This Week in AI: Claude Sonnet 5, LongCat-2.0, and the AI Agent Act

The week of June 29 through July 6 was defined by three overlapping threads: frontier labs racing to ship more capable agentic models, governments reasserting control over who can access the most powerful systems, and open-weight releases that stretched what developers can run or fine-tune themselves.

Models

Introducing Claude Sonnet 5 — Anthropic positioned Sonnet 5 as its most agentic mid-tier model yet, with stronger planning, tool use, and coding that approaches Opus-class performance at a lower price. It is now the default for free and Pro users, with introductory API pricing through August.

Redeploying Fable 5 — After U.S. export controls briefly cut off access to Claude Fable 5 and Mythos 5, Anthropic announced the restrictions were lifted and began restoring global availability starting July 1. The episode underscored how quickly policy can interrupt production AI deployments.

Leanstral 1.5: Proof Abundance for All — Mistral released an Apache-2.0 Lean 4 proof-engineering model that saturates miniF2F and reports strong PutnamBench results. The team also highlighted real bug discoveries in open-source Rust code through automated verification pipelines.

LongCat-2.0 on Hugging Face — Meituan published the weights for its 1.6-trillion-parameter MoE coding model under the MIT license, completing a rollout that began with a June 30 announcement. The model targets long-horizon agentic coding with a native 1M-token context window.

Agents & Tools

Claude Science, an AI workbench for scientists — Anthropic launched a dedicated research environment that unifies literature review, analysis pipelines, auditable artifacts, and flexible compute across genomics, proteomics, and related domains. It runs existing Claude models rather than gating a separate frontier release.

Research

Introducing GeneBench-Pro — OpenAI published a 129-problem benchmark designed to test judgment-heavy computational biology, from messy datasets to revised analytical plans. Even top frontier models solved fewer than one-third of tasks, highlighting how far agents remain from autonomous research-grade analysis.

Policy & Regulation

Warner unveils AI AGENT Act discussion draft — Senator Mark Warner released a framework to guarantee consumer AI agents fair access to major online platforms, create an FTC registry of trusted agents, and task NIST with interoperability standards. The draft signals growing congressional attention to agent market structure, not just model safety.

More details on Fable 5's cyber safeguards and jailbreak framework — Following Fable 5's global return, Anthropic published deeper documentation on its cyber safeguards and proposed an industry-wide jailbreak severity scoring framework with Amazon, Microsoft, Google, and other Glasswing partners.

Next week, watch whether GPT-5.6 moves beyond its limited preview, how quickly Anthropic restores Fable 5 across cloud partners, and whether open models like LongCat-2.0 and Leanstral 1.5 gain traction in production agent workflows beyond benchmark headlines.