AI Bulletin · week of June 22

This week’s bulletin gathers six news blocks on AI applied to work, from 22 to 26 June 2026, plus a closing framework on the state of the AI economy. The week was marked by governments pressing on the frontier models, a fresh wave of hires moving to Anthropic, and several moves in infrastructure, video and agents.

1. The White House asks OpenAI to delay its next model

The White House has issued an official administrative request asking OpenAI to delay the public deployment of its next-generation frontier model, citing national-security and structural-safety concerns. According to TechCrunch, officials are pushing for an extended red-teaming window to audit the system’s advanced cyber-capability execution limits and its automated social-manipulation vulnerabilities.

In the same vein, the US administration is pressing Meta to submit its models for a voluntary review. Meta is the only major US AI developer that has not yet reached an agreement to share its models with the federal government; its policy team is negotiating with the Commerce Department.

The episode follows the withdrawal of Claude Fable 5 and Mythos 5 the previous week. This week, several signals pointed to a possible return of Fable 5: the model reportedly reappeared in Amazon Bedrock, and version 2.2.190 of Claude Code includes string changes that anticipate its comeback.

Sources: TechCrunch, 25 June — “The White House is asking OpenAI to slow roll the release of its new model over safety concerns”; reports via TLDR AI, 24 and 25 June — “US Presses Meta to Agree to AI Reviews as Security Concerns Rise” and “Fable 5 has now reportedly also reappeared in Amazon Bedrock”.

The bulletin mascot, as a reporter, raises a hand in a "wait" gesture beside a large hourglass, with the silhouette of a domed government building through the window — The White House asked OpenAI to delay the deployment of its next frontier model and is pressing Meta to submit its own for review.

2. Anthropic draws talent from Google and DeepMind

Nobel laureate John Jumper is leaving DeepMind for Anthropic after nine years at the company. Jumper co-led the AlphaFold team and won the Nobel Prize for predicting protein structures. His departure follows DeepMind’s struggles to sell its coding tools to businesses.

A few days later, Bloomberg reported that Gemini researchers Jonas Adler and Alexander Pritzel were also leaving Google for Anthropic. The moves continue a wave of high-profile departures, after Noam Shazeer’s exit the previous week, amid growing competition between the leading AI companies.

Sources: TechCrunch, 20 June — “Nobel laureate John Jumper is leaving DeepMind for rival Anthropic”; TechCrunch (citing Bloomberg), 24 June — “AI researchers continue to leave Google for its rivals”.

The bulletin mascot points at a wall map with two office buildings and a dotted arrow along which silhouetted figures walk from the smaller building to the larger one — After Noam Shazeer's exit, this week John Jumper (DeepMind) and two Gemini researchers joined the talent flow toward Anthropic.

3. Infrastructure and chips: SpaceX, Reflection AI and the Jalapeño chip

SpaceX signed a deal worth up to $6.3 billion with Reflection AI for access to its supercomputer Project Colossus. The agreement lets Reflection use Nvidia GB300 GPUs to train open-source AI models and reflects rising demand for compute; SpaceX aims to monetise its infrastructure by offering it to outside AI companies.

OpenAI and Broadcom unveiled Jalapeño, the first accelerator in a planned family of LLM inference chips optimised for performance per watt. According to the companies, the processor was designed in nine months with AI-assisted development and is intended for gigawatt-scale data-centre deployments.

In the background, an analysis this week placed Amazon and Google at the head of the race for power and data-centre capacity in the US through 2030, with Google closing the gap on Amazon.

Sources: CNBC, 22 June — “SpaceX signs computing power deal with open-source AI startup Reflection worth up to $6.3 billion”; reports via TLDR AI, 25 June — “Jalapeño: OpenAI’s new chip” and “As AI companies race for power, Amazon and Google have the lead”.

The bulletin mascot inspects a glowing chip, with server units on the oak desk and the silhouette of a rocket lifting off through the window — An infrastructure-heavy week: SpaceX's deal with Reflection AI for its Colossus supercomputer and the Jalapeño inference chip from OpenAI and Broadcom.

4. Generative video: Seedance 2.5 and the push from Chinese models

ByteDance unveiled Seedance 2.5, its new video generation model, able to create 30-second 4K clips from a single prompt. Users can supply up to 50 images, videos or audio clips as references to gain more control over the result. The model will arrive in China next month; ByteDance has not announced a date for other countries.

Earlier this week, VentureBeat placed HappyHorse 1.1, Alibaba’s video model, at No. 2 in the global rankings, ahead of OpenAI’s Sora and earlier versions of Seedance. HappyHorse is available on Alibaba Cloud Model Studio with a 40% launch discount for the first two weeks and covers text-to-video, image-to-video and editing.

Sources: CNET, 24 June — “ByteDance introduces new Seedance 2.5 video model”; VentureBeat, 23 June — “Alibaba’s AI video model rises to No. 2 in global rankings”.

5. Launches: Mistral OCR 4, GPT-5.5-Cyber and Liquid LFM 2.5

Mistral released OCR 4, a document-intelligence tool that extracts structured content with bounding boxes and confidence scores. It supports 170 languages, deploys in a single container and, according to the company, offers a 4x speed advantage with high accuracy, especially in low-resource languages.

OpenAI introduced a set of security tools: an updated Codex Security plugin, GPT-5.5-Cyber in limited release, the Daybreak Cyber Partner Program and an open-source initiative called Patch the Planet. Daybreak is described as a defensive cyber stack distributed through partners rather than giving broad, direct access to the model.

Liquid AI published LFM 2.5, a 230-million-parameter model with a non-transformer architecture. Despite its compact size, the company says it reaches parity with transformer models three times larger on edge reasoning and generation tasks.

Sources: Mistral, 24 June — “Mistral OCR 4: SOTA OCR for Document Intelligence”; TestingCatalog, 23 June — “OpenAI launches new security tools and updates GPT-5.5-Cyber”; Liquid AI, 26 June — “LFM2-5-230M”.

6. Agents and assistants: Claude Tag, Gemini and bidirectional voice

Anthropic introduced Claude Tag, a Slack-based workflow that lets teams assign tasks to Claude, connect it to tools and codebases and keep context across channels. The company says the system has become a core part of its internal operations, where its product team uses it to generate much of its code and to support analytics, support and debugging tasks.

Google launched native computer-use capabilities for Gemini 3.5 Flash: the model processes continuous screenshots to execute clicks, scrolls and typing across different software environments. OpenAI, for its part, began rolling out a bidirectional voice mode in ChatGPT, with a new audio model (Bidi 1) that speaks and listens at the same time. And Perplexity launched Computer for Counsel, a legal-operations tool to automate administrative research, document gathering and contract triage.

Sources: Anthropic, 24 June — “Claude Tag”; Google, 25 June — “Introducing computer use on Gemini 3.5 Flash”; TestingCatalog, 24 June — “OpenAI Prepares Bidirectional Voice Mode for Rollout on ChatGPT”; Perplexity — “Computer for Counsel”.

7. Framework of the week: the state of the AI economy

An Exponential View analysis that circulated this week put figures on the size of the market: the generative AI economy has produced $110 billion in sales over the past twelve months and is growing at a rate that puts annualised revenue above $175 billion. The piece argues that the supply side of the market is well understood, but the demand side is much harder to measure, and it examines how much of that revenue covers investment spending and what happens as token prices fall and token quality improves.

In parallel, a technical review of scaling laws recalled their role as one of the central empirical findings of deep learning: the relationship between compute, loss, model size and data, how it is used to allocate compute optimally, and also its limits.

Sources: Exponential View — “The state of the AI economy”; Lilian Weng, 24 June — “Scaling Laws, Carefully”.

Closing

The six blocks cover the White House’s request for OpenAI to delay its next model and the pressure on Meta, the wave of hires moving to Anthropic from Google and DeepMind, the infrastructure deals and the Jalapeño chip, the push from Chinese video models with Seedance 2.5, the launches from Mistral, OpenAI and Liquid AI, and the batch of agents and assistants from Anthropic, Google, OpenAI and Perplexity. The framework of the week gathers the published figures on the size of the AI economy. The items reflect what the companies and the cited reports have communicated; they do not include productivity data verified in specific organisations.