$16.3 billion in half-year revenue, an $80.7 billion net loss, and no IPO price yet. Anthropic’s first public prospectus has handed the AI market a rare look inside a frontier lab—and the numbers need reading, not merely repeating. Most of that extraordinary net loss came from noncash fair-value changes tied to convertible financing, while the underlying operating loss narrowed. The answer at the top of this AI news today brief is therefore nuanced: Anthropic is growing at remarkable speed, but its proposed Nasdaq debut is still an unfinished transaction with unusual economics and governance. Elsewhere, OpenAI launched a cheaper computer-use model, Nvidia attached a $3 billion investment to a deeper Meta partnership, Siri AI added five languages, IBM put agent identity into general availability, and one chip startup bought another to accelerate an unproven efficiency target. The common theme is commercialization—who gets access, who controls the agent, and who pays for the compute.

AI news today: Anthropic opens its books for Nasdaq

Anthropic filed its Form S-1 with the SEC on October 26 and applied to list Class A shares on Nasdaq under the ticker ANT. The official filing leaves the number of shares and price range blank, and the registration statement is not yet effective. In other words, this is a proposed offering, not a completed listing.

The filing reports $16.286 billion in revenue for the first half of 2026, up from $1.163 billion a year earlier. Operating loss improved to $693 million from $2.135 billion, but net loss widened to $80.748 billion; $72.388 billion came from changes in the fair value of convertible notes, derivatives and SAFEs. TechCrunch’s reading also highlights that seven founders would collectively retain 50.1% of shareholder voting power. Investors now have to value growth, compute needs and concentrated control together. That tension follows naturally from our recent look at Anthropic’s agent oversight bottleneck.

GPT-6.1 Terra competes on task economics

OpenAI released GPT-6.1 Terra for visual reasoning, writing and computer use. Standard API pricing is $1 per million input tokens and $5 per million output tokens; OpenAI says that is 50% below GPT-5.6 Terra and GPT-6.1 Sol. The model is also rolling out in ChatGPT Work and Codex, but not ordinary ChatGPT conversations. Access depends on plan, workspace controls and the setup required for browser or computer use.

The performance claims in OpenAI’s launch post are company benchmarks: at medium reasoning effort, it says Terra scored 15 percentage points above its predecessor on Agents’ Last Exam at roughly one-third the cost. For builders, the useful question is not whether Terra wins every benchmark; it is whether a faster, cheaper model can complete enough screen-based work accurately to improve cost per finished task. Watch independent evaluations and failure rates, especially before granting write access.

Nvidia ties a $3 billion investment to Meta’s Muse stack

Nvidia and Meta announced a deeper partnership around Meta’s Muse agents and models. Meta intends to expand deployments of Nvidia’s Vera Rubin systems, Vera CPUs, Spectrum-X networking and BlueField-4 infrastructure. The companies will optimize training and inference together and say they will consider additional work on OpenShell for safer agent deployment and Sentry for behavior monitoring.

The most concrete number is financial: Nvidia says it will invest $3 billion in Meta’s equity offering. The release treats completion and benefits as forward-looking, so do not count the investment as closed or the technical gains as proven. Strategically, the arrangement tightens the loop between a major chip supplier and one of its largest AI customers. For competitors, that raises a familiar question: can infrastructure partnerships remain neutral when capital and product roadmaps increasingly overlap?

Siri AI adds five languages but keeps important limits

Apple expanded Siri AI to French, Japanese, Korean, Portuguese and Spanish with its 27.2 software releases. The beta runs on Apple Intelligence-compatible devices across iOS, iPadOS, macOS, watchOS and visionOS. Apple’s availability notes matter as much as the language list: Siri AI remains unavailable in the EU on iPhone, iPad and Apple Watch, is unavailable to users under 13, and some server-side features carry daily limits.

Apple says Chinese (traditional), Norwegian, Turkish and Vietnamese support should arrive by year-end, and that expanded access to some usage-limited features will eventually carry a fee. This is a real expansion, but not a universal launch. Product teams should watch how well actions, personal context and safety behavior transfer between languages—not assume that translating the interface reproduces identical capability.

IBM gives enterprise agents their own identity layer

IBM made Agent Identity generally available, positioning AI agents as first-class identities rather than anonymous processes borrowing a user’s credentials. The product is designed to register agents, record ownership and approved tool connections, preserve on-behalf-of delegation, issue short-lived task-scoped access and connect actions to an audit trail. IBM links the control plane to Verify, HashiCorp Vault and watsonx Orchestrate.

IBM’s announcement describes intended controls, not independent proof that every deployment is secure. The direction is still consequential. Once an agent can call APIs or delegate work, “which model?” becomes less important than “whose authority did it use, for what task, and can that authority be revoked?” This operational focus also complements yesterday’s brief on the SEC’s advanced-AI examination priorities.

Unconventional AI buys silicon experience, not a finished miracle

Unconventional AI reached a definitive agreement to acquire Israeli inference-chip company NeuReality. The boards have approved the transaction, which is expected to close in the fourth quarter subject to customary conditions. Terms were not disclosed. Reuters reports that roughly 40 NeuReality employees will join; the buyer says it will keep an engineering presence in Israel.

The reason is speed. Unconventional says NeuReality’s system, silicon and software experience will shorten its path to production. Its headline target—a 1,000-fold improvement in AI compute efficiency—is an ambition, not a demonstrated product result. What comes next is silicon evidence: working chips, measured workloads, power numbers and a manufacturing plan. Hiring a team can compress a roadmap; it cannot repeal physics.

Watch & Learn

Editor’s note: IBM Technology’s verified 14-minute lesson, How Will AI Agents Manage Identity & Build Trust in Complex Systems, explains identity propagation, token exchange and transitive trust using OAuth 2.0, OpenID Connect and APIs. It is a useful primer for architects and security leads who need to understand why one permanent credential is a poor fit for multi-agent workflows.

AI, Translated: computer use

Computer use is an AI model’s ability to interpret what appears on a screen and operate software through actions such as clicking, typing and scrolling. For example, a model might read an invoice PDF, open accounting software and enter the approved fields. Unlike a fixed automation script, it reasons over changing interfaces—but that flexibility can also make it brittle. You should care because every click turns a model answer into a real-world action, making permissions, previews and rollback controls essential.

Try This Today: audit a screen with GPT-6.1 Terra

Goal: turn one cluttered interface into a prioritized visual QA list in about ten minutes. Official OpenAI documentation says GPT-6.1 Terra is available in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans. It is not available in ordinary Chat, and workspace, browser and computer-use controls can affect access.

1. Open ChatGPT Work and choose GPT-6.1 Terra under Advanced.
2. Upload a screenshot of a dashboard, form or landing page with no confidential data.
3. Ask for evidence before recommendations, then verify each issue against the image yourself.

Copy-ready prompt: “Inspect this screen without taking any action. List usability, accessibility and visual-consistency issues in a table with: exact on-screen evidence, likely user impact, confidence, and the smallest safe fix. Separate observations from assumptions. Rank the top five fixes by impact divided by effort, and finish with three checks a human reviewer should perform before approving changes.”

One thing to remember

The AI business is shifting from impressive demos to capital, distribution and accountable execution. The winning number will be useful work completed under control—not tokens generated or valuations announced.