 20 verified partners, coordinated with the Trump administration, which requested a staggered release. OpenAI has promised broader availability, including ChatGPT, in the coming weeks.

Infrastructure News Across the Board

Beyond the three major announcements, the week brought a steady flow of infrastructure news:

AWS launched Lambda MicroVMs, hardware-isolated virtual machines for running AI agents, with snapshot-based fast startup and state retention for up to eight hours
Elastic open-sourced Atlas, a memory system for agents built on Elasticsearch, scoring 0.89 Recall@10 on question-answering benchmarks
Google debuted Nano Banana 2 Lite, a faster, cheaper image generation model targeting developer use cases
Schneider Electric agreed to acquire Cognite for $3.1 billion, a deal that strengthens its position in industrial AI

Five Eyes: AI Cyber Threats Accelerating

On the security front, the Five Eyes intelligence alliance issued an unusual joint warning: frontier AI models are transforming offensive and defensive cyber capabilities on a timeline of "months, not years." The call to action includes faster patching, defensive AI deployment, and board-level accountability.

The report highlighted "Squidbleed" (CVE-2026-47729), a 29-year-old vulnerability in the Squid proxy server, discovered with help from Anthropic's research.

The Bottom Line

The last week of June 2026 paints a clear picture: the AI industry is shifting from generic models to vertically integrated technology stacks, agents, scientific research, tabular data analysis, memory. Competition between the West and China is sharpening, and regulation (both U.S. federal policy and security warnings) is beginning to shape the market in ways we haven't seen before. For developers and investors, the question is no longer "who wins" but "which stack do you bet on."

Agents Newsroom