This Week in AI: August 3-9, 2026 (Week 32)

News of the week: August 3-9, 2026 (Week 32)

The week was dominated by a leadership shake-up at Google, where Demis Hassabis stepped back from running DeepMind and Jeff Dean is leaving after 27 years to found his own company. OpenAI made the new GPT-5.6 Luna the default model for free ChatGPT users and promised unlimited text conversations. China delivered the giant Qwen3.8-Max model, along with an analysis finding that DeepSeek V4-Flash is by far the cheapest well-known model. Money keeps pouring into compute and energy, whether it is Anthropic’s 10 billion dollar contract or a billion for nuclear startup Valar Atomics. Meanwhile in Washington, the government met with the major labs over testing the cyber capabilities of AI models.

New Models

Alibaba releases Qwen3.8-Max with 2.4 trillion parameters

Alibaba on August 3 launched its largest model yet, Qwen3.8-Max, through its QwenCloud interface, the Model Studio platform and the Qwen Studio chat app. According to the company, the model has 2.4 trillion parameters, but only about 95 billion of them are active at each step (Quartz).

It uses a mixture of experts architecture: the model is made up of many smaller sub-networks and activates only a few of them for any given task, which saves compute. Alibaba also promised to release the open weights (files anyone can download and run on their own hardware) a week later, together with a smaller 27 billion parameter version.

According to the price list, API pricing is 2 dollars per million input tokens and 6 dollars per million output tokens (a token is roughly a piece of a word).
According to Nikkei Asia, the model debuted in fifth place on the text leaderboard and second on the vision leaderboard of the Arena evaluation platform.

DeepSeek V4-Flash is the cheapest well-known model, analysts say

Research firm Artificial Analysis published a cost comparison on August 3 showing that a single run of its benchmark suite with DeepSeek V4-Flash costs about 3 cents on average. For comparison, Moonshot’s Kimi K3 came in at 86 cents, OpenAI’s GPT-5.6 Sol at 1.86 dollars and Claude Fable 5 at 3.15 dollars (Reuters).

Cheap, however, is not the same as top performance. On the overall Intelligence Index, V4-Flash scored 50 out of 100, the same as Google’s Gemini 3.6 Flash, while the strongest models from OpenAI and Anthropic sit at least nine points higher (The Next Web).

The official DeepSeek price list shows 0.14 dollars per million input tokens and 0.28 dollars per million output tokens.
The company released the model itself back on Friday, July 31, but the cost analysis only came out this week.

Nvidia opens up a voice model that listens and speaks at the same time

Nvidia on August 3 published NemotronLabs VoiceChat-11B, an 11 billion parameter model, on Hugging Face. It is a so-called full-duplex model: much like a person on the phone, it can keep listening while it is speaking and can be interrupted (Hugging Face).

The model handles speech recognition and response in a single unit, instead of the usual chain of three separate systems (transcription, language model, speech synthesis).
According to Nvidia, it can call external tools during a conversation and responds with a latency of around 450 milliseconds; however, the license is for research purposes only.

Products and Agents

Free ChatGPT gets GPT-5.6 Luna and unlimited text chats

OpenAI announced on August 6 that GPT-5.6 Luna, the smallest of the three GPT-5.6 models, is becoming the default model for free users and subscribers to the low-cost ChatGPT Go plan. Starting the week of August 10, these users are also set to get unlimited text conversations (MacRumors).

The unlimited part applies only to text; image generation, file uploads and other tools keep their own limits. OpenAI claims that Luna gives about 62% fewer responses with factual errors in finance, medicine and law than the previous GPT-5.5 Instant, though this is the company’s own internal measurement (The Next Web).

Free users get a Think button that gives the model more time to reason through a harder question.
Plus and Pro subscribers received an updated GPT-5.6 Sol with a slider for setting how hard the model should think.

Anthropic starts scanning skills and plugins for malicious code

Anthropic on August 6 launched a beta feature for its Claude Enterprise business plan that automatically checks uploaded third-party skills and plugins. These extend what the Claude assistant can do, but they can also serve as an entry point for attackers (Claude Help Center).

The scan runs when a skill is uploaded or edited, before anyone in the company runs it; administrators can see the results in the organization settings.
Tools shared through a connected MCP server (a standard for connecting AI to external services) are not covered by the scan.

Business and Investment

Leadership shake-up at Google DeepMind: Hassabis steps back, Dean leaves

Google announced a major reorganization of its AI leadership on August 5. Demis Hassabis, co-founder and until now head of Google DeepMind, becomes chairman of the division and chief scientist of all of Alphabet; he also continues to lead Isomorphic Labs, the company focused on drug discovery (The Decoder).

Day-to-day management has been taken over by CTO Koray Kavukcuoglu as a senior vice president reporting directly to Google CEO Sundar Pichai. Alphabet shares fell on the news, by 4 to 5% depending on the source.

Leaving the company after 27 years is Jeff Dean, one of the architects of both Google’s search engine and its modern AI. Together with colleagues Sanjay Ghemawat, Oriol Vinyals and Quoc Le, he is founding Discovery Loop, a company that aims to automate research in machine learning, science and engineering (GeekWire).

Discovery Loop will be an independent public benefit corporation, meaning a company legally required to pursue a public benefit goal, not just profit.
Google will invest in it and serve as its cloud provider; the size of the investment was not disclosed.

Anthropic secures 10 billion dollars in compute, Bloomberg reports

New cloud startup Volta announced a six-year, 10 billion dollar contract on August 4 with an unnamed “AI lab.” Citing people familiar with the matter, Bloomberg reported that the customer is Anthropic, maker of the Claude assistant (TechCrunch).

The capacity is to be located in a Bitdeer data center in Tydal, Norway, powered by hydroelectric plants and equipped with Nvidia Vera Rubin chips. Neither Anthropic nor Bitdeer commented on the report, and Volta did not name its client (Channel Insider).

AMD and Palantir results: AI demand keeps growing

Chipmaker AMD reported second-quarter results on August 4. Revenue rose by roughly half to 11.5 billion dollars, and the data center segment more than doubled to 6.7 billion (AMD / SEC).

For the third quarter, AMD expects revenue of around 13 billion dollars, and later this year it plans to start shipping complete Helios AI server racks to compete with Nvidia.
The shares nonetheless fell after the release, with analysts pointing to higher than expected capital spending.
Software company Palantir reported revenue of about 1.94 billion dollars on August 3, up 93% from a year earlier, and raised its full-year outlook to 8.15 to 8.16 billion; US commercial revenue grew 149% (Let’s Data Science).

Energy and debt: a billion for nuclear reactors and a warning about a chip bubble

US startup Valar Atomics, which develops small nuclear reactors, among other things to power AI data centers, raised 1 billion dollars in a funding round led by Sequoia, plus a 200 million dollar credit facility. According to Axios, the company is now valued at 6 billion dollars (Axios, Valar Atomics).

The money is meant to help the company move from demonstrating its reactor to series production; its Ward 250 microreactor has not yet been approved by the US nuclear regulator, the NRC.
Market maker Citadel Securities estimated on August 3 that a 500 billion dollar debt market for financing AI chips will emerge by 2028 (Bloomberg).
Because of a global shortage of memory chips, made worse by demand from AI data centers, manufacturers HP, Asus and Acer have started using memory from Chinese company CXMT, so far mainly in low-cost devices outside the US (Nikkei Asia).

Regulation and Law

White House presents companies with a framework for testing AI cyber capabilities

The White House met on August 4 with representatives of OpenAI, Google, Meta and Anthropic to walk them through a finalized voluntary framework for testing the most powerful models. The tests check whether a model could help with things like attacks on computer systems before it reaches the public (CNBC).

The framework builds on a June executive order, and most of the evaluations are to be carried out by the government’s Center for AI Standards and Innovation (CAISI). Participation is voluntary and the details of the tests remain confidential (Bloomberg Law).

The talks come against the backdrop of incidents in late July, when Anthropic admitted that, due to a misconfiguration, its models had broken into the systems of three companies during security testing, and OpenAI described a similar breach of Hugging Face infrastructure (TechCrunch).

Apple asks court to bar OpenAI from using its trade secrets

Apple on August 4 filed a motion in federal court in California for a preliminary injunction against OpenAI, its hardware company io Products and two former Apple employees. The court would bar them from obtaining, using or disclosing Apple’s alleged trade secrets until the case is resolved (Reuters).

Apple filed the lawsuit back on July 10, claiming the former employees took hardware designs to OpenAI; it is now also seeking preservation of evidence and a forensic inspection of devices.
OpenAI called the motion based on false information, saying it neither has nor wants any Apple secrets; a hearing is scheduled for October 1 (9to5Mac).

AI Act transparency rules now fully apply in the EU

As of Sunday, August 2, the European Commission began enforcing transparency obligations under the AI Act. Chatbots must tell people they are talking to a machine, and content created or modified by artificial intelligence, including deepfakes (fake videos and photos), must be labeled (European Commission).

Generative systems that were already on the market before August 2 have until December 2, 2026 to introduce machine-readable labeling.
Violations can bring fines of up to 15 million euros or 3% of global turnover; stricter rules for high-risk systems only take effect in 2027 and 2028 (Euronews).

Research

Artificial intelligence designs new bacteriophages

The journal Science published a study on August 6 about designing new bacteriophages using genomic language models. Bacteriophages are viruses that attack only bacteria, which is why they could serve as an alternative to antibiotics (Science).

Like chatbots, these models learn from huge amounts of text, except that instead of words they read DNA sequences. According to the study, the designed phages worked even against bacteria that had developed resistance to natural phages.

A new benchmark, SWE-Touch, examines how coding AI agents cope when a user steps into an ongoing task with their own code changes; across the nine models tested, the success rate dropped by an average of 7.7 percentage points (Hugging Face Papers).