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.
Major developments this week included xAI completing training on Grok V9-Medium (1.5T parameters) with a public release expected in 2–3 weeks, alongside plans to open-source its current v8-small model by year-end. OpenRouter raised $113M at a $1.3B valuation as Xiaomi's MiMo slashed API pricing by up to 99% and MiniMax teased its M3 sparse-attention architecture. Anthropic released Opus 4.8, Stanford launched the open-source OpenJarvis agent framework, and Cursor announced its inaugural Compile conference. China's Shanghai Futures Exchange began researching AI token futures as Tomasz Tunguz argued that competitive advantage now lies in the orchestration layer rather than foundation models. The week closed with Kirkland & Ellis committing $500M to build a proprietary in-house AI platform and ElevenLabs launching Dubbing v2, a model that preserves speaker emotion, tone, and pacing across 90+ languages.
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.
Elon Musk announced that training of Grok foundation model V9-Medium with 1.5 trillion parameters is complete, with positive evaluation results.
Supplementary training included substantial Cursor data, with fine-tuning now underway and reinforcement learning scheduled to start in a few days.
Public release is expected in 2-3 weeks, marking a major upgrade over the current 0.5T v8-small model, particularly for challenging coding tasks.
He also announces that xAI will open-source its current 0.5T Grok model (v8-small) by the end of 2026, stating it should remain quite useful even after newer models launch.
OpenRouter raised $113M led by CapitalG, a source says at a $1.3B valuation, and now processes 25T tokens across 400+ models weekly, up from 5T six months ago
The future of AI is neurodiversity:
- Agents choosing the most cost-effective model/provider/tool for the task
- Agents orchestrating multiple models for the smartest result
- Advanced security and
Over the last 6 months, weekly volume on OpenRouter grew from 5T to 25T tokens as AI rapidly shifts from experimentation into production.
We’re excited for what comes next.
https://www.nytimes.com/2026/05/26/business/dealbook/openrouter-ai-models-fundraising.html
Xiaomi MiMo announced permanent API price cuts of up to 99% for MiMo-V2.5-Pro and MiMo-V2.5 models, with unified pricing across context lengths and new low rates like $0.0036 per 1M input cache hit tokens for Pro.
Token plans received major upgrades delivering 5–8× more credits at the same price points, with all existing subscriber credits fully reset as a thank-you, while MiMo-V2.5-TTS stays free temporarily.
Better inference efficiency, lower costs, broader access.MiMo-V2.5 Series API pricing is now permanently reduced — by up to 99% compared to previous pricing.
Unified pricing across all context lengths.MiMo Token Plans have also been upgraded:
• 5–8× more usable tokens
Xiaomi MiMo Api Open Platform - Token Plan Global Launch
One-time purchase unlocks both MiMo-V2.5 flagship models, plus TTS model free across all tiers for a limited time. Unleash powerful productivity with Xiaomi MiMo
Skyler Miao, Head of Engineering at MiniMax_AI, posted a teaser for the upcoming M3 model showing their new sparse attention architecture. The diagram details a two-stage GQA-based system: an Index Branch quickly scans and selects top-k relevant token blocks via block max pooling, then a Sparse Branch performs full attention only on those blocks.
This design delivers 9.7x faster prefill and 15.6x faster decoding at 1M token context versus M2, enabling efficient long-context AI without prohibitive compute costs.
main changes:
- based on GQA not MLA
- block level selection like in CSA but attention is done on the real KV, not in the compressed dimension
most ai models work like this,
every word looks at every other word to understand context
great for accuracy, terrible for speed at massive scale
MiniMax's sparse attention changes that
instead of processing an entire 1m token context deeply, the model
Cursor Announces Invite-Only Compile Conference in San Francisco
The one-day Compile event happens June 16 at Fort Mason, bringing together engineers, researchers, and builders for discussions on AI-native development. Speakers like Cursor's Michael Truell and Ryo Lu, plus guests from Every, Shopify, and indie makers, will work through ideas live on a chalkboard stage. It's waitlist-only now—sign up at cursor.com/compile and invitations go out via email—with a call for papers still open on rethinking systems and simplifying complex ideas.
Compile is a one-day event that brings together engineers, researchers, designers, and builders of all kinds to discuss the future of software.
Cursor · Compile
Compile is Cursor's inaugural conference — bringing together developers, researchers, and teams shaping the future of AI-native development.
Tomasz Tunguz of Theory Ventures with a post where in he articulates that in the AI era, the core differentiator in software is not the model itself but the “harness” layer that tames a general LLM into a reliable, domain-specific agent by combining seven capabilities (context, tools, orchestration, state, sandboxing, observability, and cost optimization).
The article argues that when everyone can access similar foundation models, competitive advantage shifts to whoever builds the best harness around the model. This harness “domesticates” a powerful but wild LLM into a dependable system that can safely execute real workflows in specific industries.
Software After AI
Software is no longer about UX & data. It is about the harness, the layer that turns an LLM into a reliable agent. Seven components define the new stack.
In response Brent Maxwell writes
In my engineering teams, nobody can tell the difference between gpt-5.5, claude-4.7, gemini-3.5-flash or composer-2.5.
There is no winner right now in the model wars.
The harnesses are totally the right thing to focus on - they are making or breaking development processes for us. When we figure out how to manate the harness better, we ship faster with less time stuck in PR. When one of our devs just uses the default agent config in the IDE, it really doesn't give them the same amount of leverage.
Shanghai Futures Exchange Explores AI Token Contracts
China’s Shanghai Futures Exchange is in early-stage research on futures contracts tied to AI tokens, the smallest unit of information processed and billed for by AI models. These contracts would let companies hedge against volatile AI compute costs along the AI supply chain, similar in spirit to commodity or energy futures.
While U.S. exchanges like CME and ICE are moving toward futures tied to GPU rental/compute capacity, China’s concept would be directly linked to AI token consumption used for pricing AI services. This represents a different abstraction layer: the U.S. focuses on hardware capacity, whereas China targets the usage-based “digital fuel” that powers AI models.
China’s daily AI token usage has exploded roughly 1,000x since early 2024, reaching more than 140 trillion tokens a day by March 2026, underscoring surging demand and cost exposure. Token-based derivatives are being framed as a potential new asset class, with figures like BlackRock’s CEO noting that futures on compute could become a distinct financial market.
Exclusive: China works on AI token futures market, sources say, in race with US
China is designing a futures market for AI tokens, sources familiar with the matter said, as the country potentially takes a different tack to U.S. exchanges developing compute power futures to tap the rapidly growing appetite to hedge AI costs.
Some links about Anthropic's Opus 4.8 launch including a tweet thread from Anthropic with some guidance on devs as they migrate to Opus 4.8 and some commentary from Dan Shipper of Every who had access to Opus 4.8 for the past two weeks
A few things worth knowing:
Anthropic just dropped Opus 4.8—and it is a MONSTER
We've been testing for about a week @every and our verdict is they could've just called it Opus 5, it's that good.
Here's our vibe check:
- Beats GPT-5.5 on Senior Engineer bench. On our toughest benchmark Opus
Vibe Check: Opus 4.8-Anthropic Should've Rounded Up to 5
Opus 4.8 tops both our Senior Engineer benchmark and our writing tests. It's the most complete model we've tested. We just wish it had an app to match.
Stanford Unveils OpenJarvis: Efficient AI Agents for Your Devices
Stanford researchers from Hazy Research and Scaling Intelligence Lab released OpenJarvis v1.0, an open-source framework that builds personal AI agents to run locally on devices. It emphasizes 'Intelligence per Watt' with swappable components like local models, engines, agents, tools for apps like Slack and email, and self-improvement features. Users get eight ready-to-run agents for tasks such as morning briefings and code review, installable via a simple one-line script across CLI, web, desktop, or messaging apps. Benchmarks show it handles most queries at interactive speeds with far lower costs and latency than cloud systems, positioning it as a privacy-focused alternative to massive data centers.
We believe the future is in the opposite direction: on-device inference, smaller models, watts instead of gigawatts.
Today we're releasing @OpenJarvisAI v1.0: a personal AI
@OpenJarvisAI fixes the harness, not the model. Result: 77% of the accuracy gap recovered, 800x lower cost per query, 4x lower latency.
Built on Lambda. Open-sourced from Stanford.
For two years, they've been making the case that AI inference belongs on hardware people already own, not just in megawatt data centers. Excited to support the
We believe the future is in the opposite direction: on-device inference, smaller models, watts instead of gigawatts.
Today we're releasing @OpenJarvisAI v1.0: a personal AI
via shrwn
Free Google x Kaggle vibe coding course
* Mid June
* 1-2hrs/day for 5 days
* Free
5-Day AI Agents: Intensive Vibe Coding Course With Google
June 15 - 19, 2026
Kirkland & Ellis LLP is one of the world’s most successful and profitable “Big Law” firms with approximately 4,000 attorneys across 23 offices. It consistently ranks at or near the top of revenue and profitability metrics.
Kirkland & Ellis is committing $500 million over the next 3–4 years to build its own proprietary AI platform and custom tools. This is one of the largest and most ambitious technology investments ever announced by a law firm.
Key details include:
Funding & Timeline: Funded entirely from the firm’s own revenue (aligning with Chair Jon Ballis’s philosophy of investing roughly 1% of revenue in new initiatives — which would be ~$100M+ in a $10B+ year). More than $100 million is expected in 2026 alone, with the balance spread over the following years.
Goals: Create a broad, firm-wide AI platform that captures and deploys the firm’s “collective intelligence” to support lawyers across practices “start to finish” on client work. The aim is to move beyond reliance on multiple third-party tools toward more integrated, customized capabilities.
Development Approach: Informed by input from 250 lawyers (including 100 partners) on real workflows and needs. A team of 180 tech professionals is involved, working with undisclosed external partners/companies to help build the technology. Crucially, the resulting tools and IP will not be commercialized or sold to other law firms — a deliberate contrast to some other firm-vendor partnerships (e.g., Freshfields’ work with Anthropic).
Technical Elements: Involves on-premise GPU environments and Microsoft Azure-based AI infrastructure, including facilities for training and inference. The firm is actively hiring for AI-related roles (dozens of positions), including high-compensation roles like AI Infrastructure Directors.
Context Within the Firm: This builds on existing efforts. Kirkland already deploys third-party legal AI tools such as Harvey across its attorneys. It has a history of building proprietary technology in-house, including SideTrack (a tool for investment fund work, particularly around MFN issues) and earlier databases like CTRAN for M&A competitive intelligence. It also maintains internal innovation teams, AI Innovation Advisors embedded in practice groups, and responsible AI governance structures.
This move sits at the center of a key debate in legal technology: buy vs. build for AI. Many top firms rely heavily on specialized platforms like Harvey, CoCounsel (Thomson Reuters), or Lexis+ AI. Kirkland’s approach prioritizes:
Greater control over data, customization, and roadmap.
Using the firm’s own vast proprietary knowledge and deal experience as a competitive differentiator (rather than everyone having access to similar generic or vendor tools).
Enhanced security, confidentiality, and alignment with internal workflows.
Long-term ownership rather than ongoing licensing dependency.
Kirkland & Ellis has form for building its own technology. The $500m AI play is its biggest yet. - Legal IT Insider
Kirkland & Ellis announced that it would spend $500m over the next three to four years developing its own custom AI tools and services.
ElevenLabs launches Dubbing v2, which it says preserves the original speaker's emotion, tone, and pacing across 90+ languages while staying synced to content
Introducing Dubbing v2: our revolutionary new dubbing model
Introducing Dubbing v2: our revolutionary new dubbing model, which carries the emotion and performance of the original speaker across every language.
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 Krea 2 Large model within the Krea tool via the following prompt
Automobile inspired modern curved building , where the curved building's façade meticulously crafted from gears, cogs, and mechanical parts. The building's façade design embodies the fusion of nature and machinery, as evidenced by the intertwining vines and mechanical components