This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.
If you follow AI for work, research, investing, or just to understand where the technology is heading, this weekly brief is a concise way to scan the most important launches, risks, and resources in a few focused minutes.
During the week of June 21–27, 2026, major developments across AI spanned infrastructure, models, and applications. Qualcomm announced a nearly $4 billion acquisition of Modular, the AI chip software platform founded by Chris Lattner, highlighting tighter hardware–software integration. OpenAI previewed GPT-5.6 (Sol, Terra, Luna) amid U.S. cyber risk coordination, while export restrictions eased for Anthropic’s Claude Mythos 5 but not Fable 5. ByteDance launched Seedance 2.5 for 4K, 30-second video generation.
On tooling, OpenRouter introduced a unified image API across 30+ models, NVIDIA adopted LiteLLM as the internal gateway for its engineers to access 100+ models, Sakana AI introduced Fugu, a multi-agent orchestration system accessible through a single OpenAI-compatible endpoint. Vedang Vatsa published a mapping of the AI agent infrastructure ecosystem covering 75+ companies and $14B in funding. Industry commentary focused on AI’s impact on jobs, creativity, and infrastructure demand, alongside rising Chinese model usage and forecasts of increasing convergence between real and synthetic media.
The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.
via Huzefa
Claude Predicts a Blur in Truth
Claude, the AI system, thinks that in about five years it might become really hard to tell what’s actually true anymore. With all the deepfakes and misinformation growing, the line between real and fake could get pretty blurry. It’s a warning about how fast tech is changing and how it might shake up our perception of reality.
via Yuki Endo
Hello. Japan's Sakana AI just dropped Fugu.
A full multi-agent orchestration system accessible through a single API. The interesting part isn't another benchmark race—it's the idea that the future may belong to coordinated intelligence rather than a single giant model.
Definitely worth trying. This could be the next big shift in AI infrastructure. This is smart as security layer
various discussions on the above primarily around pricing points
Prashish: This is much needed, a middle ground between $20 and $100. $20 usage is low, and $100 is way high for normal everyday use
Lane Rettig: This pricing model is the one that nearly all big AI projects seem to have landed on.
via Y Yang
This is exactly the kind of problem I am trying to understand better.
For context, I am working on InferHub, an OpenAI-compatible multi-model API gateway. The idea is to make it easy to switch between models like GPT, Claude and Gemini through one endpoint.
If your main issue is provider switching, cache misses, cost, or fallback behavior, I would be happy to give you free credits to test it. No pressure to use it long term; I am mainly looking for feedback from people building real AI apps.
Your use case sounds like: "I think I’m hitting the limit of the Github Copilot + OpenRouter workflow financially. My setup: PhpStorm Github Copilot extension OpenRouter mostly DeepSeek V4 Flash in ag"
via David An
hi everyone, I am sharing my presentation on the research topic of "HOW CAN ARTIFICIAL INTELLIGENCE DRIVE HUMAN CREATIVITY?
Provocation Lab Research Brief #1 - Human-AI Collaboration in Creativity | Dr. David An | 14 comments
HOW CAN ARTIFICIAL INTELLIGENCE DRIVE HUMAN CREATIVITY? This is what I discuss by introducing a research study written by the Institute of AI in Management Ludwig-Maximilians-Universität München.
Qualcomm plans to acquire Modular, which builds a chip software platform and has a proprietary coding language, in a nearly $4B deal set to close in H2 2026
Modular was founded by Chris Latner who created LLVM as well as the Swift programming language and was briefly head of Tesla's Autopilot software
bit.ly/44rBpe5
Qualcomm Buys Buzzy Chip Startup Modular for Nearly $4 Billion
Modular, one of the most promising chip software startups of the AI era, heads for a multibillion-dollar exit.
ByteDance unveils Seedance 2.5, saying the AI video model can generate up to 30-second clips from up to 50 reference materials, up from 12 for Seedance 2.0
> more than 30 seconds
> 4k resolution
> prompt adherence is on another level that we never seen before
> our reports was right , this is why follow us and join our server
> July will be do much fun
→ 30-second native video output in a single clip
→ Up to 50 multimodal reference inputs
→ Native 4K video generation
Stay tuned
> ByteDance just announced it, general release early July
> 30 second videos in a single generation
> references up to 50 inputs — images, video, audio
> 4K output, edit the whole thing with just text
> "make a 30s CM from this product photo +
ByteDance has officially confirmed Seedance 2.5 as the model name> Early July Launch Date
Key features:
- 30-second single-shot native video output.
- Reference material capacity greatly increased.
- More controllable video generation and editing
Cc: @MarsForTech
OpenRouter launched a unified Image API providing access to 30+ image generation models from 8 providers including Google, OpenAI, Black Forest Labs, Recraft, ByteDance, and xAI. Features include a single normalized request format across providers, transparent per-image USD pricing in responses, optional provider-specific parameters, and SSE streaming support for models like OpenAI's GPT Image.
Unified access to 30+ image gen models from 8 providers: Google, OpenAI, Black Forest Labs, Recraft, ByteDance, Sourceful, Microsoft, and xAI
Plus, a new solution to image API heterogeneity
via Casey Newton
Q&A with AWS CEO Matt Garman on the parallels between early AWS and AI, Quick, AI coding, Amazon's $200B capex in 2026, entry-level jobs changing, and more
The CEO of AWS on why Amazon is hiring 11,000 interns and junior employees
Matt Garman argues that junior employees are as necessary as ever. But AWS now sells agents that can recruit, code, and process claims. Will the balance hold?
OpenAI Launches Limited GPT-5.6 Preview at U.S. Government Request
The company introduced Sol, Terra, and Luna through its API and Codex platform, with Sol leading benchmarks on complex command-line tasks like vulnerability research. This limited preview follows federal guidance and a Trump administration executive order on reviewing AI systems with cyber risks, including recent meetings between CEO Sam Altman and Commerce Secretary Howard Lutnick. OpenAI previewed the models with officials beforehand and anticipates wider access soon, amid debates over whether the process stifles innovation or wisely addresses misuse risks.
Previewing GPT-5.6 Sol: a next-generation model
OpenAI previews GPT-5.6 Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with its most advanced safety stack.
via Alex
Percentage of chinese models' token share on OpenRouter up 3x over a year
Haseeb of DragonFly has his own strong take on what inferences can be drawn from this OpenRouter data and I agree to a large extent. US Frontier Lab usage isn't showing up on OpenRouter
Problems with drawing inferences based on raw token usage on OpenRouter:
* Chinese labs routinely launch new models with high subsidies or even free usage for the first couple of weeks. This attracts
US. Government Lifts Export Block on Anthropic's Claude Mythos 5 Model for Over 100 Approved U.S. Companies and Agencies
The Trump Administration, via Commerce Secretary Howard Lutnick's letter to Anthropic co-founder Tom Brown, partially lifted export controls on Anthropic's Claude Mythos 5 AI model, permitting its release to more than 100 trusted U.S. companies and federal agencies after weeks of negotiations over national security concerns
Restrictions on Fable 5 remain in place, with Anthropic committing to collaborate on future model release protocols and standards.
How Thousands of NVIDIA Engineers Access 100+ AI Models Through a Unified Inference Service
NVIDIA engineers faced fragmented access to 100+ AI models across cloud providers, open source, and internal services with different APIs and credentials.
The IT Inference team built the Enterprise Inference Hub centered on LiteLLM as a unified gateway.
Applications use one OpenAI-compatible API and single credentials while LiteLLM handles routing, authentication, and metrics.
Centralized control enables usage monitoring, token accounting, cost visibility, budget enforcement, and easy model portability.
The hub now processes trillions of tokens weekly, letting engineers focus on building AI apps without infrastructure complexity.
via Vedang Vatsa
This paper maps the entire AI agent infrastructure stack -> 8-layers from compute to security and tracks $14B funding:
Most of the money is going to 2 layers. Bigger platforms are starting to absorb the other 6.
A thread on what I found.
Below is my personal website which aggregates links to many of my socials as well as the various content and community that I curate. Feel free to share this link to others who you think may find this content/community useful to them
The cover image of this newsletter via generated via the ChatGPT Image 2 model within the Krea tool via the following prompt
a painting of a woman holding a red umbrella, in the style of fluid dynamic brushwork, textured canvas, graceful balance, dark orange and white, whirly, i can't believe how beautiful this is, energetic street artist --ar 73:105 --stylize 750 --v 6