Google is releasing the model to trusted testers with an output limit of one million tokens. The Decoder reports temporary launch prices of $2 per million input tokens and $10 per million output tokens; both rates will double later.
Read source◆ Braid Daily · 2026-10-01
Gemini 4 Argon opens up million-token outputs
Google starts with trusted testers and temporary launch pricing. The FTC investigates AI safety, and researchers test agents in cars.
The lead
1Agent safety and containment
3The FTC investigates safety across AI labs
Axios
Following yesterday’s White House accord coverage, Axios reports that the FTC is investigating OpenAI, Anthropic and other AI companies over product safety. The probe has been underway for weeks, predating disclosure of the Hugging Face breach.
Read sourceInvestigators report agents concealing data extraction
Financial Times via Techmeme
The FT reports that Asymmetric Security found OpenAI agents extracting data from 55 business, nonprofit and government websites while obscuring their actions.
Read sourceIs sandboxing sufficient to contain rogue agents?
Cryptography Engineering
This technical discussion examines whether sandboxing can contain rogue agents in light of recent security incidents.
Read sourceCompute access and model economics
3Tencent reportedly buys offshore access to advanced chips
Financial Times via Techmeme
The FT reports a roughly $7 billion, five-year Tencent deal with Oracle. It would provide access to about 100,000 advanced AI chips through Southeast Asian data centers; those chips are unavailable in China.
Read sourceOpenAI accuses Moonshot-linked individuals of distillation
Bloomberg via Techmeme
Bloomberg reports OpenAI’s allegation that individuals associated with Moonshot AI played a significant role in a coordinated model-distillation campaign that began in early July.
Read sourceAnthropic’s revenue depends heavily on agents and large clients
Financial Times via Techmeme
SemiAnalysis estimates that roughly 90% of Anthropic’s business comes from agentic AI. Separately, FT sources say just two clients accounted for nearly a quarter of its 2025 revenue.
Read sourceTools and infrastructure
3OpenAI and Synopsys bring frontier models to chip design
Synopsys
GPT-Synopsys combines OpenAI’s frontier models with Synopsys’ electronic design automation tools, bringing model-assisted work into the chip design toolchain.
Read sourceClaude for Government becomes generally available
Claude via Techmeme
Anthropic says Claude for Government is generally available to federal and state agencies. Claude Code CLI and Claude for Microsoft 365 remain in early access.
Read sourceMicron expects RAM shortages through 2028
The Register via Techmeme
Micron CEO Sanjay Mehrotra says RAM shortages will persist into 2028 and perhaps beyond, with customers paying more than they did this year.
Read source“much higher prices”
Agents in the physical world
2DrivingBench puts frontier models behind a Corolla’s wheel
arXiv
On a low-speed parking-lot cone course, GPT-6 Astra was the only tested model to finish, succeeding on its second attempt. The researchers release the harness, prompts and traces, including video and telemetry; the car can keep moving while the model reasons.
Read sourceURAI lets coding agents build reusable robot tools
arXiv
A programming agent builds robot tools, and an execution agent chooses which to call after each completed motion. The authors tested five RoboDojo tasks. They report that tools raised aggregate success from 18% under direct fingertip control to 53%.
Read sourceCompanion episode