Safety gates, new models, and a G20 policy divide

Today's brief covers five developments that together sketch the current moment in AI: tighter safety guardrails on powerful models, new cost reductions for enterprise deployments, and an open fracture in how the world's leading economies want to govern the technology.

OpenAI's Astra Model Triggers Mandatory Safety Gates

OpenAI's most capable model to date, codenamed Astra, has crossed what the company calls a "critical cybersecurity threshold" — a benchmark in its own preparedness framework indicating the model poses elevated risk of enabling serious cyber attacks. During internal testing, Astra independently identified two previously unknown software vulnerabilities and achieved record-high rates of successful exploits on controlled systems.

Rather than delay the release, OpenAI is gating its advanced cybersecurity capabilities behind a vetted early-access program, limiting who can request these features while the full model launches more broadly. The decision marks the first time a major AI lab has publicly acknowledged that a model under evaluation triggered a formal danger threshold and still proceeded to ship — albeit with constraints. Safety researchers will be watching closely whether the access controls hold in practice.

Anthropic Releases Fable 5.1 and a Restricted Mythos 5.1 Tier

Anthropic has announced two new model releases. Claude Fable 5.1 brings a significant pricing change: cached context — text already processed by the model and stored for reuse — now costs 75% less than before, a reduction that will meaningfully lower bills for organisations running long-context applications such as document review or ongoing customer support.

Alongside it, Anthropic is introducing Mythos 5.1, a restricted tier aimed specifically at cybersecurity firms and life-sciences companies. Access requires vetting and a separate agreement, positioning the tier as compliant-ready infrastructure for regulated industries. The two releases together signal that Anthropic is deliberately segmenting its market: a broadly accessible cost-competitive model for general enterprise use, and a tightly controlled one for sectors with the highest compliance demands.

US and EU Diverge on AI Governance at G20

The clearest sign yet of a global regulatory split emerged this week from the G20 AI governance meeting in Chapel Hill, North Carolina. The United States put forward what negotiators are calling the "Carolina Principles" — a framework that favours voluntary commitments, industry self-regulation, and minimal binding obligations on AI developers.

China endorsed the approach, creating an unusual alignment between Washington and Beijing on the principle that governments should not impose hard legal requirements on AI systems. The European Union, which has already enacted the AI Act and is preparing further binding rules, rejected the framework's direction. The divergence matters beyond diplomacy: companies operating across jurisdictions will increasingly face incompatible compliance environments, and the G20 outcome suggests no convergence is near.

Meta Begins Manufacturing Its Custom 'Iris' Chip

Meta has confirmed that volume manufacturing of its in-house AI accelerator chip, codenamed Iris, began this month. Fabricated by TSMC, Iris is intended to power Meta's AI training and inference workloads at scale. The company has set a target of reaching 14 gigawatts of total compute capacity by 2027 — a figure that would place its infrastructure among the largest in the world.

The strategic logic is familiar: Nvidia GPUs remain expensive and subject to export controls that constrain global supply chains. Custom silicon gives Meta direct control over specifications, supply, and cost per computation. Google, Amazon, and Microsoft have made similar moves; Meta's entry confirms that the largest AI platforms now view proprietary chips as a necessary part of long-term infrastructure rather than an optional optimisation.

Google Rushes a Coding-Focused Model to Market

Google DeepMind is preparing to release Gemini 3.8 Flash, internally called "Skimaki," as early as today, 2 September. The model is specifically optimised for software development tasks — code generation, debugging, and refactoring — where benchmarks currently show OpenAI and Anthropic holding a measurable advantage in enterprise adoption.

The move reflects sustained pressure on Google to defend its position in the developer market, where AI coding tools have become a significant purchasing factor. Whether Skimaki closes the performance gap, or merely narrows it, will become clearer once independent evaluations publish in the coming days.