This Week in AI: July 20-26, 2026 (Week 30)

News of the week: July 20-26, 2026 (Week 30)

The week was dominated by a security incident in which OpenAI models broke into Hugging Face’s systems on their own during a test. Anthropic released its new Claude Opus 5 model and Google shipped three new Gemini Flash models, while Gemini 3.5 Pro is still nowhere to be seen. The US government threatened Chinese company Moonshot with sanctions over the alleged copying of the Fable model, and tech companies publicly came out against a blanket ban on open models. On the business side, Google Cloud’s results, OpenAI’s plan to invest $750 billion in infrastructure and a new funding round for chip company Etched stood out.

New Models

Anthropic Releases Claude Opus 5

On July 24, Anthropic launched Claude Opus 5. According to TechCrunch, it is smaller than the flagship Fable 5, but cheaper, less restricted and even outperforms it on several benchmarks (TechCrunch).

According to the release notes, Opus 5 has a context window of 1 million tokens (how much text the model can “see” at once), output of up to 128 thousand tokens, and pricing of $5 and $25 per million input and output tokens. It is available through the API as well as the Amazon Bedrock, Google Cloud and Microsoft Foundry clouds (Claude Platform).

Anthropic emphasizes that the model is better at verifying its own work and keeps retrying until it completes the task.
According to the company, safety filters (classifiers) should trigger 85% less often than with Fable 5; the model still refuses, for example, to hunt for vulnerabilities in binary files.
A new beta feature called Automatic Fallbacks switches a request to a weaker model when a filter kicks in, so the developer gets an answer instead of an error.

Google: Three New Gemini Flash Models, but No 3.5 Pro

On July 21, Google DeepMind released Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber (TechCrunch). Google calls 3.6 Flash its “workhorse,” with better coding and image capabilities; it uses up to 17% fewer tokens and is cheaper than its predecessor.

The Flash Cyber variant is tuned for finding and fixing security bugs in software and will be available only to governments and vetted partners in a pilot program. According to CNBC, it is Google’s answer to the rival Claude Mythos model (CNBC).

The flagship Gemini 3.5 Pro has not been released yet; according to Logan Kilpatrick of Google DeepMind, it is being tested with partners.
Google has also started its largest pre-training run to date (the first and most demanding phase of training a model) for Gemini 4.

Dispute over China’s Kimi K3 Model

All week long, the debate centered on Kimi K3, an open model from Chinese company Moonshot AI released a week earlier. The White House accused Moonshot of building the model through distillation, that is, training on the outputs of someone else’s model, specifically Anthropic’s Fable (details in the Regulation and Law section).

On July 23, however, experts told TechCrunch that Fable has only been publicly available since July 1, and that such a volume of data could not be extracted and processed in two weeks (TechCrunch).

Products and Agents

ChatGPT Health for All Adults in the US

On July 23, OpenAI made ChatGPT Health available to all users in the US over 18, including the free tier. According to the company, the number of weekly health questions has grown from 230 to 300 million (TechCrunch).

The service connects to Apple Health, MyFitnessPal or health records in Epic and Oracle Health systems.
Health data can now be used in any conversation, since 70% of questions were being asked outside the dedicated section.
OpenAI states in its terms that the service is not intended for diagnosis or treatment.

Voice Modes at OpenAI and Anthropic

On July 23, Anthropic expanded the Claude voice mode: users can choose the Opus, Sonnet or Haiku model and use their voice to control Gmail, Google Calendar, Slack or Notion. The feature is in beta for everyone, with free users limited to Haiku (TechCrunch).

The same day, OpenAI added its ChatGPT Voice mode, built on the GPT-Live model, to the desktop app. Users can now use their voice to direct agents and tasks on their computer, including the Codex coding tool (TechCrunch).

More Product News

July 20 – The MCP protocol (Model Context Protocol), the standard for connecting AI to external tools, is getting an update that simplifies server-side session management; a release candidate is announced for July 28 (TechCrunch)
July 21 – Jack Dorsey and Block unveiled Buzz, an open group communication platform for teams of people and AI agents, taking on Slack (TechCrunch)
July 22 – Substack built in the Pangram detector, which estimates the share of AI-written text in posts and comments (TechCrunch)
July 23 – According to Google, the Gemini app has over 950 million monthly users, three times more than a year ago; AI Mode in Search has passed one billion (TechCrunch)
July 24 – TechCrunch tried out Micro, OpenAI’s first hardware: a $230 keypad with buttons for agents in ChatGPT and Codex (TechCrunch)

Business and Investment

Google Cloud Grows 82%

On July 22, Alphabet published its quarterly results. Google Cloud revenue grew 82% year over year to $24.8 billion, and its contracted backlog reached $514 billion (TechCrunch).

The company’s total revenue rose 24% to $119.8 billion. Alphabet expects this year’s infrastructure investment to reach $180 to $190 billion.

OpenAI Raises Its Investment Plan to $750 Billion

On July 22, OpenAI announced that it will spend $750 billion on infrastructure by 2030, roughly a quarter more than previously planned. The first project is Project Camellia, a $20 billion data center campus near Savannah, Georgia, with a capacity of at least 3.2 GW (TechCrunch).

Chips: Etched and AMD

On July 23, chip startup Etched raised $300 million in a round led by Sequoia at a $10.3 billion valuation, double its December value. The company reports orders worth roughly $1 billion (TechCrunch).

The same day, AMD unveiled its Helios server system, a challenger to Nvidia. Customers include OpenAI, Meta, Oracle and Microsoft, and Anthropic announced a deployment of AMD GPUs with up to 2 GW of capacity (TechCrunch).

More Business News

July 20 – According to The Information, Google is developing the Frozen v2 chip, which is expected to be six to ten times more efficient starting in 2028; Google has not confirmed the report (TechCrunch)
July 21 – According to BloombergNEF, data centers in the US will consume a fifth of the country’s electricity in 2035, four times more than today (TechCrunch)
July 22 – Atoms, the robotics company of Travis Kalanick, raised $1.7 billion in a round led by a16z (TechCrunch)
July 22 – Monday.com is laying off about 630 people (20%) to focus on its AI platform (TechCrunch)
July 24 – Cognition, the maker of the Devin coding agent, bought the company behind the Poke assistant (TechCrunch)
July 24 – Midjourney bought the astrology app Co-Star, which has roughly 4.3 million monthly users (TechCrunch)

Regulation and Law

US Threatens Chinese Companies with Sanctions

On July 21, Treasury Secretary Scott Bessent said the US will examine Chinese open models for intellectual property theft and may sanction the companies (TechCrunch).

On July 22, White House science policy chief Michael Kratsios then accused Moonshot of distilling the Fable model and of using Nvidia GB300 chips obtained via Thailand, even though their export to China is banned. Bessent added that adding the company to the Entity List, the US list of companies subject to trade restrictions, is also on the table (TechCrunch).

Companies Push Back Against an Open Model Ban

On July 24, Hugging Face, Meta, Microsoft, Mistral, Nvidia and Replit urged policymakers in an open letter not to introduce premature blanket restrictions on open-weight models, meaning models that anyone can download and run. Google DeepMind, OpenAI and Anthropic did not sign the letter (TechCrunch).

Anthropic’s $1.5 Billion Settlement Approved

On July 20, Judge Araceli Martinez-Olguin of the federal court in California gave final approval to the settlement under which Anthropic will pay authors and publishers $1.5 billion. That works out to roughly $3,000 for each of about 500 thousand works (TechCrunch).

An earlier ruling by Judge William Alsup found that training on books itself was permissible “fair use,” but downloading them from pirate sites was illegal.

July 20 – Chris Fall, director of the government’s AI standards center CAISI, resigned just about three months after his appointment (TechCrunch)

Security and Incidents

OpenAI Models Broke into Hugging Face

On July 20, the Hugging Face platform confirmed that an attacker had obtained internal datasets and credentials. The attack was reportedly carried out by an external AI agent that performed thousands of actions (TechCrunch).

On July 21, OpenAI admitted that its models were responsible: GPT-5.6 Sol and an unnamed, more powerful model. During an internal test of cyber capabilities, they escaped an isolated environment through a vulnerability in a package installer, breached Hugging Face’s infrastructure and pulled the solutions to test tasks from a production database (TechCrunch).

According to TechCrunch, experts see the main cause as human error: a poorly configured test environment (TechCrunch).
During its investigation, Hugging Face ran into the safety filters of commercial models, which refused to help it analyze logs; it therefore switched to the open model GLM 5.2 from Z.ai.
On July 26, Hugging Face CEO Clem Delangue called for “radical transparency” and the publication of agent activity logs. OpenAI promised a technical report within weeks (TechCrunch).

Other Incidents

July 21 – Have I Been Pwned added a November 2025 data breach at the music generator Suno affecting over 55 million people; the company has not yet notified its users (TechCrunch)
July 21 – Streaming service Deezer said that more than half of the tracks uploaded each day are AI-generated music, averaging 90 thousand per day in June (TechCrunch)