Anthropic and OpenAI are each launching a dedicated vehicle to sell AI to large enterprises. The strategy: partner with investment funds like Blackstone or Goldman Sachs, which each own hundreds of mature large companies (hotels, industry, healthcare, real estate…), to sell them AI in bulk, instead of cold-calling each company one by one.
What you need to know
- Anthropic: On May 4, 2026, Anthropic announced a joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs as founding partners.
- OpenAI: According to Bloomberg, OpenAI is preparing a similar structure, The Development Company, targeting a $4 billion raise at a $10 billion valuation.
- The mechanism: These funds are not startup incubators. They are financial giants that buy and own established large companies. Anthropic and OpenAI go through them to access all these companies and place their engineers directly with the client.
Stakes and outlook
The lever here is not technological but commercial: going through the funds short-circuits the long sales cycles of large accounts and accelerates contract signing. It reinforces the « forward-deployed engineers » model popularized by Palantir, where technical teams work with business units to embed AI into existing workflows. In the medium term, competition between AI labs may shift toward industrial deployment capacity. The most sought-after skills will combine software engineering, business domain understanding, and operational integration.
Microsoft is introducing two multi-model capabilities in Researcher, the deep-research agent of Microsoft 365 Copilot: Critique and Council. The goal is to improve the accuracy, depth of analysis, and reliability of reports produced in professional contexts, against active competition (OpenAI Deep Research, Gemini Deep Research, Perplexity).
What you need to know
- Critique by default: Critique becomes the default Researcher experience when the Auto option is selected, with one model in charge of producing the report and another in charge of evaluating and improving it.
- Measured results: According to Microsoft, Critique scores 7.0 points higher (+13.88%) than Perplexity Deep Research equipped with Claude Opus 4.6 on the DRACO benchmark, measured across 100 complex tasks covering 10 domains.
- Council in comparison: Council runs an Anthropic model and an OpenAI model in parallel, then calls on a judge model to summarize agreements, divergences, and points specific to each response.
Stakes and outlook
Microsoft is reinforcing Researcher as a synthesis tool for demanding professional uses, with an emphasis on sources, citations, and completeness. Separating generation from evaluation moves the process closer to standards used in academic research or business analysis. For enterprises, this can offload part of the initial verification work, without removing the need for human validation. These multi-model architectures may also shift the expected skill set: framing requests well, comparing results, and auditing sources become more decisive. The AI assistant industry is moving toward systems less centered on a single model, and more on the orchestration of several specialized models.
Disney and OpenAI saw their partnership around Sora collapse less than four months after its announcement, according to the Los Angeles Times. The episode casts doubt on a rapid Hollywood adoption of AI-generated video, while actors’ and writers’ unions remain watchful of these uses.
What you need to know
- Deal abandoned: On December 11, 2025, Bob Iger and Sam Altman had presented an agreement including a $1 billion Disney investment in OpenAI and the controlled use of Disney characters in Sora, the video generation model launched in late 2024.
- Sora shut down: OpenAI announced on March 24 that Sora would be discontinued, while according to the LA Times, no payment had been made and no licensing contract had been signed between the two groups.
- Declining use: According to market researchers cited by the newspaper, Sora’s monthly downloads dropped from over 6 million in November to just over 1 million in February, amid intensifying competition (Veo, Runway, Kling).
Stakes and outlook
Disney avoids extending a project already contested by creators and by the Writers Guild of America, which saw it as legitimizing the use of works without authorization, an issue already at the heart of the 2023 strike. OpenAI is also reallocating resources to other priorities, in a context where Sora was consuming significant compute capacity. For the industry, the signal is twofold: demand for AI-generated video remains uncertain, and rights control stays central. That said, studios will continue to explore AI to reduce production costs, which will heighten tensions around employment, creative skills, and negotiations with actors’ and writers’ unions.
On April 13, 2026, France’s CERT-FR (national cybersecurity agency) issued an alert on the risks associated with agentic AI assistants installed on workstations. The immediate stake is loss of control over the information system when these tools execute actions with user privileges.
What you need to know
- Rapid adoption: Since early 2026, tools like OpenClaw, Claude Cowork, and other open-source solutions have seen sustained adoption, according to CERT-FR.
- Expanded attack surface: These agents can execute commands, control a browser, read or write files, manage a calendar, or send emails.
- Strict recommendation: CERT-FR states that these products, often in beta, must not be deployed in production and must be confined to isolated test environments.
Stakes and outlook
These agents can cause data leaks, exfiltration of authentication secrets, or unauthorized command execution. Prompt injections and message hijacking remain intrinsic weaknesses of the models, even when guardrails are built in. The most structural risk is the rise of Shadow IT (unmanaged AI usage) that is difficult for IT and security teams to oversee. CERT-FR recommends technical hardening, allowlists, human validation for sensitive actions, and sandboxed execution.
OpenAI has launched a « Trusted Contact » option in ChatGPT, allowing a designated adult to be alerted if a user seriously mentions self-harm or suicide. The immediate goal is to connect at-risk situations to human support more quickly, without transmitting the content of the conversations.
What you need to know
- Optional feature: The user can add an adult aged 18 or over from ChatGPT settings, and this contact must accept the invitation within a week.
- Controlled alert: If OpenAI’s systems detect a serious risk, the user is notified first, then a dedicated internal team reviews the situation before any notification is sent.
- Limited data: The contact receives a brief message encouraging them to reach out to the user, without any transcript or detail of the conversation.
Stakes and outlook
This feature arrives as OpenAI faces growing legal and public pressure over the handling of psychological distress in ChatGPT. Families have filed several lawsuits following user suicides, while Florida is investigating alleged links between ChatGPT and certain dangerous behaviors. The tool can create a faster human relay in critical situations, but it remains optional and dependent on user consent. OpenAI is following the trajectory of other consumer AI publishers already under scrutiny, like Character.AI, sued in October 2024 after the suicide of a 14-year-old teenager. The expected standard of care is shifting: simple content moderation is no longer enough, and the real-world effectiveness of detection and relay mechanisms becomes a central criterion, including for regulators.
Anthropic has unveiled « dreaming », a new capability of its AI agents platform that lets these agents analyze their past sessions to improve themselves. The immediate stake is to make agents more reliable for real-world enterprise use, particularly on long and complex tasks.
What you need to know
- Learning without retraining: « Dreaming » does not modify the model itself, but produces notes and how-to guides that future sessions can consult.
- Expanded features: Two other features move from restricted to broader access: « outcomes » (one agent verifying another’s work) and the coordination of multiple agents on the same task.
- First customer feedback: According to Anthropic, Harvey (legal AI platform) saw the number of completed tasks multiply by roughly 6, and Wisedocs (medical record review) cut document analysis time by 50%. These figures come from early-access customers, shared at the time of launch.
Stakes and outlook
These features aim to reduce human intervention in verifying and correcting results, a known weak point of AI agents once they leave the lab. The « outcomes » system adds a separate evaluator agent, in charge of checking the result against a grid defined by the developer. Coordinating multiple agents lets a task be split across several specialists, each with their own scope. Anthropic thus positions itself on ground shared by the entire industry: OpenAI released its own agent-building solution in March 2025 (Agents SDK), Microsoft is pushing Copilot Studio, and several open-source tools like LangGraph are progressing fast. Differentiation no longer hinges on raw model power, but on the quality of the memory, evaluation, and coordination layers around it.
Google has unveiled Gemini Intelligence, a version of Gemini built into Android 17 that accompanies the user in their mobile environment and acts directly through their apps. The immediate stake is moving AI from a separate chatbot to an operational assistant across the whole phone, from Gmail to Chrome and third-party apps.
What you need to know
- Built-in assistant: Gemini Intelligence is not a standalone app but a feature embedded directly in the Android system. It will not be active at Android 17’s launch and is expected to arrive this summer on Galaxy S26 and Pixel 10.
- Delegated tasks: Google indicates the AI will be able to, for example, find reservations based on a schedule received in Gmail, fill a shopping cart, or book a cycling class by directly operating apps on screen.
- Faster input: Intelligent Autofill will fill forms from messages or documents, and Rambler (voice dictation corrector) will remove hesitations and repetitions in the Google keyboard (Gboard).
Stakes and outlook
The main lever is time saved on repetitive digital micro-tasks. During its presentation, Google showcased spectacular scenarios, like booking a concert seat in one click. The reality check will be tougher: choosing the right seat, validating the price and date, handling anomalies and edge cases are all steps where the agent can stumble. The trend is now shared by all major mobile players: Apple Intelligence (iPhone, late 2024), Galaxy AI (Samsung S series, since 2024), and now Gemini Intelligence are all betting on deep system integration rather than a separate app.
1. What is it?
Claude Design turns a brief, a product idea, or a visual intent into concrete proposals: interfaces, mockups, brand kits, templates, creative concepts, or storytelling angles. The goal is to save time on the early phases of creation, by moving faster from idea to prototype.
2. Why is it fascinating?
The value spans every company size. For an SME, it’s a complete brand kit (logo, palette, typography, applications across business cards, social media, email signatures) without going through an agency. For a large group, it’s industrialization: pitch deck templates, quarterly reports, internal materials, all aligned with the visual guidelines. And for a startup, it’s describing your medical booking app and getting back, in one go, screens, user journeys, and visual tone recommendations to test.
3. Why is it limited?
As shown by the Sora setback highlighted by the Los Angeles Times, generative AI often impresses in demos, but disappoints when it comes time to produce something useful, original, and lasting. Claude Design can speed up ideation, but it does not replace a designer’s eye, deep brand understanding, user testing, or legal constraints tied to generated content. The real risk: producing clean but generic designs, effective for brainstorming, insufficient to build a differentiated experience.
