Meta Raises Its AI Game as Google Opens Gemini 4 Argon

 | 

Highlights

Meta and Google Ship Frontier Models

• Meta introduced Muse Wonder and Muse Wonder Full on Monday (Oct. 12), building on the Muse Spark model it released in April. Meta said they are its most capable yet, designed for personal and professional AI agents. A feature called Active Memory lets the models remember what matters to a user across conversations, while Active Focus keeps them working through blockers on long tasks. Meta tested them on tasks in law, finance and health, including the Legal Agent Benchmark from legal AI firm Harvey.
• Google opened Gemini 4 Argon to developers, businesses and paying consumers on Tuesday (Oct. 13). Google said it ranks first on the Vals Index, a test of work in finance, coding, legal and tax. It launched at $2 per million input tokens and $10 per million output tokens, an introductory price that doubles after Nov. 30. Tokens are the small chunks of text AI companies use to meter usage. Google said security firm Wiz has already used the model to find and patch flaws that put patient records and government data at risk.

Microsoft and Cloudflare Bet on Decision Models

• Microsoft introduced Microsoft-Decision-1 on Friday (Oct. 9), a model that decides instead of writes. Decision models do not generate text. They pick from a fixed set of options and return a confidence score. Microsoft built it for routing, classification and verification. The company said it runs 35 times faster than OpenAI’s GPT-6 Sol. Input costs $0.042 per million tokens, and output is free. Microsoft’s Xbox research team found it 200 times cheaper than GPT-6 Sol, and competitive on quality, for sorting more than 10,000 pieces of player feedback.
• Cloudflare released Clef-omni on Friday, a decision model that accepts audio, video, images and text in one call. A bank could feed it a recorded customer call and get a decision back without first converting speech to text. Cloudflare said it outscored its other Clef models on BANKING77, a test of how well a model identifies what a bank customer wants. The weights are free to download. Cloudflare also cut the price of its smaller Clef-flash model by more than half.

Liquid, Noiz and JetBrains Push Open Models

• Liquid AI, an MIT spinout that builds small models for phones and laptops, released LFM2.5-24B-A2B on Tuesday. The open model is built to run AI agents on a laptop. It has 24 billion parameters, the internal settings a model learns in training, but uses only about 2.5 billion at a time. That lets it run on a personal device rather than a data center. Liquid said it handles tool use and multi-step tasks privately on-device.
• Noiz AI, a voice and video generation startup, and the Hong Kong University of Science and Technology introduced Helix World on Monday. The model generates an explorable video scene and matching stereo sound from a single image, in real time. The code and a preview version are available to download.
• JetBrains, the Czech software company behind developer tools like IntelliJ, shipped Mellum2.1 on Oct. 8. The open 12-billion-parameter model is built for coding agents. It explores a codebase, edits files and checks its own work. Its Apache 2.0 license makes it free to use and modify.

The frontier is getting more crowded and more expensive. Decision models, meanwhile, are cutting routine calls to a fraction of a cent. The next wave of enterprise AI spending will be about matching each task to the cheapest model that can handle it. AT&T is already working that way. It cut the cost of coding and other advanced AI tasks by up to 56% by routing queries to cheaper models, PYMNTS reported in September.

For all PYMNTS AI and digital transformation coverage, subscribe to the daily AI and Digital Transformation Newsletters.