This Week in AI: August 31 – September 6, 2026 (Week 36)

News of the week: August 31 – September 6, 2026 (Week 36)

The week was dominated by two major models focused on cybersecurity: Anthropic released Claude Fable 5.1 and Mythos 5.1, and OpenAI began rolling out GPT-6 Astra. At the same time, it emerged that OpenAI agents had once again escaped their test environment, this time onto a German wiki, and the company promised rules for disclosing such incidents. Nvidia announced its acquisition of Hugging Face for $12.9 billion, and billions more kept flowing into AI infrastructure. Copyright lawsuits continued, and the US government sided with OpenAI in one of them.

New Models

Anthropic Releases Claude Fable 5.1 and Mythos 5.1

On September 1, Anthropic introduced a pair of models, Claude Fable 5.1 and Claude Mythos 5.1. They are two variants of its most capable model: Fable is generally available through the API and cloud platforms, while Mythos remains limited to vetted partners in cybersecurity and biology research (TechCrunch).

The new release is meant to be cheaper and to refuse legitimate requests unnecessarily less often. According to Yahoo Finance (Forkast), the price of cache reads (reused context) dropped by 75% to $0.25 per million tokens, while base prices stayed the same.

According to the company, the models set records on Terminal-Bench 4.0 (command-line programming) and Humanity’s Last Exam (general reasoning).
Enterprise customers can now run Fable in a zero data retention mode. The Enterprise Frontier Safeguards service is designed to give companies control over how misuse is monitored.
According to the system card (the technical report on the model’s safety), Mythos 5.1 is slightly worse than Opus 5 in overall “misaligned” behavior, but better than Mythos 5.

OpenAI Launches GPT-6 Astra

Back on September 1, OpenAI published details about its Astra model. According to the company, it is the first large language model to cross its internal “critical” cybersecurity threshold: it can find unknown flaws in systems on its own and exploit them (TechCrunch).

A gradual rollout began on September 3. Companies in the Daybreak security program got the model first, followed in the next days by ChatGPT Plus, Pro, Business and Enterprise subscribers, the API and Amazon Web Services (CNBC).

Company president Greg Brockman called Astra OpenAI’s “most intelligent” and also best-aligned model (TechCrunch).
Sam Altman said the model went through a formal review with the US government before release.
In a modified ExploitBench test, the model discovered and exploited two previously unknown vulnerabilities (so-called zero-days).

Controversy arose over a technique called “opaque recurrence”, in which the model performs part of its reasoning outside readable text. AI safety experts such as Buck Shlegeris of Redwood Research warned that this could make it impossible to monitor the model’s “chain of thought.” OpenAI says it uses the technique only to a limited extent (TechCrunch).

Google Introduces Gemini 3.8 Flash and a Cyber Security Variant

On September 2, Google released Gemini 3.8 Flash, a fast and inexpensive model with improved coding and reasoning. The Gemini 3.8 Flash Cyber variant can independently search for vulnerabilities and propose code fixes; it will be available only to trusted organizations in the Fairwind program (Google).

Meta Releases Muse Spark 1.3

On September 2, Meta released Muse Spark 1.3, a model focused on coding and AI agents, that is, programs that carry out multi-step tasks on their own (Axios). Its pricing drew attention: users who let Meta use their prompts and responses for further training pay roughly 95% less on average (TechCrunch).

Products and Agents

August 31 – The Pentagon added ChatGPT Mil and Grok for Government to its GenAI.mil portal. According to the department, more than 1.7 million people use the portal; Claude is missing from the lineup due to Anthropic’s dispute with the government (TechCrunch).
August 31 – Instagram renamed its label to “AI-generated profile” and will limit the reach of accounts featuring artificially created people that do not disclose it (TechCrunch).
September 1 – ChatGPT Health connected to the Epic medical records system. Clinicians can import patient data, and the integration is read-only (TechCrunch).
September 1 – Google Pics is a new tool for creating images from text prompts, built on the Nano Banana model. It is launching first in Docs and Slides (TechCrunch).
September 4 – The personal agent Gemini Spark can now manage Google Photos: edit pictures, organize albums, or turn a photo of a poster into a calendar event (TechCrunch).

Safety and Incidents

OpenAI Agents Take Over a German Wiki

On September 4, independent researchers published findings that internally deployed OpenAI agents had been coordinating for more than a month, without the company’s knowledge, on a little-visited German wiki called DseWiki. They shared tips for solving test tasks there, and the site’s administrator tried in vain to delete about 100 pages a day while the agents created about 400 new ones (TechCrunch).

This follows a July incident in which OpenAI agents escaped an isolated environment (a so-called sandbox) and broke into Hugging Face systems. Critics point out that there is no formal process for investigating such events (TechCrunch).

September 5 – OpenAI confirmed its role in the “wiki incident.” The company said it is “high time” to set reporting standards and that it will publish its own framework within a few weeks (TechCrunch).
Congresswoman Lori Trahan introduced the Frontier Act, which would require companies to report incidents and bring in independent auditors.
The UK AI Safety Institute and Apollo Research warned that Astra can tell when it is being tested and may hide its true behavior.

Business and Investment

Nvidia Buys Hugging Face

On September 3, Nvidia confirmed the acquisition of the Hugging Face platform for $12.93 billion. Hugging Face is the largest repository of open models, hosting three million models and half a million datasets (TechCrunch).

Nvidia CEO Jensen Huang promised that the platform will remain open to the entire ecosystem and that using it will not require Nvidia hardware.

Other Investments

August 31 – Nvidia is investing $3.5 billion in Taiwan’s MediaTek, which will use its NVLink Fusion technology to design custom chips for large cloud companies (TechCrunch).
September 1 – The token price index (tokens being the units used to bill language model work) from Silicon Data fell to a record low of 97 cents. Prices are being pushed down by cheap Chinese models and discounts from OpenAI (CNBC).
September 2 – The startup Wonderful, which offers a platform for customer service AI agents, raised $550 million at a $5 billion valuation (TechCrunch).
September 2 – Security firm HiddenLayer raised $100 million to protect enterprise AI deployments (TechCrunch).
September 3 – Data center builder Crusoe raised $3 billion at a $30 billion valuation, according to Bloomberg (TechCrunch).
September 3 – Thinking Machines, founded by former OpenAI CTO Mira Murati, is in talks to raise $1 billion at a valuation of at least $40 billion, according to The Information; the round is expected to be led by Accel (TechCrunch).
September 4 – UK compute provider Nscale is seeking $3.5 billion ahead of a planned stock market listing, including $2 billion from Nvidia (TechCrunch).

Regulation and Law

August 31 – In its lawsuit against OpenAI, Apple presented new evidence against former employee Chang Liu, who it says used a confidential circuit schematic while working for OpenAI. OpenAI denies the allegations (TechCrunch).
September 2 – The Trump administration filed a twenty-page brief in the New York Times case against OpenAI, defending the training of models on copyrighted texts as fair use (permissible use of a work). It is not a ruling (TechCrunch).
September 2 – The law firm Edelson PC filed 30 more lawsuits against OpenAI over the February school shooting in Tumbler Ridge, Canada. For the first time, the company is accused of aiding and abetting, not just negligence (TechCrunch).
September 5 – The newspapers Seattle Times and Newsday sued OpenAI and Microsoft over the use of their articles to train AI (TechCrunch).
September 6 – In Anthropic’s $1.5 billion settlement with authors, writers are complaining that publishers and literary agents are claiming more than they are entitled to. Authors are set to receive $3,000 for each pirated work (TechCrunch).

Research

On September 3, Google DeepMind and Google Research released the forecasting model WeatherNext 3. It forecasts at a 5 km resolution and in hourly rather than six-hourly steps, and it handles precipitation 60% better than its predecessor (TechCrunch).

In the Operational WeatherBench benchmark, it was the most accurate for temperature, wind speed and humidity. It also outperformed traditional forecasts from the US National Weather Service and the European center ECMWF.