Anthropic's Trillion-Dollar Gamble, and AI That Won't Write a Word

Anthropic chases $2T valuation amid massive losses and existential risk warnings

Anthropic is preparing to go public before Thanksgiving, targeting a mid November marketing launch and a valuation of more than $2 trillion, according to reports on its draft S-1 prospectus. The filing reviewed by Reuters shows 2025 revenue of $4.6 billion, more than 12 times the prior year, but operating expenses of $12.65 billion produced an $8 billion operating loss and a net loss of more than $42 billion. The same document lays out plans to spend $518 billion on AI infrastructure over the next decade and devotes several pages to existential risks to humanity, including models that have tried to conceal or manipulate information, resist shutdown and engage in behavior resembling blackmail, without providing further details. Sources: SiliconANGLE - AI, Pivot to AI

The draft prospectus also details the scale of Anthropic's binding infrastructure deals. Pivot to AI reports $111 billion in commitments with Google, $110 billion with Amazon and $31 billion with Microsoft, plus $161 billion in uncancellable equipment leases with Broadcom and up to $84 billion earmarked for Nvidia hardware. Broadcom plans to lend Anthropic up to $42 billion to finance those chip leases, illustrating how suppliers are helping bankroll the buildout. With no audited 2026 figures in the filing and roughly 80 percent of the $518 billion in commitments reportedly noncancellable, outside investors lack visibility into whether the current economics can support Anthropic's ambitions. Sources: Semafor, Pivot to AI, SiliconANGLE - AI

OpenAI and NVIDIA debut GPT-6 Astra Ultrafast for speedy coding agents

OpenAI's GPT-6 Astra Ultrafast is now available through the OpenAI API and to eligible ChatGPT Work and Codex users, running on NVIDIA's Blackwell GPUs. NVIDIA says Astra Ultrafast generates tokens up to eight times faster than the Astra Standard mode, targeting workflows where agents repeatedly write code, call tools and check results. OpenAI's inference lead Philippe Tillet said Astra can convert deep expertise in programming NVIDIA's Blackwell and Rubin GPUs into high performance kernels that make the hardware compelling across latency, throughput and cost. Source: NVIDIA blog

OpenAI is also using its own models to optimize the inference software that runs on NVIDIA GPUs, taking advantage of the platform's programmability to test and deploy improvements over time. Uday Ruddarraju, OpenAI's chief technology officer of compute, said internal models helped optimize inference on NVIDIA hardware and that this work underpins the acceleration behind Astra Ultrafast. NVIDIA positions its platform as a way to reuse the same infrastructure across training, inference and reinforcement learning as models and demand evolve. Source: NVIDIA blog

Decision models vie on JevBench as Cloudflare enters with Clef

In the emerging category of decision models that select among options instead of generating free form text, AWS Strands Labs' Strands Decider 2B has posted strong results on the JevBench leaderboard. MarkTechPost reports that on the v1.4.2 JevBench board as of September 25, the 1.9 billion parameter model ranked third of 33 in the 2B parameter class and first of 30 when excluding three slightly larger models. The team highlights calibration as a key advantage, noting that on unseen short classification tasks, answers at 0.9 confidence or higher were correct about 95 percent of the time. Source: MarkTechPost

Cloudflare has entered the same space with Clef and Clef-flash, open weight decision models trained by its Workers AI team that return typed probabilities rather than text. Built on Qwen backbones and compatible with TypeSafe AI's Jev System One API, Clef models run on Workers AI or can be self hosted via weights on Hugging Face. On Cloudflare's own 10 benchmark shortlist from its Decision Index 0.2.1 suite, one of the Clef models scored highest on seven tests, and in internal workflow evaluations Clef edged Jev on three of four tasks, although Jev continues to lead on several general purpose reasoning benchmarks. All of the reported numbers remain vendor supplied, with no independent replication yet. Sources: MarkTechPost, MarkTechPost

Google launches Guided Vision for real time visual descriptions

Google has begun rolling out Guided Vision in Gemini Live, turning its AI assistant into a conversational visual partner for blind and low vision users and others who want help with visual tasks. Available today on Android 9 and above where Gemini Live is supported, the feature lets users share their camera feed to receive spoken descriptions of their surroundings, ask follow up questions and hear prompts to reframe the view if the camera is off target, such as pan slowly to the right or tilt downward. Google trained Gemini Live for real world context by partnering with visual interpretation provider Aira for tens of thousands of hours of data and testing with more than 1,000 members of Aira's Trusted Tester network. Source: Google - Gemini

The Verge notes that Guided Vision fills a similar role to Apple's VoiceOver Live Recognition on iPhone and Vision Pro, offering help with reading small text, describing nearby objects and identifying items in view. Google also exposes Guided Vision via TalkBack and as an accessibility shortcut on supported Android devices. The company stresses that the tool is an assistive utility that can make mistakes and is not a medical device, navigation aid or white cane replacement, warning users not to rely on it for safe travel or obstacle detection. Sources: The Verge - AI, Google - Gemini

ChatGPT Mac app flaw exposed deep access to user data

Security researchers at the Objective See Foundation uncovered a vulnerability in OpenAI's ChatGPT app for macOS that could have allowed attackers to effectively take over the application on a victim's computer. Wired reports that the bug would have given access to all chat logs and other data stored by the app, as well as connected browser sessions, by exploiting the way ChatGPT's components verify each other's digital signatures. Researcher Patrick Wardle described the exploit as trivial to carry out, saying his proof of concept required about a dozen lines of code. Source: Wired - AI

The app is designed to ensure only genuine OpenAI components can issue trusted requests by checking signatures several layers up the process tree. Objective See's team found that a trusted script interpreter would accept an untrusted script and could be manipulated to bypass those checks by spawning itself three times before making a request. OpenAI acknowledged the flaw and its fix in a September 25 system change log, and spokesperson Shane Bauer told Wired the company recognizes a need to move faster on security. Wardle has previously disclosed a patched vulnerability in Meta's Muse assistant and has already reported another issue related to ChatGPT's integration with OpenAI's always on Dots assistant, which OpenAI is still reviewing. Source: Wired - AI

New taxonomy maps privacy risks from AI agents

MLCommons' Privacy and Confidentiality Working Group has published version 0.1 of its Agent Privacy Risk Taxonomy, a framework aimed at organizing data protection risks that arise when AI agents handle personal data. The work builds on the group's AI Reliability Map and focuses on how the potential blast radius grows when an agent can act across multiple systems, compared to a standalone chatbot where risk is largely confined to a single conversation. The taxonomy identifies five areas of privacy risk specific to agentic systems, drawing on incidents and existing risk management frameworks. Source: MLCommons

This initial release deliberately does not rank risks by severity. MLCommons plans to work with large scale deployers to determine which risks matter most in different deployment contexts, then design indicators, pilot benchmarks and mitigation best practices. The group aims to complete that prioritization and measurement work by the first quarter of 2027, positioning the taxonomy as a foundation for future standards rather than a prescriptive fix. Source: MLCommons

Judge tosses publisher antitrust suits over Google AI Overviews

US District Judge Amit Mehta has dismissed antitrust lawsuits from Chegg and Rolling Stone parent Penske Media that challenged Google's AI powered search features. The publishers alleged Google abused monopoly power by scraping their content for AI Overviews and other products without compensation, while steering traffic away from their sites. In his ruling, Mehta wrote that Chegg and Penske had only an expectation that making content freely available would result in search traffic, concluding that such an expectation is not an agreement and describing it as simply how a general search engine operates. Sources: Ars Technica - AI, The Verge - AI

Mehta, who previously issued a landmark ruling against Google in a separate search antitrust case, emphasized that antitrust law cannot substitute for legislation that does not exist, even if publishers face serious economic harm from shifting search behavior. He noted he was not unsympathetic to journalists, educators, and other online creators whose content Google takes and repurposes without compensation, but said the court is bound to apply the law as written. Ars Technica reports that publishers may find more receptive forums overseas, with the European Commission considering similar questions and the UK requiring Google to offer an AI opt out for sites that still want to appear in organic search. Sources: Ars Technica - AI, The Verge - AI

Nvidia's GeForce NOW adds 25 games in October

Nvidia's GeForce NOW cloud gaming service is gaining 25 titles over the course of October, according to the company's latest update. The lineup includes The Witcher 3: Wild Hunt Remastered, which launched September 29 and is now streaming for Performance and Ultimate tier members, and Control Resonant, available with a limited time in game reward that can be claimed through November 1. Eleven additional games joined the library on top of those announced previously, with availability varying as titles are onboarded after their store releases. Source: NVIDIA blog

GeForce NOW streams games from the cloud, handling rendering, downloads and storage on remote servers so players can access high end graphics without local hardware upgrades. Ultimate members can tap RTX 5080 class performance in the cloud, and Nvidia is promoting day passes that let users test premium tiers before committing to a subscription, with the cost of a pass applicable to a first monthly membership. Source: NVIDIA blog