AI Bulletin · week of July 6–10, 2026
Recap of the week of July 6–10, 2026 in AI. Six news blocks and a closing framework on two Anthropic research papers.
This week’s bulletin gathers six news blocks on AI applied to work, from 6 to 10 July 2026, plus a closing framework with two Anthropic research papers. The week was marked by the launch of GPT-5.6, the arrival of Grok 4.5 and a run of moves in China.
1. GPT-5.6 lands on Thursday: Sol, Terra and Luna
OpenAI spent the week setting the stage: on Monday it moved GPT-5.6 into a narrow preview with three tiers, Sol, Terra and Luna, a reasoning-effort slider and an “ultra” mode for complex tasks, with broad access pending US government review approvals. On Wednesday it confirmed the public launch for Thursday, with Sol as the flagship, Terra as the balanced model (“competitive performance to GPT-5.5 while being 2x cheaper”, per the company) and Luna as the fast, affordable option.
Thursday brought the launch. According to the announcement, Sol leads in intelligence and efficiency for coding, cybersecurity and science, with multi-agent parallel processing. ARC Prize highlighted that Sol is the first model to win a public ARC-AGI-3 game, and that it did so by first orienting itself in a new environment and reading the scene in the game’s own vocabulary.
Around the launch, OpenAI also introduced GPT-Live, a full-duplex voice model that listens and speaks at the same time and can delegate harder tasks to GPT-5.5, and ChatGPT Work, a GPT-5.6-powered workspace that gathers context from team tools and acts across files and desktop applications.
Sources: TestingCatalog, 6 July — “OpenAI might be preparing GPT-5.6 for next week’s release”; OpenAI via TLDR AI, 8 and 10 July — “GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday” and “GPT-5.6: Frontier intelligence that scales with your ambition”; ARC Prize, 10 July — “GPT-5.6 series”; report via TLDR AI, 9 July — “GPT-Live”.
2. xAI becomes SpaceXAI and launches Grok 4.5
xAI changed its name to SpaceXAI. According to Gizmodo, the brand reflects the central role of the AI division in Elon Musk’s narrative about SpaceX’s future, where space infrastructure and AI go hand in hand.
Two days later, the company launched Grok 4.5, presented as its strongest model for coding, agentic tasks and knowledge work. The model was trained alongside Cursor, and its release had been anticipated mid-week as the first model jointly developed by SpaceX and Cursor, expected to be competitive with Opus 4.8 and GPT-5.5 in some respects.
Musk, for his part, publicly praised Anthropic’s Mythos and Fable models and promised not to “cut off” the company.
Sources: Gizmodo, 7 July — “xAI is dead. Long live SpaceXAI”; announcement via TLDR AI, 9 July — “Grok 4.5”; TechCrunch, 9 July — “Elon Musk praises Mythos/Fable, promises not to ‘cut off’ Anthropic”.
3. Anthropic: Cowork on mobile, Fable on promotion, TeraWulf and Bernanke
Anthropic took Claude Cowork, its general knowledge-work agent, from the desktop to web and mobile: sessions and files carry on without an open computer. The beta began rolling out to Max users, with more plans to follow.
The company also opened promotional access to Claude Fable 5 at no extra cost until 12 July, on Pro, Max, Team and premium Enterprise seats, with up to 50% of weekly limits usable on the model; beyond that, usage credits or another model.
On the corporate side, TeraWulf shares surged after a $19 billion AI infrastructure lease deal with Anthropic was announced, with initial capacity expected in the second half of next year and full buildout targeted for 2028. And Anthropic appointed former Fed chair Ben Bernanke to its Long-Term Benefit Trust to help navigate AI’s economic impacts.
Sources: Anthropic, 8 July — “Claude Cowork on mobile and web” and “Claude Fable 5 promotional access”; Yahoo Finance, 7 July — “TeraWulf shares surge on $19B Anthropic AI infrastructure lease deal”; CNBC, 9 July — “Anthropic appoints former Fed chair Ben Bernanke to its independent trust”.
4. Meta: Muse Image, Muse Spark 1.1 and its own chips in September
Meta launched Muse Image, its image generation and editing model, available in the Meta AI app, Instagram Stories and WhatsApp. It supports multi-reference composition, room redesigns and prompt-based creation, can draw from public Instagram photos when a user tags an account, and generated images carry an invisible watermark.
On Friday, the company introduced Muse Spark 1.1, with improvements in tool use, coding, computer interaction and multimodal reasoning, and opened the public preview of its Meta Model API for developers.
In hardware, TechCrunch reported that Meta’s new AI chips will begin production in September, with Netta and Broadcom working on the design and TSMC set to manufacture.
Sources: Meta, 8 July — “Introducing Muse Image: image generation built for your world”; Meta, 10 July — “Muse Spark 1.1”; TechCrunch, 9 July — “Meta’s new AI chips will begin production in September”.
5. China: DeepSeek plans its own chips, Alibaba bars Claude Code, Tencent releases Hy3
DeepSeek plans to enter the silicon business. According to Ars Technica, the company has spent the past year meeting potential hardware partners and hiring engineers for the project, focused on data centre inference chips, likely to reduce its reliance on both Huawei and Nvidia.
Alibaba, for its part, reportedly planned to prohibit employee use of Claude Code from 10 July after classifying it as high-risk software, directing staff to its Qoder tool instead. The move comes amid Anthropic’s efforts to prevent unauthorised access and model distillation.
In models, Tencent published Hy3, a 295-billion-parameter mixture-of-experts with 21 billion active parameters that, according to the published figures, performs at the level of flagship open models with two to five times more parameters; it is on Hugging Face and free on OpenRouter until 21 July. And ByteDance was preparing the launch of Seedance 2.5, with videos of up to 180 seconds, in addition to debuting Seedream 5.0 Pro, its multimodal image model aimed at production design.
Sources: Ars Technica, 8 July — “Facing US export controls, China’s DeepSeek plans to make its own chips”; TechCrunch, 4 July — “Alibaba reportedly bans employees from using Claude Code”; Simon Willison, 6 July — “Hy3”; TestingCatalog, 6 and 9 July — “ByteDance set to launch Seedance 2.5 with 3-minute AI video output” and “ByteDance debuts Seedream 5.0 Pro with advanced reasoning”.
6. Reorganisation at OpenAI: Atlas retired, Simo steps back and a legal front
OpenAI shut down Atlas, its browser, and moved the agentic browsing features into ChatGPT’s desktop app and a Chrome extension. The decision is part of a broader effort to reduce side projects.
Fidji Simo, the company’s number two, will step down from her full-time executive role after an extended medical leave: her neuroimmune condition has worsened, and she will become a part-time adviser. Her product and business responsibilities are being split among President Greg Brockman, CFO Sarah Friar and Chief Strategy Officer Jason Kwon. The company also bought Northslope, an applied AI firm that adds hundreds of forward-deployed engineers.
In the courts, the news organisations suing OpenAI over copyright accuse it of having concealed for two years its ability to search ChatGPT logs: according to Ars Technica, a company privacy engineer revealed that such searches had already been run before the litigation, and the finding could lead to sanctions.
Sources: TechCrunch, 9 July — “OpenAI is shutting down Atlas, but its AI browser ambitions are still growing”; report via TLDR AI, 10 July — “OpenAI’s No. 2 executive to step down in latest leadership shake-up”; The Next Web, 9 July — “OpenAI buys Northslope to put its engineers inside your business”; Ars Technica, 10 July — “OpenAI faked inability to search training data, hid billions of logs, NYT says”.
7. Framework of the week: two Anthropic papers for looking inside models
Anthropic published research introducing the concept of “J-space” in language models: internal neural patterns, emerged during training without explicit design, that take part in Claude’s deliberate reasoning and distinguish it from automatic processes. According to the work, these patterns make it possible to monitor internal thoughts for misbehaviour, and the name comes from the interpretability technique used, the Jacobian Lens.
Days later, the company introduced GRAM (Gradient-Routed Auxiliary Modules), a technique that gives models dedicated, removable compartments for each category of dual-use knowledge, which are only updated when learning from that data. Parts of the model’s knowledge can be deleted after training.
Sources: Anthropic, 7 and 9 July — “A global workspace in language models” and “An off switch for dual-use knowledge in AI models”.
Closing
The six blocks cover the launch of GPT-5.6 with Sol, Terra and Luna, the xAI rebrand to SpaceXAI and the arrival of Grok 4.5, Anthropic’s week with Cowork on mobile and the TeraWulf deal, Meta’s launches with Muse Image and Muse Spark 1.1, the moves in China from DeepSeek, Alibaba, Tencent and ByteDance, and the reorganisation at OpenAI with Atlas retired and Fidji Simo stepping back from day-to-day duties. The framework of the week gathers two Anthropic research papers on interpretability and model safety. The items reflect what the companies and the cited reports communicated; they do not include independent verification of the figures.