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.
The week of June 7–13, 2026 saw model releases dominate the AI landscape, with Anthropic introducing Claude Fable 5, Z.ai launching GLM-5.2 with a 1 million-token context window, NexEcosystem announcing the Nex-N2 open-weight agentic family, Moonshot releasing Kimi-K2.7-Code for coding efficiency, and Google upgrading NotebookLM with Gemini 3.5 and agentic capabilities. In infrastructure, Harvey AI and Trajectory Labs post-trained NVIDIA's Nemotron 3 Ultra for legal tasks, and an authentication layer was released to connect agents like Hermes and OpenClaw to commercial APIs.
The week closed with a significant policy disruption as the US government issued an export control directive suspending all non-US access to Fable 5 and Mythos 5, forcing Anthropic to disable the models worldwide. Meanwhile, investor Ray Dalio warned of an AI bubble, OpenAI reportedly considered price cuts to compete with Anthropic, and Cognition's coding benchmark found most frontier models produce largely unusable output.
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
In an interview on Bloomberg Television, Dalio said that all great technological changes create bubbles and that AI was showing the typical signs of this pattern. Investor Ray Dalio says the AI bubble will burst in the next few years unless startups start making billions. Currently, AI companies spend about 8 times more than they earn. The sector needs to generate $2 trillion annually by 2030 to justify current investments.
Dalio compares the AI bubble to past tech bubbles like dotcoms, where stock prices rose on hype but lacked real profits. He acknowledges AI will change the world but doubts its economic viability so far.
My friend Ian Rogers of Ledger just published on Rick Rubin's Tetragrammaton on why AI will be many things, but never human. He argues that while AI will become pervasive, persuasive, and highly “personality-like,” it remains a tool rather than a person, and the real danger is humans forgetting that distinction and granting it moral or legal status.
Google upgrades NotebookLM, which now runs on Gemini 3.5 and Antigravity, to deliver new agentic capabilities and more advanced reasoning for AI Ultra users
Massive upgrades deliver agentic capabilities in chat, more advanced reasoning, and a suite of new output formats. Tackling complex, multi-step research problems has never been easier.
Rolling out now to Google AI Ultra subscribers.
Do better research with NotebookLM
NotebookLM's latest upgrades deliver new agentic capabilities and more advanced reasoning to tackle complex research projects.
Gemini 3.5 and Antigravity come to Google NotebookLM
NotebookLM is getting a big upgrade, but it's only for AI Ultra and enterprise accounts right now.
Cognition, creators of the AI software engineer Devin, launched FrontierCode, a new coding benchmark that evaluates whether AI-generated code would be merged by expert open-source maintainers rather than just passing functional tests.
Tasks were hand-crafted by maintainers from 36 major repositories with over 40 hours invested per task, using unit tests, rubrics, scope checks, and novel verifiers like mutagent to assess quality, style, test effectiveness, and codebase adherence.
Leading models show major gaps on the 50-task Diamond set, with Claude Opus 4.8 topping at 13.4%, GPT-5.5 at 6.3%, and open models lower, while reducing misclassification errors by 81% versus SWE-Bench Pro.
Models write sloppy code that works but isn’t maintainable. Our eval is first to measure: would you actually merge this code?
Cog had IOI
Models write sloppy code that works but isn’t maintainable. Our eval is first to measure: would you actually merge this code?
via Ben
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.
Comment from Anya Shapina on Fable's release
Day 1 using Fable - so far, extraordinary! Quality, speed, rule-following. Use case - building agents on MastaAI (ts). Haven't check token burn :) Anyone else?
Harvey AI partnered with Trajectory Labs to post-train NVIDIA's open-weight Nemotron 3 Ultra on their Legal Agent Benchmark, lifting its all-pass rate from 0% to 5.8% and rubric pass rate to 81%, placing it between Claude Sonnet 4.6 and Opus 4.6.
Post-training shifted held-out task reliability from ~70% to ~95% rubric pass rates within 24 hours of the base model's release, using the same data and platform recipe applied to prior Nemotron versions.
The resulting legal agent matches closed-model performance bands while operating at 1/8th to 1/50th the per-token cost, demonstrating rapid domain specialization of open models for high-value verticals like law.
1) Open-weight models can reach frontier legal performance.
On our Legal Agent Benchmark (LAB), Nemotron 3 Ultra started at a 0% all-pass rate. After post-training, it
via the Wall Street Journal so it may be behind a paywall
Sources: OpenAI is considering drastically lowering its price for tokens in anticipation of similar cuts the company expects at Anthropic
SemiAnalysis bought all consumer plans from Anthropic and OpenAI, then pushed them to weekly limits with intense coding tasks. Their tests showed OpenAI's $200 Pro 20x plan equals up to $14,000 in monthly API value, while Anthropic's $200 Max 20x reaches $8,000—many times the subscription cost. The firm notes these heavy subsidies hook developers, though limits evolve and skeptics question long-term sustainability as companies adjust rates.
via Ming Guang Yong
For folks with hermes and openclaw, I have build an authentication layer (https://mybrandmetrics.com) to various apps that I use like Google Ads, Facebook Ads, Google workspace (calendar, sheets, drive, gmail), Youtube, Instagram, Tiktok (mainly on the social media side) that can connect agents to (openclaw and hermes agent). It's currently free. I have been using it for my own Google Workspace use cases like finding emails, copying content from emails into google sheets, creating calendar invites with attachments. Looking for some folks to test and give feedback. Pls try it out and DM me
This authentication layer enables agents to connect to APIs and data in these apps. I also have the associated skills at https://clawbus.com/skills
NexEcosystem announces Nex-N2, a new open-weight family of agentic models emphasizing unified "Agentic Thinking" that fuses reasoning, coding, search, and tool use into one coherent loop with adaptive depth scaling for efficiency.
The self-reported numbers are spicy
75.3 on Terminal-Bench 2.1
80.8 on SWE-Bench Verified
90.7 on GPQA Diamond
Currently available for free via OpenRouter
Nex-N2 is here!A family of agentic models that doesn't just think, it acts!
Coding, search, tool use. All fused into a single agentic reasoning loop.
- Adaptive Thinking, auto-scales reasoning depth per step. Saves ~20% tokens, zero performance loss.
- Coherent Thinking, one
modelscope.ai/models/nex-agi…
modelscope.ai/models/nex-agi…● Models: Nex-N2-Pro 397B total, 17B active; Nex-N2-mini 35B total, 3B active
●
Couple of links about AI created micro-dramas which are picking up in China and may become popular in the US.
Roope Rainisto , Creative AI lead at Supercell who also lurks in this group has created a product named Beat Bandit which generates scripts for different formats including micro-dramas [site has some rendering issues currently which I've let Roope know]
https://x.com/ollieforsyth/status/2058773143091896679
Serialized, vertical, and made for the phone. Episodes run one to two minutes, seasons run 100 to 150.
In China it's already obvious. More viewing time now goes to short dramas
Followup message from Roope himself
Yeah, working on BeatBandit - you are the showrunner and that's your writers' room - use the team a little, use the team a lot.
I've been running it through Claude Fable for a few days now through the MCP - pretty brutal workflow to allow Claude to read the code, simulate the user, evaluate the result, pinpoint things in the code that are not optimal - rinse and repeat. Fable is smarter than me, tough pill to swallow but it is what it is. :)
Kimi-Moonshot announces the open-source release of Kimi-K2.7-Code, a specialized coding model with strong gains over K2.6: +21.8% on Kimi Code Bench v2, +11% on Program Bench, +31.5% on MLS Bench Lite, plus 30% fewer reasoning tokens and better long-horizon task success.
Benchmark charts show K2.7-Code outperforming K2.6 while closing the gap or exceeding GPT-5.5 and Opus 4.8 (xhigh) on coding and agent metrics, with a performance-vs-tokens plot highlighting reduced overthinking and higher efficiency.
The model is immediately available via Kimi API and Kimi Code, with 6x high-speed mode incoming and a new beta program for early testing of upcoming features
Kimi-K2.7-Code, our latest coding model, is now released and open-sourced!
Improved coding & agent performance over K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite. Reasoning efficiency: Less overthinking, with 30% lower
Anthropic reveals a US government export control directive citing national security has forced the suspension of Fable 5 and Mythos 5 access for all foreign nationals, including their own foreign employees.
To comply, the company must immediately disable these models for every customer worldwide, while access to other Claude models remains fully operational.
Anthropic calls the order a misunderstanding, apologizes for the disruption, and states it is working urgently to restore service as soon as possible.
The net effect of
Statement on the US government directive to suspend access to Fable 5 and Mythos 5
The US government has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States.
Z.ai released its flagship GLM-5.2 AI model to paid Coding Plan users, highlighting its strengths in coding tasks with a 1 million-token context window and adjustable thinking levels for complex workflows. Full API access and open-sourcing under the MIT license arrive next week
GLM-5.2 is now available to all GLM Coding Plan users, including Lite, Pro, Max, and Team plans.
docs.z.ai/devpack/latest…
As our new flagship model, GLM-5.2 delivers
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 Nano Banana Pro model within the Krea tool via the following prompt
The Great Wave : "Blossoming Cataclysm" Photorealistic : State-of-the-art photorealistic , leveraging advanced AI imaging capabilities. Landscape : Supernova erupting into a colossal red flower , marrying intense beauty with awe-inspiring realism. Message: Artistic convergence of cosmic power and giant blooms , evoking a vivid and visceral explosion. Format: A visually arresting masterpiece , showcasing the zenith of modern imaging techniques, ideal for gallery and prestigious exhibitions.