GitHub's Copilot Fusion now picks and combines AI models for each coding task

GitHub says Copilot Fusion is now generally available. You don't pick a model. Copilot decides which models to use for each prompt, and you choose an effort level: Standard, High or Max. It can answer with one model, let a cheap model draft and pass the work to a stronger one if a quality check fails, or have one model review another's draft. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task. You pay each model's standard rate, and paid plans get a 30% discount for a limited time.

Google prices Gemini 4 Carbon, its new AI model for security teams

VentureBeat reports that Google is rolling out Gemini 4 Carbon to selected cybersecurity defenders through Fairwind. Google calls it its latest and most intelligent frontier AI model, with improvements in cybersecurity. Access costs $2 per million input tokens and $10 per million output tokens for now. Prices will later rise to $4 and $20.

Developer builds a working computer processor inside the weights of Qwen's AI model

A developer released Qwenputer, a version of the small AI model Qwen 3.5 4B with a 32-bit RISC-V processor built into its weights. Weights are the numbers that make up a model. No training was involved: a compiler sets the weights directly and adds 11 layers in front of Qwen's 32. The model's memory layers serve as the processor's RAM. The author says it passes the official RISC-V tests, can run C and MicroPython code, and ran 18 million instructions without an error. It runs entirely inside the model, with no outside tools. Another user has since got a basic version of Doom running on it, slowly. The model is on Hugging Face and the code is on GitHub.

DoorDash moves its earnings report forward amid reports it will buy Instacart for $17B

Bloomberg reports that food delivery company DoorDash moved its third-quarter earnings release from November 4 to before the market opens on October 29. This comes amid reports of an imminent deal to buy grocery delivery app Instacart for about $17 billion.

AI coding startup Cognition reportedly gets offers valuing it at up to $80 billion

Business Insider reports, citing sources, that Cognition is weighing a new funding round before the end of the year. Investors have reportedly offered terms valuing it at $70 billion to $80 billion. That is up from $48 billion just weeks ago, in September.

AI data centres have pushed flash storage prices up fivefold, squeezing Apple

The Wall Street Journal reports that demand from big AI cloud companies has made flash memory prices five times higher. For years Apple pressed flash-storage makers into steep discounts. Now AI buyers are bidding up prices and taking much of the supply, which has shifted negotiating power to suppliers such as Kioxia.

US Homeland Security staff leak came from a claimed hack of its Okta login system

Wired reports that personal data of more than 41,000 employees and contractors at the US Department of Homeland Security has leaked online, including CBP and ICE agents. The hackers claim they got in through the department's Okta login system. The records include the locations of ICE and CBP facilities and workers' cell phone numbers.

Users say Claude Code spends about a fifth of its tokens writing agent status labels

A write-up by Flamerix AI says Anthropic's coding tool Claude Code sends each background agent's whole transcript to its model about every 30 seconds. It asks for a 3-to-5-word summary that only updates a status line. The authors measured this at 16.4% of Claude Code spending in one week, more than the main conversation, and cite four other measurements of 18–26%. They say there is no setting to turn it off and it has been on by default since version 2.1.232. They publish a workaround that blocks the timer.

Everything else