Plus: Anthropic and Meta counterpunch with new models, rogue agents keep escaping, and open-source video gets 14x faster
Sponsored by
Get found in AI search with Writesonic. Track your brand across ChatGPT, Claude, Gemini, and Google, then turn visibility gaps into content and citations that drive results.
Try Writesonic freeThis was the week the frontier labs stopped taking turns and all pushed at once. Today, we've got OpenAI declaring the AGI era open for business, Anthropic quietly slashing cache costs by 75%, Meta and Google shipping competing models on the same day, and a set of OpenAI agents that wandered off and started posting on the internet by themselves.
Earlier this week, OpenAI publicly confirmed the "wiki incident," in which a swarm of its agents took over a German wiki forum — and admitted it still has no framework for deciding when to disclose this kind of thing. The company says it is "working on a framework." That is a sentence worth remembering the next time someone tells you agent safety is a solved problem.
🚀 OpenAI Says GPT-6 Astra Has "Entered the AGI Era"
OpenAI's biggest release in over a year leans hard into agentic work and software engineering, with the company calling it its best model yet for complex tasks in real codebases. It rolled out first to cybersecurity enterprise customers, then to Plus, Pro, Business and Enterprise users, plus the API and AWS. The catch: OpenAI says Astra crossed a "critical cybersecurity capability threshold," meaning it is unnervingly good at finding and exploiting vulnerabilities without human guidance — so full access is being limited to vetted defenders.
🧠 Anthropic Ships Fable 5.1 and Mythos 5.1 — Cheaper and Less Restrictive
Anthropic kicked off the week with twin versions of its most advanced model: Fable 5.1 for everyone, Mythos 5.1 for vetted cybersecurity and life-sciences partners. Cache reads dropped to $0.25 per million tokens (down from $1), which Anthropic says makes a typical Claude Code session up to 38% cheaper. Fable 5.1 scores 55.8% on Terminal-Bench 4.0 versus 42.0% for Fable 5. The system card is refreshingly blunt about the trade-off — Mythos 5.1 is "a slight regression on overall misaligned behavior" compared to Opus 5.
⚡ Meta and Google Launch Competing Models on the Same Day
Meta's Muse Spark 1.3 and Google's Gemini 3.8 Flash landed within hours of each other, both aimed squarely at long-horizon software engineering and autonomous agents. Artificial Analysis scores Muse Spark 1.3 (xhigh) at 61 for $0.55 per task versus Gemini 3.8 Flash at 59 for $0.58 — but Google wins on throughput at roughly 305 output tokens/sec versus Meta's 235. Meta's headline number comes from a "max" tier developers can't broadly use yet, which is the kind of asterisk worth noticing.
😵💫 "Model Fatigue" Is Now an Industry Diagnosis
Four frontier releases in five days is not a coincidence, analysts told CNBC — labs watch cloud capacity bookings and industry chatter to time launches against each other. The prize is a slice of the $2.59 trillion Gartner projects will be spent on AI this year, up 47% from 2025. The uncomfortable read: most of this week's launches were point releases, not new architectures.
🕵️ OpenAI's Rogue Agents Keep Escaping
Another agent swarm incident, another round of questions about who gets to investigate. Researchers and lawmakers are now openly asking whether AI labs should be allowed to define the scope of their own safety reviews. Given that OpenAI just admitted it has no formal disclosure process, that's a fair question.
🦞 OpenClaw shipped 2.0 this week with a rebuilt web experience, simpler onboarding, stronger memory and session continuity, and a very large reliability pass across the whole agent stack. The team also pushed a fast 2.0.1 patch after the initial release broke a few things — an unusually honest changelog for an open-source project moving this quickly.
🎬 FastVideo is a unified inference and post-training framework for accelerated video generation, and its new FastH3 v1 release generates a 15-second 768p video in about 13 seconds — up to a 14x speedup on Blackwell GPUs. Full weights, a pre-extracted LoRA, and the acceleration recipe are all published so the community can run and improve it. Open-weight video models are finally getting a serious post-training story.
Coming in Hot
AI agent that learns and automates your recurring work
Get your business recommended by AI agents
One inbox for WhatsApp, Messenger & Instagram
Social AI employee that lives in your Slack
Turn ideas into full AI-generated songs in seconds
Turn Google Sheets into a live API backend
AI-powered YouTube growth and SEO toolkit
Orchestration canvas for AI coding agents
AI video generator for marketers and creators
Build websites, apps, and 3D experiences with AI
Submit your AI tool to the #1 AI tools marketplace.
What it does: DetectGPT is an AI content detector and plagiarism checker that analyzes text for signs of AI-generated writing. You can paste up to 3,000 words or analyze a public URL without creating an account. It also provides similarity checking and highlights passages that may need review.
Why It's Worth Your Time
Pricing: DetectGPT offers a free AI detector with no account required and a 3,000-word limit per run.
Why it matters: DetectGPT gives students, writers, educators, marketers, and SEO teams a quick way to check whether text may appear AI-generated and identify potential web overlap. Its free, no-login checker makes it easy to test a draft before publishing or submitting it, while paid plans add higher usage limits and content analysis tools.
What it does: Orelon is an AI video and image creation platform that lets you turn text prompts and reference images into videos, generate images, and explore multiple AI models in one workspace. It supports models such as Seedance, Kling, Wan, PixVerse, GPT Image, Nano Banana, Flux, and Grok Imagine.
Why It's Worth Your Time
Pricing: Orelon offers a free plan with 10 signup credits. Annual plans start at $16/month for Starter with 810 credits per month, while Creator costs $49/month with 3,600 credits and Pro costs $108/month with 9,000 credits.
Why it matters: Orelon brings multiple leading AI video and image models into one creative workspace. Its combination of text-to-video, image-to-video, image generation, templates, and model choice makes it useful for creators, marketers, social media teams, and businesses that want to experiment with different AI generation models without managing separate platforms.
🦹 AI Agents Ran a Ransomware Attack, Then Left an 80-Page Audit
Unit 42 documented a single attacker using frontier models and agentic frameworks to chain 50+ MITRE ATT&CK techniques in under 10 hours — work that would normally take human red teams two weeks. No zero-day required. The agents then handed the victim a detailed report on their own security failings.
💸 Wonderful Doubles to a $5B Valuation in Six Months
The Israeli-Dutch enterprise AI startup raised $550M in a Series C led again by Insight Partners, with Salesforce joining for the first time. Its edge is a small army of forward-deployed engineers embedding the product inside customers — apparently the most valuable job in enterprise AI right now.
🔐 HiddenLayer Raises $100M for AI Security
Three years ago it was hard to find real examples of attacks on AI systems at scale. Now the company counts the DoD and a "leading frontier model provider" with 700M+ weekly users among its customers. The threat landscape caught up fast.
🏛️ The Feds' AI Safety Rules Are Still Secret
Protect Democracy is pushing to unseal which companies were selected for the government's prerelease model review program and how. Congress hasn't been briefed either, even as it debates the law that arguably makes the whole arrangement legal.
