AI Daily Brief

Friday, May 8, 2026

Top stories from Anthropic, OpenAI, Google, xAI, Microsoft & more

Today's Stories

  1. Anthropic Unveils Claude Mythos: Cybersecurity AI Too Powerful to Release Publicly
  2. Anthropic Surpasses OpenAI in Revenue, Eyes $900B Valuation
  3. Claude Managed Agents Gets Dreaming, Multiagent Orchestration & Security Beta
  4. OpenAI Upgrades ChatGPT Default to GPT-5.5 Instant with 52% Fewer Hallucinations
  5. OpenAI Releases Three New Realtime Voice Models for the API
  6. Microsoft Launches Agent 365: AI Agents Are Now the Operating Layer of Work
  7. Google DeepMind Takes Stake in Eve Online Developer to Train AI on Space MMO
  8. US Government to Test Google, Microsoft & SpaceXAI Models Before Launch
  9. xAI Is Officially Gone — Grok Now Lives Inside SpaceXAI

Anthropic Unveils Claude Mythos: A Cybersecurity AI Too Powerful to Release Publicly

Anthropic

Anthropic has quietly launched Project Glasswing, a controlled-access program for its most advanced model yet — Claude Mythos Preview. The company says Mythos is "far ahead" of any other model in offensive and defensive cybersecurity, and has decided against a public release over concerns it could be weaponized.

Access is currently restricted to a hand-picked set of approved organizations including AWS, Apple, Cisco, Google, JPMorgan Chase, and Microsoft, who are using it to identify and remediate critical software vulnerabilities before they can be exploited. Governments, banks, and utility companies have expressed both interest and concern since the model's capabilities became known.

Why it matters: This is the first major AI lab to publicly acknowledge restricting a frontier model's release on safety grounds — a significant precedent for the industry.

Anthropic Surpasses OpenAI in Revenue, Eyes $900B Valuation in New Round

Industry

In a stunning reversal, Anthropic hit $30 billion in annualized revenue in April 2026 — surpassing OpenAI's $25B ARR. This represents a 30x revenue increase from just January 2025, driven primarily by enterprise Claude adoption and the explosive growth of Claude Code (now at $2.5B ARR alone).

Anthropic is reportedly in talks with investors to raise a new $50 billion round at a $900 billion valuation, which would overtake OpenAI's $852B valuation. Over 1,000 enterprise customers now spend more than $1 million per year on Claude. OpenAI, meanwhile, confirmed $2 billion in monthly revenue alongside its earlier $122B raise.

$30B
Anthropic ARR
$25B
OpenAI ARR
$900B
Anthropic target val.
30×
Anthropic growth (16 mo.)

Claude Managed Agents Gets Dreaming, Multiagent Orchestration & Security Beta

Anthropic

Anthropic shipped three significant upgrades to Claude Managed Agents this week. The first is Dreaming: agents can now review their past sessions to surface patterns and autonomously self-improve between runs — giving them a form of persistent memory across jobs.

The second is Multiagent Orchestration, where a lead agent can break a complex task into subtasks and delegate each to specialist agents with their own models, prompts, and tools running in parallel on a shared filesystem. Third, Claude Security launched in public beta, bringing AI-assisted vulnerability detection to enterprise workflows. Separately, Anthropic also announced Claude Code's five-hour rate limits are being doubled, enabled by a new compute deal with SpaceX's Colossus supercomputer.

OpenAI Upgrades ChatGPT Default to GPT-5.5 Instant — 52% Fewer Hallucinations

OpenAI

OpenAI on May 5 updated ChatGPT's default model to GPT-5.5 Instant, replacing GPT-5.3 Instant. In internal evaluations, the new model produced 52.5% fewer hallucinated claims on high-stakes prompts in medicine, law, and finance — the most significant accuracy improvement OpenAI has shipped in a default model update.

GPT-5.5 Instant also features enhanced personalization that learns from your past chats, uploaded files, and connected Gmail. The personalization features are rolling out first to Plus and Pro users on the web, with Free, Go, Business, and Enterprise tiers to follow. The model is also available in the API as chat-latest.

OpenAI Releases Three New Realtime Voice API Models

OpenAI

Alongside GPT-5.5 Instant, OpenAI shipped a trio of new models purpose-built for real-time voice applications: GPT-Realtime-2 for smarter live voice reasoning, GPT-Realtime-Translate for multilingual speech translation in real time, and GPT-Realtime-Whisper for high-quality streaming transcription.

The releases signal OpenAI's push to build out its API voice stack for enterprise developers building phone agents, live interpreters, and accessibility tools — a category where latency and accuracy are critical competitive factors.

Microsoft Launches Agent 365: AI Agents Are Now the Operating Layer of Work

Microsoft

Microsoft published its 2026 Work Trend Index on May 5 with a striking headline: 78% of knowledge workers now use AI agents at least weekly, up from just 12% in 2024. Alongside the report, Microsoft announced Agent 365 — a new platform that positions AI agents as the next operating layer across Microsoft 365, replacing the traditional app-by-app workflow.

The update also introduced Copilot Cowork Mobile for iOS and Android, new native integrations with Power BI via Fabric IQ, and expanded Dynamics 365 coverage across sales, customer service, and ERP. 67% of business leaders say agent-powered automation has already reshaped at least one core business process at their organization.

Google DeepMind Takes Stake in Eve Online Developer to Train AI on Space MMO

Google

Google DeepMind announced it has taken a minority stake in CCP Games, the Icelandic developer behind the 23-year-old space sandbox MMO Eve Online. The partnership will give DeepMind access to Eve's uniquely complex player-driven economy, diplomacy, and combat data to train next-generation AI agents in long-horizon, multi-agent strategic reasoning.

Eve Online's emergent gameplay — featuring large-scale wars, market manipulation, and elaborate player alliances — has long been considered one of the most cognitively rich simulated environments available. DeepMind has previously used games like StarCraft II and chess as AI training environments.

US Government to Test Google, Microsoft & SpaceXAI Models Before Public Launch

Policy

The US Department of Commerce's Center for AI Standards and Innovation (CAISI) announced agreements this week to evaluate frontier AI models from Google DeepMind, Microsoft, and SpaceXAI (formerly xAI) before they reach the public. Testing will occur in classified environments and will cover safety, security, and best-practice benchmarks.

OpenAI and Anthropic signed similar agreements with CAISI in 2024 under the Biden administration. The expansion to Google, Microsoft, and SpaceXAI extends the pre-deployment review framework to cover all major frontier AI developers — a significant step toward a de facto AI evaluation regime in the US.

xAI Is Officially Gone — Grok and X Now Live Inside SpaceXAI

xAI → SpaceXAI

Elon Musk confirmed this week that xAI is being fully dissolved as an independent company. Grok, the X social platform's AI assistant, and all related AI products now operate under the SpaceXAI brand — a division of SpaceX. The merger, first announced in February 2026 in a $1.25 trillion combined deal, has now concluded operationally.

11 of the original 12 xAI co-founders have departed the combined entity, leaving Musk as the sole founding member. SpaceX's Colossus supercomputer cluster — originally built for xAI — is now being leased to Anthropic to power Claude workloads, an arrangement Anthropic cited when announcing its doubled Claude Code rate limits. Musk is exploring orbital data centers as the next compute frontier.