🚀 OpenAI Releases GPT-5.5 “Spud” — A New Flagship Model
OpenAI has officially launched GPT-5.5, codenamed “Spud,” its latest flagship large language model. The new model is now the default for ChatGPT and is available via the API. Early benchmarks show significant improvements in reasoning, coding, and long-context tasks over GPT-5.
Alongside the model, OpenAI introduced Daybreak, a cybersecurity platform powered by GPT-5.5 that automatically discovers and validates software vulnerabilities — reportedly uncovering 7× more flaws than traditional tools. The launch was notably quirky: OpenAI let GPT-5.5 plan its own release party, a move that drew attention from Business Insider and others.
The company also released GPT-5.5 Instant, a lower-latency variant for real-time applications. Databricks has already integrated GPT-5.5 into its enterprise agent workflows.
⚠️ Anthropic Withholds “Mythos” AI Model — Too Powerful for Hackers
Anthropic has developed Mythos, described as its “most powerful AI model ever developed,” but has withheld public release due to its exceptional offensive cybersecurity capabilities. According to reports from The NYT, Axios, and Fortune, Mythos discovered thousands of zero-day vulnerabilities across major systems, including Apple’s macOS security stack.
The decision has sent shockwaves through the cybersecurity community. US banks are reportedly rushing to patch vulnerabilities identified by Mythos. Anthropic is sharing findings with global financial watchdogs. The Guardian and NYT have published multiple pieces examining whether Mythos represents a “reckoning” for cybersecurity or a potential threat if the technology were to be compromised.
🏛️ White House Considers Vetting AI Models Before Public Release
The White House is weighing a proposal to require pre-release government review of advanced AI models. The NYT first reported the consideration, and Reuters, Bloomberg, and CNBC have confirmed the discussions. Under the plan, models from Google, Microsoft, xAI, and others could undergo federal security testing before public deployment.
The proposal has drawn mixed reactions — some policy experts call it a necessary safeguard, while the American Enterprise Institute labeled it “bad policy.” Multiple Politico reports note that the White House’s “lack of organization” has AI lobbyists on edge, and the administration has already begun testing select models in a pilot program.
💰 OpenAI Launches ChatGPT Personal Finance Dashboard with Bank Integration
OpenAI is rolling out a personal finance dashboard for ChatGPT Pro subscribers that allows users to connect bank accounts and receive AI-powered financial insights. Users can track spending, ask questions about their finances, and get personalized recommendations — all within the ChatGPT interface.
The feature, first reported by TechCrunch, positions ChatGPT as a broader consumer financial tool, competing with apps like Mint and YNAB. It uses read-only bank access via Plaid and similar middleware.
🔧 Google Partners with Marvell for Custom AI Inference Chips
Alphabet is in advanced talks with Marvell Technology to design and manufacture custom AI inference chips, according to reports from Reuters and The Information. This move would diversify Google’s AI hardware supply chain beyond its in-house TPU partnership with Broadcom and reduce reliance on Nvidia’s GPUs.
Marvell shares rose sharply on the news. The chips would focus on inference workloads — running already-trained models — rather than training, a market where Nvidia remains dominant but faces increasing competition.
🎬 Runway Raises $315M to Take on Google and OpenAI in AI Video
Runway, the AI video generation startup, has closed a $315M funding round as it positions itself to compete with Google’s Veo and OpenAI’s Sora in the AI video generation space. CNBC reports Runway’s latest model beats both Google and OpenAI in key video generation benchmarks.
Runway also launched a $10M Builders program to support early-stage AI startups, signaling its ambition to build an ecosystem beyond its core video tools.
📚 arXiv to Ban Authors for One Year if AI Writes Their Papers
The research preprint repository arXiv has announced a strict new policy: authors caught submitting papers primarily written by AI will face a one-year ban. The move, reported by TechCrunch, The Verge, and 404 Media, comes as AI-generated “slop” has increasingly flooded academic submissions.
The ban applies to papers where large language models are used to generate substantial content without meaningful human contribution. Standard use of AI for editing or grammar assistance remains acceptable.
🤖 Microsoft AI Chief: All White-Collar Work Could Be Automated in 18 Months
Mustafa Suleyman, Microsoft’s AI CEO, has predicted that AI will be capable of automating all white-collar desk jobs within 18 months. Speaking to Fortune, Suleyman described a future where most knowledge work is handled by AI agents, requiring humans to shift to oversight and creative roles.
The bold prediction has sparked debate across the tech industry, with some analysts questioning the timeline while others point to the rapid pace of model improvements as supporting evidence.
⏰ Limited-Time Free Models
Several models with limited free access are expiring soon on OpenRouter:
| Model | Expires | Pricing | Context |
|---|---|---|---|
| Arcee AI: Trinity Large Preview | May 22 (2 days) | $0.15/$0.45 per M tokens | 131K |
| Mistral: Mistral 7B Instruct v0.1 | May 30 (10 days) | $0.11/$0.19 per M tokens | 4K |
| Google: Gemini 2.0 Flash Lite | Jun 1 (12 days) | $0.075/$0.30 per M tokens | 1M |
| Google: Gemini 2.0 Flash | Jun 1 (12 days) | $0.10/$0.40 per M tokens | 1M |
🤗 Trending on Hugging Face
Notable new models on Hugging Face this cycle:
- PeterPanonly/Qwen2.5-VL-3B-Instruct-Thinking-SubQ — A vision-language model fine-tuned with GRPO for improved reasoning over sub-questions
- juergengunz/fluxer — Community favorite with 5 likes, likely a Flux model LoRA or derivative
- yalen-ai/distilbert_pii_ner_yalen — PII (Personally Identifiable Information) NER model from Yalen AI, optimized for ONNX deployment
- iamseungpil/metacot-h200-rod-pt-R10v2-0512 — Meta-CoT training checkpoint on H200 hardware
⭐ GitHub Trending: AI Edition
- yaassin12/DeepSeek-V4-Pro-App ⭐ 617 — C++ desktop app for DeepSeek V4 Pro with 1.6T MoE architecture, 1M token context, and agentic coding features
- mikesheehan54/Claude-Code-Design-AI ⭐ 404 — TypeScript-based AI UI/UX architect using Claude Code: screenshot to React, Tailwind CSS, shadcn/ui integration
💡 Key Trends
Today’s news paints a picture of an industry accelerating on all fronts: frontier model releases (GPT-5.5, Mythos), regulatory pressure (White House pre-release vetting), hardware diversification (Google × Marvell), and consumer AI expansion (ChatGPT Finance). The Mythos story in particular marks a turning point — an AI model so capable at cybersecurity that its creators chose not to release it, triggering a global conversation about responsible capability development.