This Week in All Things AI - Week 23-2026
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 May 31 to June 6, 2026, the AI landscape was defined by a wave of model releases, escalating cost concerns, and infrastructure advances. Major model launches came from MiniMax (M3), NVIDIA (Nemotron 3 Ultra), Alibaba (Qwen3.7-Plus), Google DeepMind (Gemma 4 12B), and Microsoft (seven MAI models), while Anthropic reported Claude now authors 80% of its own codebase and OpenAI updated ChatGPT's memory system. On the cost front, an unnamed firm's $500M Claude AI bill dominated headlines, spurring broader cost-control moves — Uber capped AI spending at $1,500 per employee monthly, and tools like Factory Router and Lindy's migration to DeepSeek v4 demonstrated significant savings. Infrastructure and tooling saw updates from Perplexity's "Search as Code," SambaNova's disaggregated inference, Supabase's $500M raise, OpenAI's Codex plugins, Grok Composer 2.5, and Scaledown.ai launch of small-language-model services for compression and summarization.
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.
Sunday 31st May 2026
Unnamed firm racks up $500M Claude AI bill in one month
An unnamed enterprise client spent more than $500 million in just 30 days on Anthropic's Claude AI platform after failing to set any usage limits or spending caps for employees, according to an Axios investigation published on May 28. The disclosure, made by an AI consultant whose client generated the bill, may represent one of the costliest IT governance failures on record.
[When AI costs spiral: A company accidentally spent $500 million in one month on Claude AI- what went wrong? | Company Business News \ \ An enterprise client spent $500 million in a single month on Claude AI after failing to set employee usage limits, exposing a growing crisis in corporate AI cost governance.\ \ https://www.livemint.com\ \
Xiaomi just published a deep dive into the end-to-end inference engineering behind MiMo-V2.5 and MiMo-V2.5-Pro — and it’s a strong example of how architecture + systems work together to unlock real production gains.
Inference Optimizations Behind the MiMo-V2.5 Series API Price Reductions
Read the full technical blog: mimo.xiaomi.com/blog/mimo-v2-5…
The V2.5 model family, including MiMo-V2.5 and MiMo-V2.5-Pro, is built on a Hybrid Sliding Window Attention (Hybrid SWA) architecture.
Monday 1st June 2026
MiniMax Launches M3 AI Model for Coding and Agents
Chinese AI lab MiniMax released its M3 model excelling in coding, agentic reasoning, tool use, multimodal chat, and long-context tasks. OpenCode integrated it instantly for free early access, offering 1,400 requests every five hours via terminal, IDE, or desktop app on macOS, Windows, and Linux. It posts strong benchmark results including 59.0% on SWE-Bench Pro, 66.0% on Terminal Bench 2.1, 74.2% on MCP Atlas, and competitive scores against models like GPT 5.5 and Gemini 3.1 Pro across agentic and tool-use tasks. Weights and technical report expected in about 10 days; API is immediately available with 50% off standard usage for the first 7 days to encourage early testing.
Introducing MiniMax M3: The First Open-Weights Model to Combine Three Frontier Capabilities
- Coding & Agentic Frontier: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas
- MiniMax Sparse Attention scales context to 1M
NVIDIA Unveils Nemotron 3 Ultra
The model tops U.S. open-weights rankings with an Intelligence Index score of 48, beating rivals like Gemma 4 31B while delivering over 300 output tokens per second in tests. It shines in agentic AI tasks, boasting 91% agent productivity, strong instruction following, and top coding skills, with 5x faster inference and 30% lower costs than leading competitors. Huang highlighted its openness for developers building everything from search tools to molecular simulations, paired with new agent deployment tools.
NVIDIA just announced the release of Nemotron 3 Ultra in Jensen Huang's Computex keynote: at 550B parameters (55B active), this is the largest Nemotron 3 model to date, and it is the most intelligent US open weights model.
Tuesday 2nd June 2026
SuperGrok and X Premium+ users now can use Composer 2.5 model from Cursor via Grok Build.
Qwen3.7-Plus is Alibaba's new multimodal agent model unifying vision and language for hybrid GUI/CLI operations, visual perception, reasoning, grounding, coding, and search-augmented tasks.
Benchmarks show it delivers competitive results across text, multimodal reasoning, visual understanding, agentic coding, and real-world QA, often matching or exceeding models like Claude Opus-4.6, GPT-5.4, and Gemini-3.1-Pro. Released via API on Alibaba Cloud Model Studio with demos of browser and interactive agents.
Wednesday 3rd June 2026
Uber says it has limited all employees to $1,500 in monthly token spending per AI coding tool “to responsibly encourage agentic AI adoption.”
Microsoft Launches MAI AI Models at Build 2026
The lineup covers reasoning, coding, images, voice, and transcription, all built as a multimodal system optimized for Microsoft's efficient MAIA 200 chips. MAI-Thinking-1, a 35 billion active-parameter model trained on 30 trillion tokens, scores 97% on math benchmark AIME 2025 and tops tough coding tests like SWE-Bench Pro, even beating models from Anthropic and matching Claude in blind evaluation.
OpenAI unveils new Codex plugins for tasks related to public equity investment, banking and sales, and other roles, and plans to integrate Codex into ChatGPT.
Factory Router automatically selects the best AI model for each task in the Droid coding agent, slashing costs by 20-25% while matching 99% of premium performance on benchmarks like Terminal-Bench 2.
Thursday 4th June 2026
Google DeepMind releases Gemma 4 12B, a lightweight 12B-parameter multimodal model under Apache 2.0 license designed to run locally on laptops with just 16GB VRAM or unified memory. It introduces an encoder-free unified architecture where vision uses a tiny 35M-parameter embedding module and audio projects raw signals directly into the LLM backbone.
Friday 5th June 2026
At Computex, SambaNova demoed a “disaggregated inference” architecture where GPUs handle prefill and their RDUs handle decode, yielding faster, cheaper long-horizon agent workloads than GPU-only setups, and they now have this running in production-style environments with partners.
OpenAI updates ChatGPT memory with a “more capable and compute-efficient” architecture and a summary page that lets users review and steer what it remembers.
Supabase, which provides backend tools for building AI apps, raised a $500M Series F led by GIC at a $10B pre-money valuation, up from $5B in October 2025.
Saturday 6th June 2026
Came across Scaledown which has developed purpose-built models for compression, summarization, extraction, and classification. Pricing is 0.05/M tokens with self-deployment options also available.