AI Daily 2026.11.12: Anthropic IPO Valuation Heads for $1.5T; OpenAI Q3 Revenue Tops $10B; Baidu Releases ERNIE 5.5 Flash Preview
Anthropic used an irrevocable, $100B-plus agreement with Microsoft and Google to push its IPO valuation back toward $1.5T; OpenAI crossed $10B in revenue yet posted a wider loss, showing the tension between scale and profitability is unresolved. On the model front, Baidu, Runway and Google keep widening the capability boundary; on the safety front, xAI Grok's lawsuits and 94% of login-free apps put "the stronger the capability, the heavier the responsibility" squarely back on the table.
AI Daily 2026.11.12: Capital Moves, Model Releases, and Security Scrutiny in Parallel
Capital: Anthropic Signs $100B+ Irrevocable Deal, IPO Valuation May Hit $1.5T
As disclosed by X blogger Rohan Paul, Anthropic announced in its amended S1 filing that it has reached new agreements totaling over $100 billion with Microsoft and Google. Microsoft committed to spending at least $47.2B (up to $49.2B) by June 2033; Google raised its minimum commitment from about $114.5B to about $121.3B. Both are irrevocable payments—due regardless of whether capacity is used. The filing also noted $5B in short-term compute spend last quarter, at markedly higher prices.
Bolstered by the news, demand for Anthropic remains hot, and underwriters plan to lift the IPO target valuation back to $1.5 trillion or higher. Separately, per The Wall Street Journal and The Information, OpenAI's Q3 revenue grew about 50% quarter-over-quarter to roughly $10 billion, but adjusted operating loss widened from $10.1B to $10.9B—profitability pressure remains significant, and the tug-of-war between scale and breakeven is still the shared challenge for leading labs.
Model Releases: Baidu ERNIE 5.5 Flash Preview, Runway Praxis 1, AlphaGenome 2 Land in Force
Baidu's ERNIE team released ERNIE 5.5-Flash-Preview, the first post-trained version built on the earlier Base model, with 280B total parameters, 14B activated, and a 1M context window. The API is live on the Qianfan platform, and weights will be open-sourced by November 25. Demos spanned game development, 3D modeling, knowledge work, and video creation; Baidu labeled it an intermediate post-training version to be further optimized.
Runway released Praxis 1, a world-action model with open weights under the Apache 2.0 license. A 27B-parameter MoE diffusion Transformer, it can predict video and actions simultaneously, topping the NVIDIA RoboLab-120 leaderboard at 46.2% success—ahead of FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%), and π0.5 (28.0%).
Google DeepMind released the next-generation DNA sequence model AlphaGenome 2, predicting how DNA variants affect gene expression across hundreds of human cell types. Switched to a conditional architecture and trained on over 107 million single-cell samples across 68 tissues, it identifies nearly twice as many causal variants as its predecessor on TenK10K (30% vs 16%) and generalizes to new single-cell experiments without retraining.
Product Updates: Google Opens Gemini 4 Carbon; Gemini Decisions API Ships
Google officially opened Gemini 4 Carbon, focused on deep reasoning and long-horizon agent tasks, with a million-token unified context window and three tiers of reasoning depth. It suits high-difficulty scenarios like bug review, architecture design, and scientific derivation, and can pair with Argon and Flash into a multi-agent pipeline; Google also provided pricing, migration notes, and hands-on tips.
Google simultaneously launched the Gemini Decisions API, powered by the new Gemini Decision 1 model, capable of evaluating up to 1,000 yes/no, multiple-choice, and scoring questions over text, image, audio, video, and PDF (up to 256K tokens) in a single pass within 150–250ms end-to-end latency. MiniMax also open-sourced its SOTA image model H3 Image, supporting native 4K output and up to 16 reference images for game storyboards, illustration, and product visualization.
Industry: xAI Grok Faces Class Actions; 94% of Live Built Apps Accessible Without Login
Per Bloomberg, multiple underage victims filed class-action lawsuits against Elon Musk's xAI over Grok being used to alter ordinary photos into child sexual abuse material (CSAM); lawyers estimate thousands may qualify for claims.
Security firm Claroc reported that of 37,571 still-online Grok-built apps examined, 35,578 (94.7%) were accessible without login—because the publish dialog preselects "Anyone with the link," and each app's dedicated subdomain enters a searchable passive DNS database, posing serious privacy and security risks. It is the latest strain on the AI development trust chain, following this round's OpenAI ban wave, the Claude Code credential leak, and Anthropic's Supported Regions Policy taking effect.
Research & Safety: Anthropic Publishes MAMUBench; Fable 5.2 vs Gemini 4 Carbon Tested Side by Side
The Anthropic Fellows Program built four multi-user, multi-agent simulation environments (shared API token budget, clinic calendar, group ordering or travel booking, merged queue), tested frontier models from Anthropic, OpenAI, and Alibaba, and released three of them as the MAMUBench benchmark—a new tool for evaluating multi-agent collaboration.
Separately, developer side-by-side tests of Claude Fable 5.2 and Gemini 4 Carbon concluded that Fable 5.2's standout difference is aesthetic sense—a capability dimension that cannot be directly quantified by benchmark numbers but is intuitively perceptible.
Quick Hits
- Per WSJ, NVIDIA's board is nearing a decision on whether to remove CEO Jensen Huang (source: X, Rohan Paul).
- Cursor Projects added multi-user collaboration with real-time co-editing of projects (source: X, Matt Palmer).
- a16z published an article proposing the B2A (Business-to-Agent) economy, with AI agents becoming a new traffic and distribution channel (source: a16z News).
- OpenRouter released an LLM quantization explainer, benchmarking 1,455 provider endpoints and six Qwen3.8 27B endpoints (source: OpenRouter Announcements).
What Developers Can Do Next
- Deploy ERNIE 5.5-Flash-Preview or Gemini 4 Carbon with one click via Celedog.io to build million-token-context agent apps without worrying about underlying GPU scheduling.
- Use Celedog.io's massive Agent Skill marketplace to wire robotics-action generation like Praxis 1 into existing workflows, lowering the end-to-end barrier for agent development.
- Use Celedog.io's smart model routing to switch automatically between Gemini 4 Carbon (deep reasoning) and lightweight models, allocating compute and cost by task complexity.
- Watch Celedog.io's newly added MAMUBench-related skills to quickly evaluate budget allocation and queue-strategy performance across models in multi-agent collaboration.
- For trust-chain incidents like "94% of Grok apps accessible without login," prefer a gateway on a compliant access channel: Celedog.io is served from 9 global nodes (Singapore, Germany, Australia, Indonesia and more) under a globalized operating entity—not a capital entity from an unsupported region—so global users can use the platform with confidence.
Last updated November 12, 2026
Where to go next
- Try Celedog — free credits on signup, no card required.
- API documentation
- Per-model pricing
- More Celedog Blog