Google releases Gemini 4 Argon as Meta launches Muse Wonder

Google released Gemini 4 Argon on October 13, 2026. The model is available globally through the API, Google AI Studio, and Google Cloud; access in the Gemini app and Google Antigravity depends on subscription tier.¹²³ Google positions it as a model for complex tasks across software engineering, cybersecurity, finance, law, and science.² For its use in Antigravity, the company reports more than 35% more agentic code submissions and nearly one-third fewer code rollbacks over four weeks.²⁴

Those figures describe usage in a specific Google product, not a general improvement in developer productivity or a proven reduction in faulty commits.²⁴ The published information gives no sample size or controlled comparison group, so teams cannot assume they will see the same results without testing the model themselves.²⁴ For businesses, the broader availability matters: the model can be accessed through both development tools and cloud interfaces.²³

Meta introduced the Muse Wonder model family on October 12. The company describes Muse Wonder and Muse Wonder Full as the most capable models yet from Meta Superintelligence Labs. They are intended for tasks including coding and work across applications. Meta says “Active Memory” and “Active Focus” help the models retain relevant context and stay focused on a task’s goal. Muse Wonder is available through the Model API and Muse Code; Muse Wonder Full is limited to Muse Code subscribers and is not available through the API.¹²

Artificial Analysis gives Muse Wonder Full a score of 56 on its Intelligence Index, versus 52 for Muse Wonder. Full places in the top three on six of ten individual evaluations, but those results do not establish a second-place overall ranking.³ Muse Wonder costs $4 per million input tokens and $20 per million output tokens.³

Microsoft has reorganized Copilot: Home brings Chat and Cowork together with Word, Excel, and PowerPoint capabilities, while Code lets users create apps and workflows using natural language.¹² Autopilot is designed as a persistent agent that tracks tasks and handles recurring processes even when a user is not actively interacting with Copilot.¹² Microsoft published details of the redesign in early October. Home and Code were set to arrive first through the Frontier program, while Autopilot was planned for private preview.¹³

The change extends Copilot beyond individual chat requests toward document work, software creation, and delegated processes in one interface.¹² For companies, the practical question is how generated apps and long-running agents fit existing approval processes. Autopilot is intended to run in the cloud under organizational permissions, auditing, and governance controls.³ Microsoft did not announce general availability or specific pricing.¹³

Microsoft unveiled Maia 300 on October 12 as the third generation of its custom Azure AI inference platform. Production deployments are set to begin in 2027. Microsoft intends Maia 300 to become the primary inference system for its products and services and plans to expand Azure Maia to multi-gigawatt scale starting that year.¹²

Microsoft claims Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems. It also says a full rack can deliver up to seven times the token throughput of a Maia 200 rack.² The system is designed for inference workloads including long-context reasoning and multi-step agent tasks.² The GPU comparison has an important limitation: Microsoft does not identify the competing system or its configuration, and it has not published an independently reproducible benchmark.²

Anthropic introduced Scion on October 13 as an internally developed version-control system for its monorepo. The company has migrated the entire codebase from Git. The trigger was a sharp rise in merged code changes: Anthropic averaged about 2,500 per day in May 2026 and exceeded 30,000 per day by September. It attributes the increase to its development workflow involving many AI agents. Anthropic did not announce Scion as a product for outside users.¹

Scion relies on continuously available servers rather than complete local repository copies. Clients keep a bounded cache, while servers handle operations including commits, merges, and history queries. The architecture is intended to ease a source-control bottleneck that Anthropic identified alongside builds and continuous integration. It is tailored to a very large monorepo and automated workflows that remain online. The internal migration does not establish that Git is unsuitable for typical development environments.¹²

Liquid AI introduced LFM2.5-24B-A2B on October 13, 2026, calling it the largest and final model in its LFM2.5 family. It is designed for local reasoning and agent tasks on laptops and PCs. Base and post-trained weights are available to download, allowing users to run the model on suitable hardware.¹² It has 24 billion parameters in total but activates only about two billion per token through a mixture-of-experts architecture.¹

That design aims to limit computation during generation without reducing the model’s total size to the same degree.¹ Liquid AI reports performance figures on Apple and Nvidia hardware, but the available sources do not establish independent replication.² Local deployment is therefore a target use case, not a guarantee of usable speed on every laptop. An API is also available as an alternative to self-hosting.²

OpenAI opened the Decisions API to all developers in public beta on October 6.¹ Applications can submit text, images, or both through POST /v1/decisions.¹ The API handles three kinds of questions: estimating whether a condition is true, choosing among supplied options, and scoring inputs against ordered levels.¹ The beta supports only gpt-6-luna.¹

The endpoint is aimed at tasks such as classification and routing, where an application needs a structured decision rather than a free-form response.¹ OpenAI says it makes decisions up to 10 times faster than GPT-6 Luna through the Responses API; independent measurements supporting that comparison are not available.¹ Input costs $0.10 per million tokens.¹ Companies considering production use should check pricing and endpoint behavior as the beta develops.¹

The open-source project Strata, developed by GitHub user Niko1221, aims to run Qwen3.8-Flash-Next, a 125-billion-parameter model, on consumer PCs.¹² It spreads inference across GPU memory, system RAM, the CPU, and SSD storage.¹ The code is MIT-licensed, while model weights and other components have separate licenses.¹ Windows and Linux setup scripts are available, along with a local browser interface and APIs compatible with OpenAI and Anthropic formats.³

Strata could make large models easier to test locally, but it still requires substantial memory, storage, and suitable hardware.¹ Speeds published by the project reflect specific systems and quantization settings, not guaranteed performance on other PCs.² For sensitive data, companies should inspect setup scripts, model licenses, and network connections. Local inference alone does not establish that an installation never sends data externally.³¹