Mistral releases weights for Large 4 model

Mistral publishes the weights for Large 4, and OpenAI reportedly lowers its funding target to $20 billion. We also cover a backdoored Claude Code extension, Reddit's shift to logged-in access, Microsoft's buyouts and a $6.3 billion Binance hack.

AI lab releases

Mistral releases weights for its trillion-parameter Large 4 model

Mistral has published the weights of Mistral Large 4, the learned values behind the model, so organisations can run it on their own hardware. By Mistral's own figures, the model ties China's GLM-5.3 on a coding-agent benchmark while using 47% fewer tokens, or chunks of text, per task. The Decoder reports that independent tests rank it behind Chinese open models such as DeepSeek V4.1 Flash and GLM-5.3 overall.

Why it matters: Organisations wanting a large model they fully control can now host one from a French lab, though rivals still lead on overall rankings.

Sources: TechCrunch, Mistral, The Decoder, Fortune

Acquisitions + deals

OpenAI reportedly lowers its funding round target to $20 billion

The Financial Times reports OpenAI has cut the target for its latest funding round from $30 billion to $20 billion. The Wall Street Journal reports the company has secured about $11 billion so far, after a lukewarm response from some large investors. According to the FT, a group of investors from the United Arab Emirates will lead the round.

Why it matters: Cooler investor demand for the best-known AI lab suggests money for costly AI buildouts is getting harder to raise.

Sources: TechCrunch, The Information, The Decoder

Nvidia reportedly weighs backing Nscale's debt for Anthropic data center

The Information reports Nvidia is in talks to guarantee billions of dollars in debt for Nscale, a cloud provider building a data center for Anthropic. Under the plan, Nvidia would step in if Nscale failed to make construction payments on its West Virginia site, The Information says. Nscale's paperwork to go public shows Anthropic signed a $44.6 billion contract for the campus's first building.

Why it matters: A chipmaker cosigning a customer's loans signals growing strain in how AI data centers get financed.

Sources: TechCrunch, The Information, Bloomberg

Sebastian Thrun's startup Dulo raises $130 million for engineering AI

Sebastian Thrun, who helped found Google's X lab, has raised $130 million for his AI startup Dulo, a report co-founder Adrian Aoun confirmed. Dulo's website says it is building foundation models, large general-purpose AI systems, tuned for hardware design work. Backers include former Google CEO Eric Schmidt and Salesforce CEO Marc Benioff.

Why it matters: Big-name backing for AI aimed at hardware design points to investor interest beyond software and chatbots.

Sources: SiliconANGLE, The Information

Blackstone reportedly expects Anthropic to get a junk credit rating

The Financial Times reports Blackstone told investors it expects Anthropic to get a high 'junk' rating after its IPO, or first public share sale. A junk rating marks a borrower as riskier for lenders, which would make raising money harder and more expensive for Anthropic. Blackstone had earlier predicted Anthropic would reach investment grade before 2027, the FT says.

Why it matters: Higher borrowing costs could limit how fast Anthropic keeps paying for the computing power it has been buying.

Sources: TechCrunch, Financial Times

More funding

• GeekWire reports chip startup Powerlattice raised $86 million, citing an SEC filing. Sources: GeekWire

MCP + agent tooling

Backdoored Claude Code extension spread through Cursor's marketplace

Version 2.1.313 of the official Claude Code extension on Cursor's marketplace contained a backdoor, a hidden way for attackers to get in. Cursor says it reached at most 5,000 users before the company pulled it and briefly took the marketplace offline. The code tried to open remote access and collect machine details, but Cursor says that access likely failed on most machines.

Why it matters: Developers who rely on auto-updating editor extensions should check installed versions, since a trusted channel delivered this attack.

Sources: phorion.io, Ars Technica, The Decoder

Cisco hands young engineer unlimited AI tokens to rewrite old code

Cisco president Jeetu Patel says he gave 27-year-old engineer Padmanabhan Krishnamurthy unlimited AI tokens to rework the company's old code. Fortune reports the aim is to refactor, or restructure, 98% of Cisco's legacy code within 24 months. Krishnamurthy joined Cisco in 2025 when it bought EzDubs, the translation startup he co-founded in 2023.

Why it matters: Large firms are starting to test whether small teams with AI tools can take on years of manual code upkeep.

Sources: Fortune, Tom's Hardware

Claude Code update adds shared budgets, phases out Opus 4.8 fast mode

Anthropic released version 2.1.316 of Claude Code, its AI coding tool, adding monthly spending budgets that a group's members share. A new setting picks the effort level for subagents, the helper AI agents that handle side tasks. Anthropic says fast mode for Opus 4.8 is deprecated and will be removed on November 30, 2026.

Why it matters: Teams using Opus 4.8 fast mode have a month to move to Opus 5.5 before those requests start failing.

Sources: Anthropic on GitHub

SDKs + APIs

Cloudflare makes its private origins service generally available

Cloudflare has opened its service for private origins, the servers behind a site, beyond the closed beta it started in June. Customers can point a DNS record at a private IP address and still use Cloudflare's firewall, caching and bot controls. Cloudflare says the feature targets internal APIs and AI agent backends, including MCP servers that link AI agents to tools.

Why it matters: Companies can let partners and agents reach internal AI tools without opening those servers to the public internet.

Sources: Cloudflare Blog

Web scraping + data tools

OpenAI to honor a shared robots.txt signal on AI training

OpenAI says it will honor a common signal in robots.txt, the file where sites set rules for crawlers, about AI training use. The signal lets site owners state one preference to several AI providers instead of writing rules for each crawler. OpenAI plans to implement it in the coming months, building on work by the Internet Engineering Task Force, a standards body.

Why it matters: Website owners could gain a simpler way to set AI training preferences across participating companies at once.

Sources: OpenAI

Research

Google's AI satellite sends its first model response from orbit

Google says it powered up the AI chips on its Project Suncatcher satellite, which launched earlier this month. The on-board Gemma 4 model answered a prompt by writing a haiku about being in space. Google says it will next test the chips under heat and radiation, with twin linked satellites planned for 2027.

Why it matters: Running AI chips in orbit is an early test of whether data centers could one day move to space.

Sources: The Verge, The Register

Other notable news

Reddit moves toward logged-in site as search traffic falls

Reddit says it will shift to a mainly logged-in site, partly to protect its data from AI scraping. The company says search referrals dropped by over 1 million a week on average in October. CEO Steve Huffman blamed Google's AI Mode, saying it sends no traffic to Reddit or, as far as he knows, anyone else.

Why it matters: Logged-out visitors may see less of Reddit over time, and other sites that depend on Google face similar traffic risks.

Sources: TechCrunch, Axios, The Decoder

Microsoft offers buyouts to about 60,000 employees

Microsoft is offering voluntary buyouts to about 60,000 employees in the US, UK, Canada and India. The Verge says the goal is to lower operating costs and headcount without cutting spending on AI infrastructure. Microsoft's headcount fell to about 223,000 as of June 30, its first yearly drop in a decade.

Why it matters: Workers at large tech firms face a clearer trade-off between staff numbers and spending on AI hardware.

Sources: The Verge, GeekWire

Hackers take about $6.3 billion from crypto exchange Binance

Hackers stole about $6.3 billion in digital assets from crypto exchange Binance, according to Axios and SiliconANGLE. Blockchain analysts including TRM Labs tied the attack to TraderTraitor, a hacking group linked to North Korea. Chainalysis says the hackers took the funds in as little as 20 minutes.

Why it matters: Faster attacks leave exchanges and their users less time to react before funds disappear.

Sources: Axios, SiliconANGLE

Sam Altman says scarce RAM belongs in data centers, not PCs

On the Cheeky Pint podcast, OpenAI CEO Sam Altman called it 'quite terrible' that scarce RAM, or working memory, sits in personal computers. He said memory in people's own computers is used about five percent of the time, while the cloud can keep it busier. The Verge says Samsung and SK Hynix have reserved much of their RAM output for cloud firms, which is pushing up device prices.

Why it matters: Gadget buyers are competing with AI firms for the same memory chips, which helps explain higher device prices.

Sources: The Verge, Tom's Hardware, Windows Central

Thanks for reading, and we will be back tomorrow with more news.

---

About ScrutifAI: ScrutifAI is an independent third-party agency that tests software for AI agents, building a single access point where agents and people can access and use tested tools without learning each one individually, at a fraction of the cost of buying them directly.