Last week the AI fire hose hit max pressure. New open models, billing drama, a Pentagon deal, and a cancer breakthrough all landed in the same seven days. The creator behind this roundup waded through 300+ announcements and pulled out the ones that actually move the needle.
Here’s the curated rundown.
🔓 Open models are catching up fast
The big shock came from DeepSeek V4. The original poster broke down the numbers: nearly state-of-the-art performance, 1 million token context, and pricing at $1.74 per million input tokens vs $5 for GPT 5.5 and $25–$30 for Claude Opus 4.7 output. Same ballpark capability, fraction of the cost.
The expert also flagged three more open drops in the same week:
- Nvidia Neotron 3 Nano Omni (vision, audio, language, agent-ready, runs on a DGX Spark box)
- Poolside Laguna XS2 (33B open weights) and M1 (225B)
- Mistral Medium 3.5 for remote agents inside Open Claw and Hermes
His take: open weights are now “good enough” for most real workflows. Document summarization, customer support agents, pattern finding. Frontier models are overkill for that.
🎙️ Other model news worth a glance
- Alibaba’s Qwen Image 2.0 Pro shipped (ranked 9th on Arena, decent for memes and text)
- xAI’s Grok Voice Think Fast is now powering Starlink phone support with near-zero latency
🔥 The drama corner
The Musk vs Altman trial finally started. The post’s author shared real-time updates from The Verge: Musk reportedly refused yes/no questions, scolded the defense lawyer, and one juror was visibly rubbing her head. Not a great look on the stand.
Then came the Anthropic mess. The contributor surfaced a brutal find from Theo Brown and Om Patel: if your git commit contains “Hermes.md” or “OpenClaw,” Claude Code would either refuse the request or quietly bill you extra. One user got hit with $200 in overage despite having 86% of his Max plan untouched. After the posts hit 1.4M+ views, Anthropic refunded and issued credits.
Google quietly signed a Pentagon deal for classified AI use, despite a 2014 promise to DeepMind founders that the tech would never go to military or surveillance. Over 600 Google employees signed a letter against it.
China blocked Meta’s $2B Manus acquisition, even though staff and capital had already moved.
⚡ Rapid fire updates
- Microsoft and OpenAI restructured: AGI clause gone, license now non-exclusive through 2032
- OpenAI models, Codex, and managed agents shipped on AWS the very next day
- Gemini now generates PDF, DOCX, XLSX, CSV, LaTeX, and more directly in chat
- Google Translate added AI pronunciation feedback (English, Spanish, Hindi)
- Google Photos is rolling out AI try-on using clothes it has already seen you wear
- ElevenLabs launched ElevenMusic for remixing and earning from AI tracks
- Spotify added a “Verified by Spotify” green check confirming an artist is human
- X rebuilt its ad manager with AI-driven targeting and semantic relevance
🩺 The story I love most
Mayo Clinic built an AI that spots pancreatic cancer on routine abdominal CT scans up to 3 years before clinical diagnosis. The author back-tested it on old scans of confirmed patients and the model caught the subtle signs early. This is the AI promise people actually signed up for.
Pro tip from the curator: when an open model is “good enough” for your use case, the privacy and cost savings of running it locally usually win.
👉 Watch the full video for the benchmark deep dives, courtroom quotes, and links to every model dropped this week.