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 18 new AI releases, including MAI Code 2 — Microsoft's coding model at 40% of GPT-6.1 Sol's cost, GPT-6.2 Astra and Sol — OpenAI's new flagship pair, Astra 20% cheaper and FrontierHarness Eval v1.1 — which agent harness works best for each model. 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
- MAI Code 2 — Microsoft's coding model at 40% of GPT-6.1 Sol's cost
MAI Code 2 and MAI Code 2 Flash are Microsoft's new coding models, live in GitHub Copilot. Microsoft says Code 2 reaches 98% of GPT-6.1 Sol's score at 40% of its cost, at a 50% launch discount.
- GPT-6.2 Astra and Sol — OpenAI's new flagship pair, Astra 20% cheaper
GPT-6.2 Astra and GPT-6.2 Sol are OpenAI's new frontier and everyday models. Both use fewer output tokens per task. Astra drops to $8 / $40 per 1M tokens, 20% below GPT-6 Astra; Sol stays at $2 / $10.
- FrontierHarness Eval v1.1 — which agent harness works best for each model
FrontierHarness Eval v1.1 tests 22 coding-agent harnesses with 5 open models on 62 terminal tasks. The best harness changes a lot by model, and the model explains 92.8% of the cost differences.
- How the Claude Code team verifies agent code — tests, reviewers and gates
Addy Osmani explains how the Claude Code team checks code written by Claude: end-to-end, property and differential tests, a tuned reviewer, and a commit gate that blocks unverified changes.
- Tasks in Gemini — free agent that books, shops and pays for you
Tasks in Gemini lets the Gemini app call businesses, review bills, shop and check out with Google Pay. It is free in six countries, runs on Gemini 4 Hydrogen, and asks before it buys.
- Sam Witteveen — 'RAON-1 Nano: Can a Tiny LLM Power Game NPCs?'
Sam Witteveen tests KRAFTON's Raon-1 nano, a small Apache-2.0 model distilled from the larger Raon-1 mini, and asks whether a model this small can run the characters in a video game.
- Two Minute Papers — 'OpenAI's New Breakthrough Is... Utter Chaos'
Two Minute Papers covers OpenAI's claimed proof that quantum Yang-Mills theory exists, linking the statement from the physicists' working group that reviews the AI-written proof.
- Sage S2.1 Flash — an open 320B model that reads molecules, spectra and EEG
Sage S2.1 Flash is Shanghai AI Lab's MIT-licensed 320B MoE for science agents. It takes molecules, proteins, spectra and EEG as input and scores 90.3 on MLE-bench Lite, ahead of Claude Opus 5 in the lab's table.
- OpenAI Daybreak for NATO — allied cyber agencies get OpenAI's defense models
OpenAI says NATO's core institutions now use its Daybreak cyber defense models through the NATO Communications and Information Agency, at up to 300 billion tokens a month. ENISA and agencies in France, Germany and Poland also have access.
- ChatGPT for Academic Researchers — OpenAI widens free access to 200,000
OpenAI is expanding ChatGPT for Academic Researchers toward 200,000 researchers, double the pilot's goal. 50,000 new grants of free ChatGPT Plus open on November 23, first come, first served, each for a researcher and two collaborators.
- OpenCode Browse — MCP server that lets agents use your real browser
OpenCode Browse is an MCP server and Chromium extension that lets AI agents use your real browser, with your open tabs, logins and extensions. It works with any MCP client, including Claude Code, Cursor, Codex and Zed.
- Claude Code 2.1.326 — Code Mode works with agent teams, sandbox tightens
Claude Code 2.1.326 lets Code Mode run with agent teams, keeps sandboxed commands out of sensitive home folders and Claude's sign-in files, and streams Remote Control replies to your phone or browser as they are written.
- Sam Witteveen — 'ZGX Fury - How Fast Can it Go?'
Sam Witteveen returns to HP's ZGX Fury to test which models it can run and how fast, from memory math and speculative decoding to offline coding agents, private RAG and video generation.
- Windows 11 Tracks Which Coding Agents You Run — Optional diagnostics report
Windows 11 records which coding agents and developer tools run on a PC and reports it as optional diagnostic data, says researcher Xusheng Li. The report holds no prompts or code. It can be switched off in Settings.
- vLLM v0.34.0 — a bigger KV cache by default, and Model Runner V1 is gone
vLLM v0.34.0 brings 1,061 commits from 412 contributors. The KV cache is now sized from memory free after warmup, so most models get more cache. Model Runner V1 is removed, and Mistral Large 4 support arrives.
- ChatGPT for Healthcare agents — they click through hospital software
OpenAI added computer-use agents to ChatGPT for Healthcare. They read screens, click and type across hospital and insurer software. The update also adds 21 healthcare plugins and better clinical answers.
- OpenAI CSAM Image Classifier — free detection of new and AI-made abuse images
OpenAI released a CSAM image classifier, free through its API for vetted organizations, that also flags AI-generated abuse images. ChatGPT now warns users who seek such material and pauses their access for 24 hours.
- OpenAI embedded assessments — four outside labs get inside access to test safety
OpenAI signed Apollo Research, METR, Transluce and Goodfire to assess its models from the inside, with access to checkpoints, training data, source code and monitors. METR will study OpenAI's largest reinforcement learning run to date.
- California sues OpenAI — state says its agents' Hugging Face attack was foreseeable
California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court on November 16. The suit says OpenAI tested dangerous models in a flawed environment, which led to the July attack on Hugging Face.
- Hunyuan Hy4 Flash — Tencent's cheap 1M-context model for everyday work
Tencent released Hunyuan Hy4 Flash, a 224B-parameter MoE with 12B active and a 1M-token context. It costs $0.070 per million input tokens, about a twelfth of the full Hy4. Open weights are planned but not out yet.
- Anthropic cuts its bug disclosure window to 30 days — AI finds bugs too fast
Anthropic will publish high- and critical-severity bugs it reports 30 days after notifying the vendor, down from 90, and other bugs after 45. The change starts January 1, 2027. Anthropic says AI makes rediscovery of bugs much more likely.
- GTIG AI Threat Tracker — attackers now run fully autonomous exploit agents
Google's Threat Intelligence Group found exploit_agent, an attacker framework that reads security patches, writes exploits and runs them with no human in the loop. The report also covers a Claude Code settings.json backdoor.
- Ask Maps — itineraries, route tips and appointment booking
Ask Maps now builds multi-stop itineraries, flags parking, EV charging, weather and tolls on routes, and finds or calls to book local appointments. It also expands to 30 European countries.
- US remote-access chip rule — overseas data centers need a license to serve China
The US Commerce Department issued an interim rule that closes the cloud loophole on advanced AI chips. Data centers anywhere need a license to rent large amounts of compute to companies based in China, Macau or other D:5 countries.
- SkillEvaluator 0.9 — NVIDIA's agent-skill tester adds budgets and version diffs
NVIDIA SkillEvaluator 0.9.0 tests whether an Agent Skill really helps a coding agent. The release adds grounded test-set generation, skill version comparison, token and cost budgets, and clearer evidence in reports. The repo is trending on GitHub.
- Apple Health AI Usage Study — an OpenEvidence app for 200,000 iPhone users
Apple's Health AI Usage Study is a US research app built with OpenEvidence. Up to 200,000 adults can ask health questions, get answers with clinical citations, and optionally share HealthKit data. Study data will not train AI models.
- LoopForge — Tencent's self-hosted workflows for coding agents
LoopForge is Tencent's open-source, self-hosted workspace that runs Claude Code, Codex, CodeBuddy and other agent CLIs in Docker sandboxes, through workflows that mix agents, scripts and human approval gates. MIT license.
- RUMOR — FT: Jensen Huang weighs a return as Nvidia CEO
The Financial Times reported that Jensen Huang discussed returning as Nvidia CEO with confidants. On 16 November a source told CNBC that Nvidia's board is not considering it and is focused on an investigation into Huang.
- Steve Blank: What's Next for Nvidia? — Visionary CEO or operator
Steve Blank argues that Nvidia's board must choose between an innovator and an execution-focused CEO after Jensen Huang's resignation. He says CUDA and the order book can carry Nvidia for two or three years.
- Qwen4-Max — Alibaba's Qwen4 flagship is live on the API at $1.50 / $4.50
Qwen4-Max is Alibaba's first full Qwen4 model: a new MoE with native vision and a 1M-token context. It is live on the QwenCloud API at $1.50 / $4.50 per million tokens, cheaper than Qwen3.8-Max.
- GPT-6 Astra and the Standard Model — a candidate lattice definition
UCSB physicist Nathaniel Craig says GPT-6 Astra led the detailed proof in his paper proposing a candidate lattice definition of the Standard Model. OpenAI published his account on its new science blog on 15 November.
- Wes Roth — 'Claude just changed EVERYTHING...'
Wes Roth looks at how Anthropic is changing Claude ahead of its IPO: Claude as a layer across apps and live meetings, new office tools, verification tiers, and JPMorgan's reported 250,000 weekly Claude users.
- Sebastian Raschka — understanding positional embeddings in modern LLMs
Sebastian Raschka explains how LLMs encode token position, from GPT-2 learned embeddings and sinusoidal encodings to RoPE with PyTorch code, plus partial RoPE, NoPE and mixed RoPE/NoPE layers in 2026 open-weight models.
- Haru-4B — an open speech model that listens and talks at the same time
Haru-4B is higurashi's open full-duplex spoken dialogue model: Apache 2.0, built on Qwen3.5-4B, nine primary languages. It hears and speaks at once, with about 65 ms end-to-end latency on a MacBook.
- Jakub T. Jankiewicz — all AI-generated code should be open source
Jakub T. Jankiewicz argues that all AI-generated code should be open source. Models learned from open code and Stack Overflow, he says, so closing their output is "license laundering", and purely AI-written code may not be copyrightable.
- DeepSeek App API — web search, file input and safety checks built into the API
DeepSeek now offers its app's features through the API. deepseek-flash (V4.1-Flash) can search the web, read links, parse files and run safety checks on the server, and new OCR, ASR and TTS APIs are available too.
- Nathan Lambert — Jensen Huang's exit is a setback for US open models
Nathan Lambert argues that Jensen Huang's resignation as Nvidia CEO is a major setback for US open-weight models. He expects Huang's open-model funding to dry up and says the US needs its own truly frontier open model lab.
- Chrome DevTools MCP 1.13.0 — logpoints and network mocks for coding agents
Chrome DevTools MCP 1.13.0 lets AI coding agents set logpoints in a live page and mock its network requests. The mocks can block a request, serve a local file or fetch from another URL. The release also ships a long list of fixes.
- deepseek-v3-jax — DeepSeek V3 in 190 lines of pure JAX
deepseek-v3-jax is an MIT-licensed, 190-line JAX version of the DeepSeek V3 architecture. The 36.2M-parameter model trained in about 2.6 hours on a MacBook Air M5. The author calls it an architecture demo, not a strong language model.
- Plannotator — review your coding agent's plans and diffs in a browser
Plannotator is an open-source local review tool for coding agents. You annotate an agent's plan, diff or HTML output in the browser and send the feedback back. Version 0.28.64 shipped on 13 November as the repo trends on GitHub.
