Daily AI News — 19 Oct 2026

StepFun published the open weights of Step 5 Preview on Hugging Face and ModelScope on 15 October, days after a free promotional week pushed it to second place on OpenRouter's LLM leaderboard with 18.8 trillion tokens processed. The model is a sparse mixture-of-experts design with 600 billion total parameters and 27 billion active per token, accepts text, image and video, and handles up to a one-million-token context. StepFun positions it for software engineering and professional knowledge work, with particular strength in finance, and ships it under its own StepFun Model License rather than Apache 2.0.

Chinese open-weight labs keep taking the top of the usage charts, and a finance-tuned model teams can download and self-host changes the calculus for any firm that cannot send sensitive data to a closed API.

A trading or audit team can run Step 5 Preview on its own GPUs to draft financial analysis and automate multi-step research while keeping regulated data entirely inside its own infrastructure.

Meta's Superintelligence Labs released Muse Wonder and Muse Wonder Full on 12 October, the first models the company positions at the frontier of AI capability alongside OpenAI, Anthropic and Google. Where the earlier Muse Spark aimed for low-latency and light weight, the Wonder models are larger and score higher on complex tasks like coding. Both power Meta's Muse personal-agent app, starting with subscribers on its $20 and $100 monthly plans, with Wonder priced at $4/$20 per million input/output tokens on the Meta Model API. Wonder Full is reserved for Muse Code subscribers and is not served through the API.

Meta had fallen behind after reorganising its AI effort, so a credible frontier entry resets the competitive map and gives builders another serious vendor for agentic and coding workloads.

A product team building an in-app assistant can route coding and multi-step agent tasks to Muse Wonder through the Meta Model API, comparing its cost and quality against incumbents before committing.

California Governor Gavin Newsom signed Executive Order N-9-26 directing the state's Government Operations Agency to deliver recommendations by 16 November on tougher frontier-AI rules, including a mandatory shutdown mechanism, independent verification organisations embedded onsite in labs, and a wider definition of reportable safety incidents to cover loss-of-control events. The order accelerates two laws already on the books, SB 813 and AB 1405, with hard deadlines in 2027 for an audit framework and a registry of AI auditors. Coverage this week framed the move against Newsom's 2024 veto of a similar bill and his reported 2028 ambitions.

It is the clearest signal yet that a major US jurisdiction will treat frontier models like regulated infrastructure, and the 2027 deadlines bind the state whether or not the legislature acts.

A frontier lab or enterprise buyer in California can start planning for onsite auditors and a verified shutdown capability now, rather than waiting for the 2027 legislative session to force the change.

Anthropic has reworked Projects in Claude Code so a single conversation acts as a coordinator that breaks a goal into parallel threads, each a separate cloud Claude Code session on its own git branch that writes code, runs tests and opens a pull request before reporting back. Threads share a project memory index and instructions of up to 16,000 characters, keep running after the user's machine is off, and are capped at 200 new threads per day. The beta, a redesign away from the old 'folder of files and one chat' model, is open to Pro and Max subscribers, with wider access to follow.

It pushes multi-agent orchestration from a bespoke engineering project into a default product pattern, letting one person supervise a fleet of coding agents from a phone instead of babysitting each session.

An engineering lead can hand a project a list of bugs and let separate threads fix each one in parallel, opening PRs the lead reviews later, while retiring a deprecated endpoint across several repos at once.

DeepSeek has agreed to lead a seed round valuing Winverse AI at roughly $400 million, Bloomberg reported on 12 October. Winverse, founded this year by former Tencent vision-language and embodied-AI lead Hu Han, builds systems that let robots perceive and act in the physical world. Meituan's Long-Z Investments is also backing the round. The deal deploys part of DeepSeek's recently closed war chest of more than $12 billion and follows its strategic stake in humanoid maker Unitree Robotics, marking the language-model lab's most serious move into embodied intelligence.

The hardest problem in robotics is turning plain instructions into reliable physical action, not chip supply. A leading open-model lab funding that gap signals where the next wave of AI value may land.

A warehouse or manufacturing operator watching this space can expect general-purpose robot 'brains' from well-funded labs to reach pilots sooner, widening the pool of vendors beyond the handful of humanoid makers.

Promise Robotics, a Canadian physical-AI startup, announced on 17 October a multi-year deal with D.R. Horton, the largest US homebuilder, to produce more than 1,250 homes using its 'factory as a service' platform. The system ingests Horton's existing home designs and generates production plans and instructions for robotic arms that build framing and components like windows, doors and stairs in an Arizona factory before crews assemble them on site. It is the first such deal for Horton, which built about 85,000 homes last year, and extends Promise's earlier Canadian work with Mattamy Homes.

Homebuilding has lacked manufacturing-grade production infrastructure. Applying AI and robotics to existing designs promises faster, more predictable cycle times without each builder having to operate its own factory.

A regional developer facing labour shortages could use a factory-as-a-service model to prefabricate house components from its own plans, cutting build times while keeping its existing designs and standards.

Nolla Health received authorisation from the Utah Office of Artificial Intelligence Policy for what it calls the first US pilot allowing an AI to issue initial prescriptions, starting with acne treatment through its Nolla Derm app from $4.99 a month. Patients verify identity, give consent, complete a questionnaire and take a five-angle face scan, generating a personalised plan and, where appropriate, an initial prescription in 10 to 15 minutes. For the first 100 patients, two licensed physicians independently review and approve every AI-generated prescription; the app shifts to after-the-fact review only once agreement reaches 95%.

It is an early, carefully-staged test of AI prescribing under a state regulator, chosen in a low-risk, well-understood condition. How it is governed will shape how far automated clinical decisions are allowed to go.

Someone who cannot easily reach a dermatologist can get an assessed, physician-approved acne prescription from home in minutes, closing an access gap without an in-person visit.

Google made Gemini 4 Argon generally available on 13 October, calling it its most capable model, built for complex coding, agentic, math and science tasks and anchoring the Gemini 4 generation. Announced earlier in October, Argon is larger than Google's previous 'Pro' models; the company says it is comparable to frontier models from OpenAI and Anthropic on key coding and cyber benchmarks, though it trailed on some. Google has shifted its messaging toward cost advantage and shelved plans to release Gemini 3.5 Pro, after an overhaul of its DeepMind lab and leadership changes delayed the line.

A GA frontier model from Google, pitched on both capability and price, keeps competitive pressure on the whole market and gives builders another production-ready option for demanding agentic workloads.

A developer can switch a coding or research agent to Gemini 4 Argon via the Gemini API and benchmark its cost-per-task against rivals before standardising on a provider.

Replit CEO Amjad Masad said over the weekend of 17 October that the coding platform is no longer working with Anthropic as its default model provider and that OpenAI will replace it, though Claude will remain available to customers who choose it. Replit is one of the most prominent 'vibe-coding' platforms, where users build and ship apps largely through natural-language prompts, so its default model shapes the experience for a large base of non-expert builders. The move reflects how aggressively model vendors are now competing to win the platforms that sit between them and end users.

Default-model choices at big developer platforms move real usage and revenue, and they signal which vendor a platform judges best on price, quality and reliability for day-to-day building.

A team picking a build-and-ship platform should check which model is the default and whether others are available, since that choice drives the quality and cost of everything they generate.

Stablecoin issuer Tether has invested in AIsa, an extended seed round announced on 18 October that, with HIVEN also participating, takes the company's total funding past $9 million. AIsa is building payment and spending tools designed specifically for AI agents, the plumbing an autonomous agent needs to search for a capability, pick a service and pay per use without a human opening accounts and managing API keys for every tool. The round lands amid a wave of interest in agent-commerce infrastructure, from checkout marketplaces to runtime controls that authorise and audit what agents are allowed to spend.

Agents that can act are only useful if they can transact. Building safe, auditable payment rails for them is becoming its own infrastructure layer, and stablecoin issuers see an obvious role in it.

A developer building an autonomous shopping or research agent could fund one balance and let the agent pay per use across many tools, instead of pre-provisioning accounts and keys for each service.