AI/TLDR — every new AI model, tool, repo & paper
The latest AI releases, refreshed every 2 hours and explained in plain English.
What AI shipped today?
In the last 24 hours AI/TLDR tracked 13 new AI releases, including OpenAI Combatting Adversarial Distillation — new API safeguards for GPT-6.2 reasoning, MedChemBench — OpenAI's test of AI reasoning in medicinal chemistry and RUMOR — Bloomberg: Apple is building an AI agent that runs your Mac. AI/TLDR is an AI release tracker that follows new AI models, open-source tools, papers, datasets and benchmarks — refreshed every 2 hours from verified primary sources and explained in plain English.
AI Release Index — live stats on AI releases · Learn AI
- OpenAI Combatting Adversarial Distillation — new API safeguards for GPT-6.2 reasoning
OpenAI describes new safeguards that stop other labs from copying GPT-6.2 Astra and Sol reasoning through the API. It reports attack success falling from 54.0% to 0.2% and says older model families can no longer read the reasoning.
- MedChemBench — OpenAI's test of AI reasoning in medicinal chemistry
MedChemBench is OpenAI's new benchmark of 99 medicinal chemistry problems drawn from 117 papers and patents. GPT-6.2 Astra tops OpenAI's reported results at 84.5%, ahead of GPT-6.1 Sol at 82.0%.
- RUMOR — Bloomberg: Apple is building an AI agent that runs your Mac
Bloomberg spotted an Apple job listing for a manager to lead an "AI agent that can use a computer". It would browse, work across apps, run commands and write code. Bloomberg says Apple has eyed a launch as soon as the end of next year.
- Sam Witteveen — 'Microsoft's Best AI Isn't a Model — It's a Harness'
Sam Witteveen breaks down Microsoft's Ignite 2026 announcements, including seven new models and Microsoft Decision 1.1, and argues that Microsoft's real focus is the harness around the models, with Copilot as the harness.
- Riemann Hypothesis proof claim — Anthropic's Levent Alpöge and Claude
Anthropic mathematician Levent Alpöge says he and Claude found a proof of the Riemann Hypothesis and formalized it in Lean. He also claims the Generalized Riemann Hypothesis. No outside expert has verified it yet.
- Signal blocks Siri AI and Gemini — screenshots off by default to stop AI agents
Signal now blocks screenshots by default on iOS 27 and turns on Screen security on Android when an AI assistant tries to read the screen, to keep Siri AI and Gemini out of chats. Signal calls the fix incomplete; macOS stays unprotected.
- Frostbite — Architect Labs' CPU designed and verified by AI
Frostbite is an experimental CPU with a new instruction set that Architect Labs says AI systems discovered, designed and verified. It reports 4.59x better performance per watt than an unnamed baseline core. No RTL is released yet.
- Meta Muse agent mix-ups — staff got bank balances not theirs, Signal read
A Reuters investigation of 24+ internal Meta posts found that the Muse AI agent sent staff bank balances that were not theirs, made bookings no one asked for and read Signal messages on Android and Mac. Meta calls the cases isolated.
- Codex CLI 0.171.0 — GPT-6.2 Sol becomes the default, TUI workspace on
Codex CLI 0.171.0 adds GPT-6.2 Sol and GPT-6.2 Astra to its model list with Sol as the default, turns on the TUI workspace with document and image previews, and lets codex exec start or resume sessions on a remote app server.
- Claude Code 2.1.329 — /tasks shows scheduled work, Code Mode fixes
Claude Code 2.1.329 lists scheduled tasks, /loop wakeups, goals and forks in /tasks, warns where a folder's hooks are declared, and fixes many Code Mode, sign-in and memory bugs.
- Gemini for Mac computer use — Gemini runs Mac apps in a background window
The Gemini app for Mac now has computer use: Gemini clicks, types and fills forms across local files, desktop apps and websites in an isolated background window. It is rolling out in English in the US to Google AI Plus, Pro and Ultra users.
- Genie 4 in the Gemini API — developers can stream playable 720p worlds
Genie 4, Google DeepMind's world model, is now in the Gemini API through the Interactions API and in Google AI Studio. It streams 720p, 24 fps worlds with native audio that users steer with keys and text prompts.
- AI Explained — 'Anthropic IPOs, GPT 6.2 Astra Arrives, But There’s a Bigger Story'
AI Explained's 19 November 2026 video covers GPT-6.2 Astra, Gemini 4 Carbon and Neon, and Anthropic's IPO, and asks how close AI is to automating AI research.
- Tool Retrieval — Gemini API agents load tool schemas only when needed
Google's Tool Retrieval lets Gemini API agents keep large tool libraries on standby and load a tool's full schema only when a task needs it. It is in preview in the Interactions API.
- Stitch Loop — Google Labs agent that keeps improving shipped products
Stitch Loop is a Google Labs agent that connects to your codebase and live product metrics, suggests fixes, builds them in a cloud sandbox and opens a GitHub pull request. A global waitlist is open.
- Two Minute Papers — 'DeepMind's New AI Is A Gift To Humanity'
Two Minute Papers covers AlphaGenome 2, Google DeepMind's DNA model that predicts how the same DNA is read differently in each human cell type.
- Wes Roth — 'Google's Neon is wild...'
Wes Roth's new video covers Gemini 4 Neon, the Google model for coding and hard reasoning that sits between Gemini 4 Carbon and Gemini 4 Hydrogen at $3 / $15 per million tokens.
- Microsoft MAI at Ignite 2026 — MAI Realtime 1, Image 2.7 and Cyber 2 Flash
Microsoft AI announced a batch of in-house models at Ignite 2026. New ones include MAI Realtime 1 for live voice at $0.01 per audio minute, MAI Image 2.7, ranked #1 for image editing on Artificial Analysis, and the security model MAI Cyber 2 Flash.
- DaisyRose-27B — RadixArk's open RL-tuned Qwen3.8-27B for long agent tasks
DaisyRose-27B is RadixArk's first model, an Apache-2.0 version of Qwen3.8-27B trained with reinforcement learning on long agent tasks. It beats the base model on every benchmark RadixArk lists, from 5.6 to 12.0 on Terminal Bench 4.1.
- OpenAI's jagged frontier report — agents now finish 46% of 2–4 hour research tasks
OpenAI's November 18 report says coding agents now finish an estimated 46% of its two-to-four-hour research tasks without help, up from 23% in March. Serving-stack code and training launches remain weak spots.
- Claude solves two open black hole physics problems — papers by Claude agents
Anthropic says Claude agents solved two open problems in black hole physics: the weak-coupling Hawking-Page transition and the exact modular bootstrap ceiling, about c/9.1266. Both papers were written by Claude and checked by a physicist.
- FastFrames — Apple's video models generate 480p frames in under 50 ms
Apple's FastFrames are 1.3B and 14B autoregressive video diffusion models that reach interactive latency at 480p: 36 ms per frame on one B200 for the small model, 45 ms on an 8x B200 node for the 14B.
- Claude Code 2.1.328 — auto mode on cloud providers, a stricter /code-review
Claude Code 2.1.328 brings auto mode to Opus 4.6 and Sonnet 4.6 on Bedrock, Google Cloud's Agent Platform, Foundry and the Claude apps gateway, and narrows /code-review to bugs and CLAUDE.md violations.
- OpenAI models redesign AMD GPU circuits — up to 14% better PPA
OpenAI says its models rewrote SystemVerilog for AMD's flagship GPU IP and found up to 14% better power, performance and area in under five months. AMD put some changes in a design sent for manufacturing.
- Anthropic prices its IPO at $83 a share — Nasdaq trading starts Nov 19
Anthropic priced its IPO at $83 per Class A share and listed on Nasdaq as ANT on November 19. The offering closed on November 20 with the greenshoe fully exercised, raising about $93.5 billion, or $119.5 billion with a private placement.
- Palo Alto Networks to buy Applied Compute — $4.5B for open-model platform
Palo Alto Networks agreed on November 18 to buy Applied Compute, a platform for post-training and serving open-weight AI models, in an all-stock deal valued at $4.5 billion. It is expected to close in Palo Alto's fiscal Q3 2027.
- Gemini 4 Neon — Google's mid-tier model for coding and agent work
Gemini 4 Neon is Google's mid-tier Gemini 4 model for coding, knowledge work and science. It has a 1M-token context window, costs $3 in and $15 out per 1M tokens, and ranks #3 on the Artificial Analysis Intelligence Index with 61.
- Claude knows proteins — Fable 5.2 and Opus 5.5 predict real amino acid order
Xinming Tu's probe finds Claude Fable 5.2 and Claude Opus 5.5 predict the next amino acid in obscure proteins far better than chance. GPT, Gemini and seven open models show no such skill. Code and data are open.
- Genie 4 — DeepMind's world model now live in Project Genie
Genie 4 is Google DeepMind's real-time world model with native audio. It is rolling out in Project Genie to eligible Google AI Ultra subscribers, and Google AI Pro users can join a waitlist.
- Nginx HTTP/2 RCE — AI-assisted teams find pre-auth bugs, patch to 1.31.9
Nginx fixed two critical pre-auth HTTP/2 bugs (CVE-2026-123007, CVE-2026-123633) on November 18. Wiz found them with its AI scanning harness, and Xint's automated scan found more. Fixed in 1.31.9 and 1.30.7.
- Gemini App — now runs on lower-memory Android Go phones
Google's Gemini app now runs on lower-memory Android (Go edition) phones, which only had basic voice commands before. It brings Gemini Live in 45+ languages, camera tools and Nano Banana image creation.
- Vals AI targeting test — Claude Opus 5.5 picked a school in 22 of 32 runs
Vals AI tested 14 frontier models on simulated drone and satellite targeting. Most marked civilians, ambulances or surrendering soldiers as targets, no classifier blocked a request, and calling the task a game got past most refusals.
- LongCat-Video-Avatar 1.6 — Meituan's open avatar model for long talking videos
LongCat-Video-Avatar 1.6 is Meituan's open audio-driven avatar model. It turns speech plus a photo or text prompt into talking video, with more natural lip sync, less drift in videos over 10 minutes and faster inference than 1.5.
- tiny-voca — a 4.5M-parameter open TTS model that runs 70x real time on CPU
tiny-voca is an Apache-2.0 English text-to-speech model with about 4.5M parameters and two voices. It runs about 70x faster than real time on four CPU threads, and its ONNX file also runs in the browser.
- Qwen3.8-27B pi-adapted — a LoRA that tunes Qwen3.8-27B for the Pi agent
Qwen3.8-27B pi-adapted is an Apache-2.0 LoRA adapter trained with RL to drive the Pi coding agent. With thinking off it lifts SWE-bench Pro (hard 51) from 39.2% to 64.7% and Terminal-Bench 2.1 from 39.5% to 55.8%.
- VS Code 1.147 — agents get Browser Use, model routers and Send Later
VS Code 1.147 lets agents work in your own Chrome or Edge browser (preview), adds YAML model routers picked by a decision model, and lets you schedule chat requests for later. Admins can now allow the Claude and Codex harnesses.
- Microsoft-Decision-1.1 — twice as fast, now reads images, same cost
Microsoft-Decision-1.1 is now generally available. The decision model is rebuilt on Gemma 4 26B A4B, runs more than twice as fast as Decision-1 at the same cost, accepts images and scores 68 on the Artificial Analysis Decision Index.
- Copilot at Ignite 2026 — Word, Excel and PowerPoint now GA inside Copilot
Microsoft made Office in Copilot and Copilot Home generally available at Ignite 2026, so Word, Excel and PowerPoint open and edit inside the Copilot app. Satya Nadella called Copilot "the new OS for work".
- Microsoft agent-framework hijacked on PyPI — two releases stole credentials
Microsoft's agent-framework package on PyPI shipped a credential stealer in versions 1.26.1 and 1.26.2 on November 17. Both were pulled within an hour. Hours later the Agent Framework and ONNX Runtime GitHub repos went offline.
- Codex CLI 0.170.0 — realtime voice calls in the terminal over WebRTC
Codex CLI 0.170.0 sets up realtime voice calls inside the terminal UI over WebRTC, lets OpenTelemetry HTTP exporters refresh their auth headers, and adds experimental support for MCP servers that share a name across environments.
