Lead story / launch

Google opens Gemini for robot control, but reliability remains unclear

Google opened Gemini 4 Argon to developers, businesses and paying consumers on October 13, pitching one model for robot control, virtual practice environments and training-data evaluation. In one demonstration, Argon builds a virtual trash bin so a Boston Dynamics Spot robot can practice balancing on three legs before trying the pedal on real hardware. Two additional AI agents check its foot and gripper positions.

Another demonstration starts with a photo of party props: Argon creates a virtual replica, devises 77 tasks and generates more than 4,500 training episodes for a robot-control model. Google reports improvements on a real robot, but does not disclose the attempt counts behind the success rate. Human guidance also remains substantial: a cup-righting example uses a camera image with a centimeter grid and ready-made movement commands.

The benchmark results underline the distinction between reasoning about physical tasks and performing them. Google reports an 87% score on its 400-question ERQA test, ahead of Claude Opus 5.5 and GPT-6 Astra. But on Artificial Analysis’s benchmark involving direct control of a real robot arm, Argon placed third with 55 out of 100 points, behind GPT-6 Astra at 69 and GPT-6.1 Sol at 65. These are different tests of different skills.

Google AI Ultra subscribers can access Argon in AI Studio and Antigravity, while developers can use the paid Gemini API. PYMNTS reports introductory pricing of $2 per million input tokens and $10 per million output tokens, doubling after November 30. The demonstrations connect simulation, action and feedback into one workflow, but missing attempt counts and unpublished individual failure rates leave its reliability unclear.

models

TIME recognizes Meta’s speech tool for more than 1,600 languages

TIME named Meta’s open-source Omnilingual ASR to its Best Inventions of 2026 list. Meta reports support for more than 1,600 languages, including over 500 never previously served by speech recognition, and says communities can add languages with a handful of samples. Nigerian health practitioners are already using it in community clinics, though TIME reports no measured patient outcomes.

benchmark

Apple shows four linked Macs running an AI model too large for one

Apple’s MLX benchmarks show four Mac Studios generating more than 36 tokens per second from Muse Spark Open. The compressed model occupies about 493GB, but its working memory pushes it beyond a single 512GB Studio. Sharing work can speed generation while communication overhead slows prompt processing. For a long input, doubling Thunderbolt cables per link cut the initial wait from about 50.5 to 40.6 seconds.

open release

NVIDIA releases open-weight AI for business documents and local agents

NVIDIA released Super VL for document-heavy workflows and Bolt 74B-A9B for local agents. NVIDIA claims task-specific wins over named Claude and GPT rivals, but VentureBeat says it supplied no task definitions, sample sizes or protocols. Bolt’s compression lets it fit in 24GB of memory; a smaller E4B version is due October 23. The OpenMDW license permits commercial use, fine-tuning and redistribution.

Also worth knowing

•  Anthropic removes the waitlist for Claude Opus 5.5 fast mode

•  Rivian explores outside AI assistants, with steering and braking off-limits

•  Microsoft pairs junior engineers with mentors to learn AI-assisted development

•  Airbnb adds AI support chat in more than 50 languages

•  Worker sues Sutherland, alleging firing after request for fewer AI webcam checks

•  OpenAI offers eligible U.S. university labs up to 30 free AI seats

•  Bluesky expands Attie to build websites and games for communities

•  Jensen Huang urges NVIDIA to work with Chinese AI firms on safety

•  USC adds oral quizzes as students use AI for engineering projects

•  Volvo expands NVIDIA partnership to automate driving beyond electric cars

•  OpenAI previews ChatGPT Sites controls for private-network access

•  Google makes website links more prominent in AI search answers

•  Meta releases downloadable Muse Spark Open for coding and multistep work

•  Amazon launches a $74.99 Bee wearable for conversation notes

•  Small study finds patient feedback cuts false seizure alerts in simulation

Daily tool drop

6 AI tools worth knowing today

The Internet Had a Point

Gemini Argon Runs Out of Tokens

Reader check-in

Help shape tomorrow's briefing

One click tells us what to keep, improve, or tighten.