The frontier model race moves to price

Anthropic priced Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens. OpenAI lists GPT-6 Sol at $2 and $10 for standard short-context input and output; Luna costs $0.10 and $0.50. Axios reports that lower unit prices can expand usage enough to raise total spending, supporting both adoption and lab revenue.

China rejects the proposed slowdown terms

Rest of World

Pre-summit reporting finds Chinese officials and engineers treating Dario Amodei's proposal as an attempt to preserve the US lead. The dispute concerns who sets the terms of any coordinated slowdown.

AI safety is on the Trump-Xi agenda

CNBC

This pre-summit account says AI safety will be part of the September 24 meeting, while both governments continue to prioritize capability and national advantage. It sets a narrow expectation for cooperation rather than a broad agreement to slow development.

Most AI incidents had no usable stop

Oren Perez

The study reviewed 1,400 AI incidents and retained 1,213 cases. About 80% had no usable stop. It separates the problem into technical affordances, interruption authority, epistemic triggers, and epistemic standing; in cases without a usable stop, the missing element was legal four times in five.

Banks expect more fraud and disputes from agentic commerce

Techmeme / Gizmodo

Six major banks, including Bank of America and Capital One, warn that more autonomous commerce agents could increase scams, fraud, and disputes. The concern places payment accountability beside the agent's ability to act.

Rocket adds a plan checker for mainframe agents

SiliconANGLE

Rocket Software's forthcoming EVA 2.0 adds PlanGuard, which evaluates an agent's plan before it reaches critical mainframe resources. The product allows investigation and eventual automation while preserving a separate approval boundary.

Ema raises $77 million for autonomous enterprise workers

SiliconANGLE

Ema raised a $77 million Series B to expand autonomous workers across human resources, IT, and finance. Capital is moving into agent deployment while banks and infrastructure vendors add separate checks around what those agents can do.

China becomes the top destination for elite AI talent

Techmeme / The Information

Carnegie China data says China's share of elite AI talent rose from 27.1% in 2022 to 40.6% in 2025. Over the same period, the US share fell from 46.4% to 34.2%. The reported crossover reverses the prior direction of travel for this measure.

Alibaba plans three new cloud regions

Techmeme / Bloomberg

Alibaba plans its first cloud regions in Turkey, Finland, and the Netherlands over the next 12 months. The expansion takes Chinese cloud infrastructure farther into Europe and the Middle East during heightened US-China technology tensions.

ByteDance supplied nearly 75% of Nscale's 2025 sales

Techmeme / Financial Times

Financial Times reporting based on sources and US filings says ByteDance used Nscale's Norway facility to access Nvidia chips. It accounted for nearly three quarters of Nscale's 2025 sales, connecting Chinese model demand to offshore compute capacity and concentration risk for the provider.

The whiteboard defense for agent-written systems

The PrimeTime

The proposed standard is that an engineer should be able to explain a shipped system and defend its data structures, failure cases, and trade-offs. That level of understanding is broader than line-by-line recall and stricter than accepting a successful run.

Harness-Zero moves harness behavior into model weights

Haoran Ye et al.

A supervising agent converts guidance from an optimized harness into demonstrations in the target harness's action space. The paper reports macro-average task success rising from 23.3% to 44.3% after distillation. That result exceeds the 41.7% achieved with the specialized harness still attached.

A robot backdoor hides in foot-ground contact

Jiageng Zhang, Doniyorkhon Obidov, and Kaichen Yang

StepTrigger uses pressure and foot-contact patterns from dense terrain to activate a compromised robot planner, outside the text, vision, and action-history channels most monitors inspect. In offline tests, it preserved 98.75% of clean behavior and rejected 92.50% of false triggers. It activated on 76.25% of true triggers.