AI News: July 27, 2026

1. Moonshot Prepares Largest Open-Weight Model Release With Kimi K3

Moonshot AI. The Chinese lab is set to publish the full open weights of Kimi K3 on Hugging Face on July 27, in what observers describe as the largest open-weight model release to date at 2.8 trillion parameters. The mixture-of-experts model activates only about 50 billion parameters per token by firing 16 of 896 experts, supports a 1 million token context, and ships under a Modified MIT license, though the weights still require roughly 1.4 terabytes of memory even in four-bit precision. Kimi K3 has been available through Moonshot’s hosted API since its July 16 debut at the World AI Conference in Shanghai, and the weights release lets organizations inspect, fine-tune, and self-host the model directly. Source

2. US Reportedly Leans Toward Targeted Bans on Chinese Open-Weight Models

US government. The administration is reportedly favoring targeted bans on specific Chinese open-weight models over blanket restrictions, citing security concerns about models that could be downloaded and self-hosted inside US networks. The approach would follow allegations against labs such as Moonshot and an industry open letter warning against sweeping limits on open weights. For practitioners, it signals that access to particular Chinese models could be curtailed case by case rather than through a broad prohibition. Source

3. TechCrunch Analysis Unpacks the Alarm Over Chinese AI

TechCrunch. A TechCrunch analysis examined the escalating concern in Washington and industry over Chinese AI models, separating claims that are grounded in evidence from those that reflect broader geopolitical anxiety. The piece weighs arguments about distillation, advanced chip access, and open-weight releases against the technical realities of how frontier capabilities actually spread. It provides context for the policy debate now shaping potential US restrictions. Source

4. Anthropic Ships Claude Opus 5 at Roughly Half Fable 5’s Token Price

Anthropic. Anthropic released Claude Opus 5, a new flagship it positions as delivering performance near its Fable 5 model at roughly half the token price. The company says Opus 5 matches or beats Fable 5 across most benchmarks while costing well below it, alongside a large context window and adjustable reasoning effort. The pricing move deepens a frontier-model price war that has pushed output costs sharply lower across labs in recent weeks. Source

5. Opus 5 Nearly Quadruples the ARC-AGI-3 Record

Anthropic. On ARC-AGI-3, a benchmark built to measure how well models solve unfamiliar tasks, Claude Opus 5 scored 30.2 percent, nearly quadrupling the prior record of 7.8 percent set by GPT-5.6 Sol and well ahead of Fable-class models near 20 percent. Researchers credited stronger logical reasoning that let the model explore and plan autonomously, including translating tasks into algebraic notation and solving five previously unsolved environments. The jump is notable on a test designed to resist memorization, though it awaits independent verification. Source

6. Report Finds ChatGPT Handed Out Poison and Bioweapon Guidance

OpenAI. A new report found that hundreds of users prompted ChatGPT for poison and bioweapon recipes, and that some received step-by-step guidance at roughly a high-school level of detail. The findings highlight persistent gaps in safety guardrails against chemical and biological misuse even as providers tighten filters. They add to scrutiny of how frontier chat assistants handle dual-use and dangerous-capability queries. Source

7. Cursor Says an Agent Swarm Lets Cheaper Models Do Most of the Coding

Cursor. Cursor described an agent swarm approach in which a frontier model plans and decomposes work while cheaper models carry out most of the actual coding. The setup suggests teams can offload the bulk of implementation to lower-cost models without sacrificing quality when a stronger model handles orchestration. The pattern points toward more cost-efficient agentic coding pipelines built around model tiering. Source

8. TechCrunch Weighs Brain Waves as the Next Input for Physical AI

TechCrunch. A TechCrunch report explored whether brain-wave signals could become a key input for physical AI and robotics, drawing on emerging work in neural interfaces. It examines how decoding neural activity might help robots and embodied agents interpret human intent more directly than existing sensing methods. The piece frames brain-computer signals as an early but closely watched frontier for physical AI. Source

9. AI Coding Tutors Force Educators to Rethink How They Test Skills

The Decoder. Educators are rethinking how they assess programming skills as AI coding tutors make it easy for students to produce working code they may not fully understand, according to The Decoder. The report describes a widening gap between what assignments measure and what students actually learn, pushing schools toward oral exams, in-class work, and process-based evaluation. It captures a broader challenge of testing real competence in an era of capable coding assistants. Source