The Week in AI: June 29 – July 5, 2026 (Week 27)

News of the week: June 29 – July 5, 2026 (Week 27)

The week belonged mainly to Anthropic. The company released the Claude Sonnet 5 model and the Claude Science product, and the US government lifted export restrictions on its Fable 5 and Mythos 5 models. Cloud giants Microsoft and Amazon are investing billions in teams of engineers who help companies actually deploy AI. At the same time, Washington is tightening the reins: it is preparing voluntary rules for releasing the most powerful models, and according to the FT, OpenAI has proposed that the government take a stake in AI companies.

New Models

Claude Sonnet 5: a cheaper path to agents

Anthropic released Claude Sonnet 5 on June 30, the successor to Sonnet 4.6. According to the company, it is the “most agentic” Sonnet model yet, with performance approaching the more expensive Opus class (Anthropic, TechCrunch).

An agent in this sense is an AI that carries out multi-step tasks on its own, for example editing code, running tests and fixing bugs. TechCrunch described the release as a cheaper way to run such agents.

Price: until August 31, an introductory price list applies: 2 dollars per million input and 10 dollars per million output tokens (a token is a small piece of text, roughly part of a word).
Availability: the model is available in all plans and in the API as claude-sonnet-5; it has become the default model in the Free and Pro versions.
Safety: Anthropic also published a system card, a document with the results of the model’s safety testing.

Meituan unveils the trillion-parameter LongCat-2.0

Chinese company Meituan, best known for food delivery, unveiled the LongCat-2.0 model with 1.6 trillion parameters on June 30. It uses a Mixture-of-Experts architecture, in which only a smaller part of the network, roughly 48 billion parameters, is activated at each step (NYU Shanghai RITS).

The key point is that, according to the company, the model was trained and runs exclusively on chips designed and manufactured in China, without a single GPU from Nvidia. The model weights are to be released openly under the MIT license, but their publication came only after the week covered here.

Products and Agents

Claude Science heads for labs and pharma

Alongside Sonnet 5, Anthropic launched the Claude Science application on June 30. It is designed to automate a range of tasks in biology and chemistry and is aimed primarily at researchers at pharmaceutical companies (Bloomberg, STAT).

The product builds on the April acquisition of the startup Coefficient Bio and on the arrival of John Jumper, Nobel laureate in chemistry and a leading figure behind AlphaFold, who moved from Google DeepMind to Anthropic in June.

California gives officials Claude at half price

Governor Gavin Newsom announced a deal on June 29 under which California state agencies can use Claude at a fifty percent discount. The deal includes free staff training and technical assistance from Anthropic (Office of the Governor of California).

According to Newsom, AI is not meant to replace the human work of agencies but to help staff work faster. The state had already used Claude, for example, in developing the internal tool Poppy.

Smaller product news

July 2 – Cognizant and OpenAI launched a cyber defense service based on GPT-5.5, designed to find vulnerabilities and propose validated fixes (Cognizant).
July 2 – OpenAI restricted model fine-tuning: new jobs can only be started by organizations that have actually used a fine-tuned model in the past 60 days (OpenAI Developers).
July 1 – Cloudflare split AI bots crawling the web into Search, Agent and Training categories, so website owners can allow search while blocking data collection for training (Business Wire).

Business and Investment

Microsoft and Amazon send engineers straight into companies

Amazon Web Services announced on June 30 a new organization of so-called forward deployed engineers with a budget of one billion dollars. Small teams of five to six engineers will work inside customer companies in 45-day sprints to build and launch agentic systems on the customers’ own data (Reuters / The Star).

Two days later, on July 2, Microsoft responded with the Microsoft Frontier Company unit, with a budget of 2.5 billion dollars and roughly 6,000 specialists. It will be led by Rodrigo Kede Lima, until now head of the Asia division (Microsoft).

Both companies are thus copying a model previously introduced by OpenAI and Anthropic. It shows that the main obstacle to AI adoption today is not the models themselves but integrating them into business processes.

Nvidia to finance clouds and take a cut of revenue

Nvidia unveiled a model on July 1 under which smaller AI cloud operators will get its servers with credit support. In return, Nvidia will receive not only the regular hardware price but also a share of the revenue from the compute capacity the clouds sell (Nvidia, The Next Web).

The first partners are Sharon AI and Firmus. Critics point out that Nvidia is becoming more of a lender and would feel a double hit if demand for AI cooled.

Meta wants to sell excess compute

According to Bloomberg on July 1, Meta is building a cloud business as part of its Meta Compute initiative. It is considering selling access to models on its infrastructure, including its own Muse Spark models, or renting out bare compute capacity (Bloomberg).

Meta would thus enter the fight with AWS, Azure and Google Cloud. The company plans to invest 115 to 135 billion dollars this year, so unused capacity could bring in revenue.

In the same week, on July 2, Mark Zuckerberg told employees at an internal meeting that development of AI agents over the past four months has not accelerated as quickly as the company expected. He expects more significant returns on the investments within three to six months (Reuters / The Star).

Other deals and funding

July 2 – Kuaishou announced separate funding and a restructuring of its video AI Kling AI; reports say it is a round of around two billion dollars with investors including Tencent and Alibaba (Yicai Global).
July 2 – According to The Information, Anthropic is in talks with Samsung about developing its own AI chip; the talks are at an early stage and the company has not confirmed them (TechCrunch).
July 2 – Amazon is designing its own AI chips for Echo and Fire TV devices, said devices chief Panos Panay (CNBC).
July 2 – ElevenLabs, a voice synthesis company, is exploring an employee share sale at a valuation of roughly 22 billion dollars (Reuters).
July 2 – Tripo AI, a developer of models for 3D content, raised 150 million dollars in a Series A3 round (GlobeNewswire).
July 2 – Zoom is acquiring Common Room, a company that uses AI agents to gauge the interest of potential customers (GlobeNewswire).

Regulation and Law

US lifts export restrictions on Fable 5 and Mythos 5

The US Department of Commerce on June 30 withdrew the export controls that on June 12 had blocked foreign nationals from accessing the Claude Fable 5 and Claude Mythos 5 models. Secretary Howard Lutnick said Anthropic had agreed to actively identify and address the models’ security risks (Fox Business, Anadolu Agency).

The models were offline for about 18 days. Starting July 1, Anthropic began redeploying them with new protective classifiers, filters designed to catch attempts at misuse (Anthropic).

White House prepares rules for releasing the most powerful models

According to the Financial Times on July 1, the White House is in advanced talks with OpenAI, Google and Anthropic on a voluntary framework that would unify the review and release of so-called frontier models, the most capable AI systems (Financial Times).

The talks follow the June postponement of the full launch of GPT-5.6, which the government asked OpenAI for. Neither the companies nor the White House commented on the report to Reuters.

OpenAI proposes a government stake in AI companies

According to the FT, followed up on July 2 by CNBC and The Guardian, OpenAI CEO Sam Altman proposed that leading AI companies set aside 5% of their equity for a government fund similar to the Alaska Permanent Fund. For OpenAI, this would amount to roughly 42.6 billion dollars (CNBC, The Guardian).

The proposal is at an early stage and could require approval from Congress. Reuters reported the same day that the Trump administration and Anthropic had not discussed any government stake (Reuters).

FTC wants to penalize hidden “steering” of AI answers

The US Federal Trade Commission (FTC) on July 1 published a proposed policy statement on so-called suppression of AI accuracy. Companies that covertly distort the outputs of their systems to serve undisclosed ideological goals could, according to the FTC, be engaging in deception under Section 5 of the FTC Act (Inside Privacy / Covington).

The proposal is not a binding regulation and does not in itself create new obligations.
The commission is accepting comments until July 31; the FTC also states that state laws do not create an exemption from this rule (Bloomberg Government).

The Week in Numbers

2.5 billion dollars and 6,000 people for the new Microsoft Frontier Company unit.
1 billion dollars for AWS engineering teams embedded with customers.
1.6 trillion parameters in the LongCat-2.0 model, trained only on Chinese chips.
5% of AI companies’ equity could go to the US government under the OpenAI proposal.
50% discount on Claude for California state agencies.