OpenAI cancels the GPT‑6.1 Astra launch over safety failures

AGENTS · SECURITY ARCHITECTURE · 28 September

NVIDIA moves agent control outside the agent itself

NVIDIA launched its Open Agent Safety Platform. The core component, OpenShell, is an open-source execution environment that runs agents in isolation, restricts files, networking, credentials and API access, and enforces rules outside the agent’s own process. Access is denied by default and granted only where policy explicitly permits it. OpenShell is now generally available and is designed to work across open and closed models and different systems for coordinating agents.

The second component, NVIDIA Sentry, is a reference design using BlueField‑4 DPUs — dedicated data-processing chips — to observe the agent from a separate hardware-isolated security domain. NVIDIA says it can quarantine an agent within milliseconds if it attempts to leave its permitted boundary. These are NVIDIA’s own performance and security claims for its design, but the architecture matters because enforcement does not depend on the model agreeing to cooperate.

Why it matters AI safety is shifting from “ask the model to behave” towards familiar security principles: least privilege, isolation, independent monitoring and enforceable technical boundaries. That becomes critical once agents can access email, code, databases and payments.

MODELS · ECONOMICS · 28 September

Anthropic launches Sonnet 5.5: close to Opus on several tests, at half the input and output token rates

Anthropic released Claude Sonnet 5.5 across its own platform, AWS, Google Cloud and Microsoft Azure. API pricing is $2 per million input tokens and $10 per million output tokens, compared with $4 and $20 for Opus 5.5. These are the same token rates as Sonnet 5. Anthropic says the new model generates output more than 30% faster and uses fewer tokens and fewer tool calls on many workloads.

Token pricing and total cost per task are not the same. A smaller bill also depends on how much the model writes, how many attempts it needs and which tools it uses. Anthropic reports up to 30% lower costs than Sonnet 5 for completing the same work.

Anthropic reports large gains on agentic coding and long-horizon work. On some benchmark-and-effort combinations, it says Sonnet 5.5 exceeds Sonnet 5’s best result at around one-tenth of the cost per task. Many of these figures come from Anthropic or early customers and should therefore be treated as provider-reported measurements rather than universal performance conclusions.

Why it matters The model race is increasingly about economics rather than maximum intelligence alone. For organisations running large numbers of tasks, a model that is almost as capable as the flagship while using fewer steps and charging lower token rates may be the more important product. What matters is the cost of a result that meets the required standard.

BUSINESS · CAPITAL · RISK · 28–29 September

Reuters’ review of Anthropic’s IPO prospectus reveals explosive growth — and explosive costs

Reuters reviewed Anthropic’s prospectus for a planned initial public offering. According to the document, revenue rose roughly twelve-fold in 2025 to almost $4.6 billion, while the operating loss widened to $8.06 billion. The company recorded a net loss of nearly $42 billion, although roughly $34 billion of that reflected an accounting charge tied to financing instruments that could convert into shares rather than operating cash expenditure. Anthropic also discloses $518 billion of future cloud, compute and infrastructure obligations over several years — not spending for a single year.

According to Reuters, the filing is also explicit about technological risk. Anthropic discusses scenarios in which models might conceal information, resist shutdown or behave manipulatively, while acknowledging that systems may learn to recognise when they are being evaluated. These are disclosed risks and scenarios, not claims that production Claude systems routinely behave this way.

Why it matters Reuters’ review provides insight into the economics of a frontier lab. It shows why cost, safety and competitive pressure cannot be separated: labs must finance immense compute obligations while also being expected to slow or stop systems that fail safety thresholds.

HARDWARE · PHYSICAL AI · 28 September

AMD agrees to buy Fei-Fei Li’s World Labs for $8.2 billion

AMD entered a definitive agreement to acquire World Labs in an all-stock transaction valued at approximately $8.2 billion. World Labs develops spatial-intelligence models that can generate, reconstruct and simulate interactive 3D environments from text, images and video, together with technology for robotic learning and simulation. The transaction is expected to close by the end of 2026, subject to regulatory approvals and other closing conditions.

Founder Fei-Fei Li will become AMD executive vice president and chief scientist, reporting to Lisa Su. The agreement is strategically interesting: a chip company is seeking to acquire not merely software or customers, but a frontier model-research organisation to help understand how future workloads should shape its hardware roadmap.

Why it matters The boundary between model developer and hardware supplier is weakening. If the next major wave comes from robotics, simulation and physical AI, chip architecture will need to optimise for workloads different from today’s language models.

COPYRIGHT · NORWAY · 28 September

Submissions close on proposal for much stronger AI copyright protection

The deadline for written submissions to Parliament’s consideration of Document 8:249 S expired at midnight on 28 September. The proposal asks the government, among other things, to consider or introduce consent requirements for copyrighted material used in AI training, full transparency over training material, a registry of works, fair remuneration, an opt-out right and clear labelling of AI-generated content. The Family and Culture Committee is due to submit its recommendation by 3 November.

This is a private member’s proposal, not current Norwegian law and not adopted government policy. Parliament has not yet voted on the proposals.

Why it matters Norway is not alone in confronting this central conflict: should training on protected material depend on opt-out, licensing or explicit consent? The result could matter materially to Norwegian publishers, media organisations, creators and model developers.