The Week in AI: July 6–12, 2026 (Week 28)

News of the week: July 6–12, 2026 (week 28)

The week was dominated by the public launch of OpenAI’s GPT-5.6 model family, which was preceded by a review from the US Department of Commerce. Meta began selling access to its model through a paid API for the first time and announced production of its own AI chip, while Microsoft started replacing partner models in Office with its own. On the legal front, Apple’s trade secrets lawsuit against OpenAI and a Chinese warning about the Claude Code tool drew attention. Financial markets watched SK Hynix’s record Nasdaq debut, and Anthropic signed another multibillion-dollar data center lease.

New Models

OpenAI launches GPT-5.6 in three sizes

OpenAI announced the GPT-5.6 Sol, Terra and Luna models on Tuesday, July 7, and released them publicly on Thursday, July 9. Sol is the new flagship, Terra is a balanced model for everyday work, and Luna is the cheapest option (OpenAI, Neowin).

The road to release was unusual. Since late June, only a limited circle of vetted US partners had access to the models, and the wider launch was approved by the US Department of Commerce only after further testing and talks with government agencies (TechXplore).

API pricing (per million input / output tokens, the chunks of text a model processes): Sol $5 / $30, Terra $2.50 / $15, Luna $1 / $6 (ODSC).
Alongside the models, OpenAI introduced the ChatGPT Work agent, which combines the Codex coding tool and a browser into a single app for multi-hour tasks, with integrations for Google Drive, Slack and Salesforce (ODSC, iThinkDiff).

GPT-Live: a voice assistant that can listen and talk at the same time

On July 8, the day before GPT-5.6, OpenAI released the voice models GPT-Live-1 for paying users and GPT-Live-1 mini for free accounts. They are so-called full duplex models: they can speak and listen simultaneously, so you can interrupt them just as you would a person (OpenAI, eWeek).

At launch, voice mode could not be combined with video or screen sharing. OpenAI presents it as a step toward making voice a full-fledged way to control AI.

Meta sells its model through an API for the first time

On July 9, Meta released the Muse Spark 1.1 model from Meta Superintelligence Labs and at the same time launched a public beta of the Meta Model API. It is the company’s first paid self-service interface that lets developers integrate the model into their own applications (Meta AI, American Bazaar).

The model focuses on coding, computer use and working with external tools. It marks a departure from the earlier strategy of open Llama models: the Muse Spark 1.1 weights are closed, and access is so far limited to US developers, with $20 in free credit.

Other models of the week

July 6 – Tencent released the full version of its open Hy3 model with 295 billion parameters, of which only 21 billion are active for each query (a mixture-of-experts architecture, in which the model uses only selected “expert” parts). The Apache 2.0 license permits commercial use as well (VentureBeat, Gigazine).
July 8 – xAI made available the Grok 4.5 model, aimed at coding and agentic tasks at a low price of $2 / $6 per million tokens. Elon Musk described it as an “Opus-class model, but faster and cheaper” (xAI, ODSC).

Products and Agents

Microsoft starts replacing OpenAI and Anthropic with its own models

According to a July 7 report by Bloomberg, Microsoft is moving some features in Excel and Outlook from OpenAI and Anthropic models to its own MAI models. The main reason is cost (Bloomberg, SiliconANGLE).

The head of Microsoft AI, Mustafa Suleyman, said Anthropic is “extremely expensive” and that many companies are looking for alternatives. So far it involves tens of thousands of queries a week, a small fraction of traffic, and Microsoft says it is not ending the partnerships.

Claude Cowork on the web and mobile

On July 7, Anthropic made Claude Cowork, a tool for independently carrying out longer work tasks, available on the web and in its mobile apps as well. The beta is rolling out gradually, starting with subscribers to the Max plan (Claude, Claude release notes).

July 10 – Anthropic reworked Claude’s memory: remembered information is now stored in categorized entries (Claude release notes).

Business and Investment

SK Hynix debuts on Nasdaq with a record offering

South Korean memory chipmaker SK Hynix, a key supplier of HBM memory for AI accelerators, began trading on the Nasdaq on July 10 in the form of American depositary receipts (ADRs). The offering raised about $26.5 billion and is the largest US share sale by a foreign company in history (Korea JoongAng Daily, Quartz).

The receipts were priced at $149, opened at $170 and closed their first day at $168.01, up about 12.8%. The debut shows how strong investor appetite remains for the “shovels” of the AI gold rush.

Meta prepares to manufacture its own AI chip

According to an internal memo seen by Reuters on July 9, Meta plans to start producing its own AI chip in September. It is designing the chip with Broadcom, TSMC will manufacture it, and testing took only six weeks (Reuters).

This year Meta plans to deploy computing infrastructure with 7 gigawatts of power capacity and to double it to 14 GW in 2027.
It aims to release a new chip roughly every six months and will spend up to $145 billion on AI infrastructure this year.

Anthropic to lease a data center for $19 billion

On July 6, TeraWulf announced a twenty-year lease of its campus in Hawesville, Kentucky, to Anthropic. The contract is expected to bring in about $19 billion over the initial term (TeraWulf, The Block).

The campus is planned to have 401 MW of IT capacity, with operations starting in the second half of 2027. TeraWulf shares rose about 17% after the announcement.

July 9 – Ollama, a tool for running open AI models on your own computer, raised $65 million in a Series B round led by Theory Ventures. In total, the company has raised $88 million (HPCwire).
July 9 – Former US Federal Reserve chair Ben Bernanke joined the Long-Term Benefit Trust, an independent body that oversees Anthropic’s adherence to its mission and elects a majority of its board of directors (Anthropic, CNBC).

Regulation and Law

Apple sues OpenAI over trade secrets

On July 10, Apple filed a lawsuit in federal court in Northern California against OpenAI, its hardware subsidiary io Products and two former Apple employees. It claims that departing engineers took confidential hardware information that was meant to help OpenAI develop its own devices (TechSpot, OECD.AI).

According to the lawsuit, more than 400 former Apple employees now work for OpenAI. Apple is seeking an injunction against the use of its secrets and the return of materials; OpenAI denies the allegations.

The New York Times accuses OpenAI of hiding evidence

On July 9, The New York Times and other publishers filed a motion for sanctions against OpenAI in their copyright dispute. They claim the company falsely stated for two years that it was unable to search its training data and conversation logs (TechCrunch, Ars Technica).

According to the plaintiffs, a deposition revealed that OpenAI had internally searched its data for news articles. For now these are the claims of one side, and the court has not ruled on the motion.

China warns about Claude Code

On July 8, China’s national vulnerability database, run by the Ministry of Industry and Information Technology (MIIT), warned that Claude Code versions 2.1.91 through 2.1.196 may send users’ location and identity data to Anthropic’s servers without consent. It recommended uninstalling these versions (South China Morning Post, Quartz).

Anthropic responded that it was an experimental safeguard against abuse, for example against unauthorized resellers and distillation (copying a model’s capabilities by training another model on its outputs). According to the company, Claude Code has never been officially available in China. Alibaba banned its employees from using the tool as of July 10 (GovInfoSecurity).

Research

Anthropic finds a “global workspace” inside Claude

On July 6, Anthropic published a study on interpretability, that is, on what happens inside a neural network. It describes a small internal region called J-space, where the model holds concepts during multi-step reasoning, much like working memory (Anthropic).

The researchers uncovered it with a tool called J-lens, drawing inspiration from the neuroscientific global workspace theory. When they suppressed J-space, complex reasoning collapsed, while ordinary conversation kept working.

The study explicitly makes no claims about whether the model is conscious. Anthropic does, however, see a use in model oversight, because J-space can capture “thoughts” that never appear in the response.

AI beats the world’s best programmers

At the AtCoder World Tour Finals 2026 in Tokyo (July 9), a system from OpenAI solved all five problems in an exhibition contest of the algorithm division and scored 8,300 points. The best human scored 4,300 points (The Decoder, Understanding AI).

The system also had a large lead in the heuristic division. As recently as 2025, AI finished second behind a human.

Security and Incidents

July 9 – According to the GPT-5.6 system card, the UK AI Security Institute found universal jailbreaks (techniques that bypass a model’s safety guardrails) for GPT-5.6 Sol during pre-release testing. They unlocked long agentic tasks involving vulnerability hunting and exploit writing; OpenAI says it fixed the reported attacks before release (ODSC).
July 8 – China’s warning about Claude Code sparked a debate about how much telemetry AI tools may send without users’ knowledge (see the Regulation and Law section).