The day's most important AI news: breakthroughs, releases, funding, and policy — curated for developers, founders, and investors.
Today's Stories
Claude Fable 5 launches as first generally available 'Mythos-class' model, with vetted-only Mythos 5
Vendor: Anthropic
Anthropic released Claude Fable 5, the public version of its highly anticipated Mythos model, featuring a 1M-token context window, up to 128k output tokens, and state-of-the-art performance in software engineering, vision, and scientific research. Pricing is $10 per million input tokens and $50 per million output tokens. The unrestricted Claude Mythos 5 variant — the same underlying model with cyber safeguards removed — ships only to vetted cyber defenders through Project Glasswing, while consumer queries on sensitive cyber/bio topics automatically reroute to the less capable Opus 4.8.
AWS Graviton5 goes GA with EC2 M9g/M9gd instances, claiming 25% gains for agentic workloads
Vendor: AWS
AWS announced general availability of its fifth-generation Graviton5 processors via new general-purpose EC2 M9g and M9gd instances, calling Graviton5 its most powerful and energy-efficient chip with up to 25% better compute performance than Graviton4. The processors are purpose-built for the complex, long-running tasks of the agentic AI era.
Google upgrades NotebookLM with Gemini 3.5 and Antigravity, cuts AI Plus price, then hits June 10 outage
Vendor: Google
Google upgraded NotebookLM with its Gemini 3.5 model and the Antigravity capability, adding a cloud computer and source-finding tools, and cut the price of its AI Plus plan while doubling included cloud storage. On June 10, Gemini experienced a major outage with users scrambling for workarounds.
Apache Spark 4.0 reaches general availability on Amazon EMR
Vendor: AWS
AWS announced general availability of Spark 4.0 across EMR Serverless, EMR on EC2 and EMR on EKS, adding Spark Connect, the Variant data type, SQL scripting, Python API improvements and streaming enhancements. SageMaker Unified Studio Notebooks also gained EMR Serverless support.
Mistral ships Medium 3.5, Voxtral TTS open-weights, and remote coding agents
Vendor: Mistral
Mistral released Medium 3.5, a 128B dense flagship unifying instruction-following, reasoning and coding optimized for self-hosting on as few as four GPUs, alongside Voxtral TTS — a frontier open-weights text-to-speech model — plus remote coding agents in Vibe and a Le Chat Work mode for multi-step tasks.
AI routing startups raise big rounds as token-cost backlash grows
Vendor: Other
AI-routing companies that direct tasks to the most cost-effective models are seeing a funding surge: OpenRouter raised $113M at a $1.3B valuation, and Concentrate AI emerged from stealth with $5M+. Coinbase's CEO and Hugging Face's Julien Chaumond endorsed routing as a key cost lever as cheaper models like DeepSeek surge.
Cohere launches North Mini Code, its first developer model, on Hugging Face
Vendor: Other
Cohere released North Mini Code, its first model designed specifically for developers — a 30B-parameter Mixture-of-Experts model with 3B active parameters, available on Hugging Face under an Apache 2.0 license. It targets agentic software engineering and complex code generation, positioning among top open-source coding models in its class.
DiffusionGemma delivers 4x faster text generation, optimized for NVIDIA GPUs
Vendor: Google
Google DeepMind introduced DiffusionGemma, a diffusion-based open text-generation model that is up to 4x faster than token-by-token approaches by processing up to 256 tokens per step. NVIDIA optimized it across GeForce RTX, RTX PRO, and DGX Spark systems, with NIM microservices streamlining deployment from development to production.
OpenAI plans agent-focused ChatGPT redesign, declaring 'chat is dead'
Vendor: OpenAI
OpenAI is reportedly rebuilding ChatGPT around AI agents and coding tools rather than chat, driven by an internal view that 'chat is dead,' per a Financial Times report. The 'superapp' is expected to integrate Codex and AI agents, launching first on web and mobile to boost competitiveness with Anthropic and grow revenue.
Nvidia acquires Kumo AI, pushes deeper into software amid D-Matrix challenge
Vendor: NVIDIA
Nvidia acquired Kumo AI to bring predictive AI to enterprise business data, continuing a pattern of software acquisitions including Run:ai and Illumex. Meanwhile, Microsoft-backed upstart D-Matrix is challenging Nvidia's inference dominance, and Nvidia is doubling down on CPUs and AI PCs to compete with Apple silicon on memory bandwidth.
Grok V9 rolls into Tesla cars and X, leveraging Musk's distribution flywheel
Vendor: xAI
xAI has begun integrating Grok V9-Medium, its largest model yet at 1.5 trillion parameters, into Tesla's connected-car fleet and the X social network. The deployment leverages Elon Musk's distribution advantage to reach hundreds of millions of X accounts and millions of internet-connected Teslas. Grok functions as an in-car voice and navigation assistant but does not control Tesla's self-driving system.
xAI brings Colossus 2 online — ~1-million-GPU supercluster now pre-training Grok 5
Vendor: xAI
xAI announced the operational launch of Colossus 2, featuring roughly 1,020,000 NVIDIA GB300 GPUs across Memphis and Atlanta with 1.4GW total power and 140kW per-rack liquid cooling. Grok 5 is now pre-training on the cluster with completion expected by August 2026, reshaping the training-compute dynamics of the industry.
NVIDIA releases Cosmos 3 open omni-model for physical AI at GTC 2026
Vendor: NVIDIA
NVIDIA released Cosmos 3, billed as the first open omni-model for physical AI reasoning and action across video, robotics and industrial applications. Jensen Huang showcased it at GTC 2026 alongside Adobe, Cohere, Google DeepMind, Meta, Microsoft, OpenAI and Tesla, and it is hosted on Hugging Face.
Meta will use off-site activity to personalize feeds and AI responses
Vendor: Meta
Meta announced it will integrate data from ad partners — including website activity collected via Meta Pixel — to personalize both user feeds and responses from its AI chatbot, framing it as a step toward 'personal superintelligence.' Meta says it isn't collecting new data and that users retain privacy controls.
Meta pivots from open-source Llama to closed 'Avocado' frontier models
Vendor: Meta
Meta is preparing new foundational models code-named 'Avocado,' marking a shift from open-source Llama toward proprietary frontier models released first in closed form. The effort is led by Scale AI CEO Alexandr Wang following Meta's $14B Scale AI investment.
Meta partners with Reliance on first AI-enabled data center in India
Vendor: Meta
Meta and Reliance Industries announced an expanded strategic partnership under which Meta will lease its first AI-enabled data center in India, deepening Meta's AI infrastructure footprint in one of its largest user markets.
Apple debuts 'Siri AI' and Core AI on-device framework at WWDC, powered partly by Google
Vendor: Apple
At WWDC 2026, Apple unveiled 'Siri AI,' an entirely new, more conversational and personal assistant delivered as a dedicated app and powered in part by more capable Google-built models. Apple also introduced Core AI, an on-device framework for running full-scale LLMs optimized for Apple silicon, plus Xcode 27 agentic coding integrating Anthropic's Claude Agent and OpenAI's Codex. Features are in developer testing now, with public beta later this year.
DeepSeek reportedly completes major training run on Huawei chips, bypassing Nvidia
Vendor: DeepSeek
Chinese researchers, working with Huawei and Shenzhen institutions, reportedly completed full-parameter post-training of DeepSeek's ~1.6-trillion-parameter V4-Pro model using over 1,000 Huawei Ascend 910C chips, without any Nvidia hardware. The milestone signals that domestically produced accelerators can support intensive training-class AI workloads, a notable advance for China's AI hardware independence.
OpenAI confidentially files draft S-1, targeting $1T+ IPO valuation
Vendor: OpenAI
OpenAI submitted confidential draft S-1 registration documents to the SEC, formally beginning its path to public markets and joining Anthropic, which filed on June 1. The valuation is expected to exceed $1 trillion, and OpenAI expects to go public within the next year as frontier labs race to capitalize on surging investor interest.
Alibaba's cloud division begins AI-driven 'quiet' layoffs as Beijing pushes adoption
Vendor: Alibaba
An engineer at Alibaba's cloud division said AI-driven headcount reductions have begun, unfolding through gradual cuts and attrition rather than a single mass round. The trend reflects broader 'quiet' layoffs across Chinese tech firms as Beijing promotes AI adoption while avoiding visible job losses that threaten social stability.
Microsoft restricts Claude Fable 5 internally over data-retention terms, even as it ships it to customers
Vendor: Azure
Microsoft is limiting employee use of Anthropic's newly released Claude Fable 5 over Anthropic's data-retention requirements, even as it quickly added the model to GitHub Copilot and the Microsoft Foundry catalog for customers. The move highlights enterprise tension between adopting and trusting third-party frontier models.
Guided Learning in Gemini boosts Sierra Leone students' math scores by 1.2-1.7 years
Vendor: Google
A randomized controlled trial by Google DeepMind, Fab AI, and Sierra Leone's Ministry of Education found that Guided Learning in Gemini produced math gains equivalent to 1.2-1.7 years of typical learning within eight weeks across 1,763 junior secondary students. The Socratic pedagogical design emphasized building conceptual understanding over giving direct answers.
Meta's Muse Spark AI model replaces Llama 4 on its smart glasses
Vendor: Meta
Meta's Muse Spark model now powers Meta AI on most of its smart glasses, a significant upgrade over Llama 4 that narrows the gap to leading AI systems. Announced in April, Muse Spark is the first publicly released model from Meta Superintelligence Labs.
AWS launches Neuron Agentic Development to automate Trainium kernel tuning
Vendor: AWS
AWS announced Neuron Agentic Development, a collection of AI agents and skills for developers building on AWS Trainium and Inferentia, aiming to replace manual hand-tuning of kernels with an agent-driven workflow. The release deepens AWS's effort to make its custom AI silicon more accessible to developers.
Samsung reverses 2023 ban, rolls out ChatGPT, Gemini and Claude companywide
Vendor: Samsung
Samsung Electronics is integrating external generative AI services — ChatGPT, Gemini, and Claude — across its Device eXperience division and other affiliates this month, reversing a 2023 ban prompted by data-leakage concerns. The rollout follows a two-month validation with 2,500 employees, and Samsung will continue developing its in-house Samsung Gauss model under a security control layer.
Amazon secures $17.5 billion loan facility to fund AI capex ramp
Vendor: AWS
Amazon secured a $17.5 billion delayed-draw term loan facility, letting it withdraw funds as needed amid an AI-driven capital-expenditure ramp. Earlier in the week Amazon also filed for a five-part debt offering in Canada worth up to C$14 billion.
GPT-5.4 and GPT-5.5 now available on Amazon Bedrock in US East
Vendor: OpenAI
AWS expanded availability of OpenAI's GPT-5.4 and GPT-5.5 models in the US East (N. Virginia) Region on Amazon Bedrock, targeting reasoning, coding, computer use, document workflows, and long-running agentic tasks. GPT-5.5 is described as OpenAI's most capable model for advanced coding and research.