AI Industry Daily News
A roundup of the AI industry's day, centered on Codex Windows support, grok-build-0.1, Claude Opus 4.8, Command A+, and Rosalind Biodefense.
Today's highlights
Databricks Unveils Genie Ontology, Reporting 84.5% First-Pass Accuracy on Data Tasks
At Data+AI Summit 2026, Databricks introduced Genie Ontology, a layer that automatically builds a knowledge graph from internal data, citing 84.5% first-attempt accuracy on enterprise data analysis tasks.
Key topics and reactions
Anthropic Keeps Security Model Mythos Private After It Finds 271 Vulnerabilities
Anthropic reported that its most advanced model, Mythos, identified and was able to exploit 271 vulnerabilities during testing, with about 99% still unpatched at the time. Rather than ship the model broadly, the company restricted access to a defensive-only program called Project Glasswing.
The decision has drawn attention as an example of scoping a release based on a model's capabilities, with Anthropic citing safety considerations as the reason for limiting availability.
The move fits a broader pattern at Anthropic, which also released a new safety-focused development SDK, of deliberately narrowing distribution for high-capability systems.
Meta Caps Employee AI Token Use After Internal Costs Approach Billions
According to an internal memo, Meta is setting limits on AI token usage for roughly 6,000 employees and encouraging a shift from external tools to in-house alternatives such as MetaMate. The company had previously embedded "AI-driven impact" into performance reviews, but reversed course as internal usage costs were projected to reach billions of dollars in 2026.
The change follows a surge in usage and a "tokenmaxxing" culture, including an employee-built leaderboard called "Claudeonomics" that ranked top users with titles like "Token Legend." Tracking suggested employees consumed roughly 60 trillion to 73.7 trillion tokens over about 30 days, which various estimates valued anywhere from about $900 million to $9 billion at public prices.
Meta is building an internal platform called "AI Gateway" to track usage and spending in real time, with anomaly alerts and team-level visibility, and plans formal team and individual budgets from 2027. Uber, Amazon and Walmart have made similar moves, with Uber reportedly setting a $1,500 monthly cap per tool.
Anthropic Resets Claude Usage Limits Across All Plans for the Weekend
Anthropic reset the 5-hour and weekly usage caps for Claude across every plan, opening additional usage allowance for the weekend.
The change was welcomed by developers and other heavy users, for whom usage limits had been a frequent constraint.
Google, Amazon, Meta, Microsoft and OpenAI Developing Custom ASICs With Broadcom
Google, Amazon, Meta, Microsoft and OpenAI are reported to be developing their own ASICs in parallel, with Broadcom playing a leading role, as part of an effort to reduce dependence on NVIDIA GPUs.
The custom chips are focused on inference, where the companies expect improved cost efficiency. Optimizing inference cost has become a central front in infrastructure competition.
Category highlights
Tesla Files 'MEGAPOD' Trademark for AI Compute Data Center Hardware
Tesla filed a standard-character trademark for "MEGAPOD" with the USPTO on June 18, 2026 (application 99893717, status LIVE/PENDING), describing modular, self-contained data center hardware for AI computing. The filing covers servers, AI processing hardware, networking gear, power distribution units and cooling, plus downloadable management software, but lists no benchmarks, prices, capacity or release timing. The move extends Tesla's energy work with utility-scale Megapack batteries, used at sites like xAI's Colossus and Meta's Wyoming facility, into compute hardware itself.
Nous Research's Hermes Agent Tops OpenClaw in New GitHub Contributors
Nous Research's autonomous Hermes Agent surpassed the earlier open-source agent OpenClaw in new GitHub contributors over the past 30 days, according to reports. OpenClaw, an MIT-licensed personal agent that acts as a gateway to messaging apps and delegates actions to models like Claude, GPT, Gemini and Ollama, reached over 100,000 GitHub stars and as many as 370,000-plus. Hermes Agent, released in late February 2026, features a self-improvement loop that generates reusable skill files, supports a migration command from OpenClaw, and reportedly passed 110,000 stars in 10 weeks with more than 142 contributors by June 2026.
Anthropic SDK, MistralAI Funding and xAI Robotics Model Headline Model News
Beyond Mythos, Anthropic released a new safety-focused development SDK, while MistralAI completed a large European funding round, intensifying competition. xAI announced a new model aimed at physical-world understanding and robotics.
NVIDIA, Genesis, Sony AI and DeepMind Announce Robotics and Protein Tools
NVIDIA unveiled MotionBricks for real-time physics simulation, Genesis introduced the autonomous robot ENO, and Sony AI announced an agent called ACE. Anthropic rolled out Claude Code Artifacts, and DeepMind published a new protein-folding method extending AlphaFold.
Tesla Reports FSD Supervised Crash and Braking Figures on Dutch Roads
Tesla reported that FSD Supervised on public roads in the Netherlands had 3.5x fewer collisions, 14.9x fewer emergency brakes and 8x fewer hard steering events than manual driving, with v14.3.4 cited for precise defensive maneuvers. Critics questioned comparing US national averages with Tesla's own data, noted continued ghost braking and thousands of ongoing NHTSA complaints, and pointed to insufficient independent verification of self-reported figures.
Key trends
PixVerse Adds Captain Tsubasa Template as Kling Teases Full Music Video
PixVerse added an official Captain Tsubasa template to its web version, letting users generate anime soccer clips by uploading a photo and selecting a special move. Kling AI released a music video teaser and announced a full MV three days later, while open-weight models like LTX-2 and one-prompt agents like N2 widen the field at both the professional and consumer ends.
Data+AI Summit, ICML, Cannes Lions and YC Hackathon Run Concurrently
Data+AI Summit 2026 recordings are available on demand, ICML 2026 runs July 6-11 at COEX in Seoul with Luma exhibiting, CapCut is hosting a panel at Cannes Lions, and a Y Combinator hackathon is underway with MiniMax noted. The clustering of events has concentrated research, recruiting and marketing announcements.
Signal President Warns AI Agents Can Bypass Encryption Protections
Signal's president warned that AI agents can circumvent the protections provided by encryption, raising questions about how agentic systems interact with secure messaging.