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 11 new AI releases, including Claude browser use tool — Anthropic's agent APIs are now generally available, Claude Academy — Anthropic opens a course hub for learning AI and Rafal Cymerys — AI-written docs now get skipped before they get read. 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
- Claude browser use tool — Anthropic's agent APIs are now generally available
Anthropic moved computer use, the Skills API and the Files API to general availability on the Claude Platform, and added a new browser use tool that reads page structure instead of guessing from pixels alone.
- Claude Academy — Anthropic opens a course hub for learning AI
Claude Academy is a learning hub from Anthropic with courses, tutorials and use cases for working with AI. It opens with two courses, one tutorial and five product tracks covering Claude.ai, Cowork, Code, Tag and Claude Platform.
- Rafal Cymerys — AI-written docs now get skipped before they get read
Rafal Cymerys writes that he now skips work documents that look AI-written, the same way people ignore banner ads. He lists the tells: Claude-style phrasing, routine features sold as breakthroughs, and long text that says little.
- Thomas Ptacek — coding agents make native apps cheap, so stop building TUIs
Thomas Ptacek argues that terminal UIs only ever existed because of modems and because Unix developers avoided GUI toolkits. Coding agents now write idiomatic SwiftUI, so a small native Mac app costs about as little as a CLI.
- OpenAI regional processing — pick an inference region per API request
OpenAI now lets a single API call choose where it runs. Swap api.openai.com for a regional domain such as eu.api.openai.com or jp.api.openai.com and that request is processed in that region. Ten regions are live.
- Benchmark Fitting — Hugging Face scores how much ASR models memorize tests
Hugging Face and Hume AI added a Benchmark Fitting tab to the Open ASR Leaderboard. Three probes check whether a speech model transcribes the audio or repeats the reference. Six of 11 models dropped audible words the reference lacked.
- GPT-5.6 Sol price cut — OpenAI drops API rates over 20% for three months
OpenAI cut GPT-5.6 Sol API pricing by over 20% for three months. Input falls from $5 to $4 per million tokens, cached input from $0.50 to $0.40, and output from $30 to $20. Codex credits and ChatGPT Work get the same rates.
- NoBuzz — a Claude Code skill that rewrites Claude's replies in plain English
NoBuzz is a Claude Code skill, /debuzz, that sends Claude's last reply to the Antigravity CLI and prints back a plain-English version. Three modes — colleague, manager and director — set how much detail survives.
- 1littlecoder — 'I Tested Ox Alpha (stealth model)'
1littlecoder's new video is a hands-on test of Ox Alpha, the anonymous reasoning model that appeared on OpenRouter on August 20 with a 1,048,576-token context window and no charge for tokens.
- Claude Mythos 5 comes to Claude Security — Anthropic's cyber model for defenders
Claude Security scans now run on Claude Mythos 5, in public beta for Claude Enterprise customers. Each finding carries a CWE category, severity, confidence and a suggested fix. Anthropic also opened a $35 million Defender Advantage Fund.
- Codex CLI 0.149.0 — a dashboard for every running agent task
Codex CLI 0.149.0 adds an interactive `codex agents` dashboard for searching, starting, opening, renaming and stopping tasks. The same build adds `codex queue`, which sends a message into an existing local or remote session.
- NVIDIA AVO — an agent system scores 100% on ARC-AGI-3's public set
NVIDIA AVO scored 100.00 RHAE on the 25-environment public set of ARC-AGI-3, clearing all 183 levels in 6,624 actions. Claude Opus 5 on its own scores about 30% on the same set. The private competition sets were not tested.
- 4DAnyone — turn one handheld video of a person into a 4D model
4DAnyone rebuilds a moving person in 4D from a single ordinary video. The Ant Research system generates multiview-consistent video first, then lifts it into 4D Gaussian Splatting. Code and checkpoints are public.
- DeepSeek V4-Flash-Vision-Exp — an experimental V4 model that reads images
DeepSeek V4-Flash-Vision-Exp is an experimental multimodal model on the DeepSeek API that takes images alongside text. It scores 83.9 on Terminal Bench 2.1 and matches DeepSeek V4-Flash on text-only work.
- Ox Alpha — an anonymous 1M-context coding model, free on OpenRouter
Ox Alpha is a reasoning model that appeared on OpenRouter under the provider name Stealth, with no lab claiming it. It has a 1,048,576-token context window, takes text, images and video, and costs nothing during the test window.
- SWE-bench Science — the best coding agent fixes under half of 119 science bugs
SWE-bench Science is an MIT-licensed benchmark of 119 repository-level tasks drawn from 98 scientific codebases across 20 domains. On the published leaderboard, Claude Opus 5 leads with 47.90% Pass@1.
- Claude Code 2.1.238 — plugin marketplaces can mint their own auth headers
Claude Code 2.1.238 adds headersHelper, a command that mints HTTP headers such as a short-lived token for private plugin marketplaces. The same build fixes unbounded memory growth in long interactive sessions.
- EnvHarness — Google's wrapper layer makes static agent benchmarks adapt
EnvHarness wraps an existing agent benchmark so its behaviour can be reshaped without editing the benchmark itself. The paper reports up to a 9.0-point gain on held-out tasks with 9.8% fewer steps. Code is Apache-2.0.
- Waymo's custom AI chip — a 5nm ASIC that runs 1,000+ TOPS in the car
Waymo announced its own 5nm ASIC, made by TSMC, that runs over 1,000 TOPS of machine-learning work inside its robotaxis. The chip fuses lidar, radar and 13 camera streams in real time and is in production in Waymo's latest robotaxi generation.
- Slack Code — AI coding agents get a channel the whole team can watch
Slack Code turns a coding agent's work into a shared Slack channel with tabs for the plan, the code diffs and a live preview. Claude, Devin, GitHub Copilot and Vercel work at launch; ChatGPT is coming soon.
- Antigravity IDE Extensions — Google's agent moves into VS Code and JetBrains
Google put Antigravity inside VS Code, Visual Studio, JetBrains and Zed as one-click extensions that share context with the Antigravity desktop app and CLI. A companion release adds budget caps, token pools and audit logging for enterprises.
- LFM2.5-DSpark — Liquid AI's draft models decode up to 3.18x faster
Liquid AI released DSpark draft models for three LFM2.5 checkpoints. Each is about 300M parameters and speeds up decoding through speculative decoding without changing the output: 2.67x on an H100 for LFM2.5-2.6B.
- Semantica v0.6.6 — security release hardens ingestion, adds CrewAI support
Semantica v0.6.6 fixes a privately disclosed batch of vulnerabilities across backup, export, outbound requests and triplet-store backends, and hardens ingestion against SSRF. The same release adds CrewAI integration and graph retraction.
- Simon Willison — ChatGPT search now scopes one query in six to a single site
Simon Willison flags Promptwatch data showing ChatGPT's use of site:-scoped search queries jumped from 0.37% to 16.8% of all fanout queries on 8 August 2026, right after OpenAI updated GPT-5.6 Sol in ChatGPT.
- Apple Messages plugin — ChatGPT can read and send texts on a Mac
OpenAI's Apple Messages plugin lets ChatGPT read and search iMessage, SMS and RCS chats on a Mac and send replies through the Messages app. It runs in Codex and ChatGPT Work on the macOS desktop app, on all plans.
- Huzzah — an editor where you write pseudocode and an LLM fills in the code
Huzzah is an experimental editor where you write pseudocode in .hz files and an LLM turns it into real source code. Each edit sends only the pseudocode diff as the prompt, and the pseudocode stays on disk as a record of intent.
- Vomit — a local model rewrites Claude Code's replies before you read them
Vomit is a Go tool that hooks into Claude Code, buffers each reply and hands it to a local model to be rewritten in plainer language before it reaches the screen. Zach Ahn runs gpt-oss:20b for the rewriting pass.
- Grok leaks chat data — encrypted prompt injection slips past its filters
Adversa AI showed an attack that makes Grok send a user's name, rough location, plan and chat prompts to an attacker's server. The instructions sit on a web page as AES-256 ciphertext, so guardrails cannot read them. xAI has not patched it.
- Google Preferred Sources button — publishers can now embed it on their site
Google now gives publishers an embeddable Preferred Sources button. Readers who click it mark a site as a favorite, so its stories show more often in Top Stories and can carry a "preferred" badge in AI Mode and AI Overviews.
- Ramp Router — an LLM gateway that cut customer inference bills 40% on average
Ramp Router is an LLM gateway that takes one API call and sends it to the cheapest model that still meets your quality bar. Ramp says customers on it spend 40% less on inference on average. Routing is free through 2026.
- Fireship — 'DeepSeek just cooked again... Big AI is big scared'
Fireship uploaded 'DeepSeek just cooked again... Big AI is big scared' on 20 August 2026. The title points at DeepSeek; the video description leads on OpenAI, which it says paused its biggest training run in history.
- NeuroQuant PET — FDA clears automated amyloid scoring for dementia
Cortechs.ai's NeuroQuant PET received FDA 510(k) clearance on August 18, 2026. The software measures amyloid in each brain region from a PET scan and reports SUVRs and Centiloid scores for dementia workups.
- LFM2.5 Q4_0 — Liquid AI's 4-bit models keep 96%+ of full accuracy
Liquid AI released Q4_0 GGUF checkpoints for LFM2.5 at 230M, 350M, 1.2B and 2.6B. Quantization-aware distillation keeps 96.5-97.4% of the BF16 benchmark average and decodes 3-33% faster on phones, mini-PCs and a Raspberry Pi 5.
- Mistral Agentic Search — models search, open and grep their way through docs
Mistral Agentic Search gives models file-style tools — search, open, navigate, read, grep — so they dig through a document set over several turns instead of one retrieval pass. On FinanceBench, Mistral Medium 3.5 rose from 26.7% to 86%.
- RollTab — a 125M piano model that autocompletes on an iPhone
RollTab is a free iOS app that autocompletes piano playing from a MIDI keyboard. A 125M-parameter decoder-only transformer runs on the phone at about 108 notes per second on an iPhone 15, trained on roughly 300 million note events.
- Meta AI for Mac — a native desktop app with screen sharing and dictation
Meta AI now runs as a native Mac app. It can look at any window you attach to the chat, and a keyboard shortcut dictates speech into any app on the Mac. The 1.0 beta is free and needs an Apple Silicon Mac on macOS 15 or later.
- Binance Agent OS — an MCP server that lets AI agents place real trades
Binance Agent OS gives AI agents a standard way into Binance trading, market data, wallets and payments. Any MCP client over Streamable HTTP can connect, including Claude Code, Cursor, Codex and ChatGPT, with limits set per sub-account.
- ChatGPT Ads reach Europe — 31 markets get labeled ads on Free and Go
OpenAI is taking ChatGPT Ads into 31 European markets, the largest step yet for its advertising pilot. Only Free and Go users see ads; Plus, Pro and Enterprise stay ad-free, and Europeans choose personalized or contextual targeting.
- ai-memory v1.29.0 — long-term memory that follows agents across CLIs
ai-memory is an MIT-licensed Rust tool that keeps a git-versioned markdown wiki of what a coding agent learned, so you can stop Claude Code mid-task and continue in Codex or Cursor without re-explaining the project.
- Simon Willison — lines of code count again, but review capacity does not
Simon Willison argues that lines of code became a useful productivity measure once coding agents write them. The limit he hits is review, not typing: he can produce code 100 times faster but cannot hold 100 times more of it in his head.