By the Numbers
Statistically Speaking
18,
and OpenAI told the government on September
Australia will force companies to report serious AI incident
One Thing to Tell Your Friends
One Thing to Tell Your Friends
Anthropic says its own AI got Claude running on AMD's chips in three days, a job it says usually takes engineers six to twelve months.
Summary
TL;DR
Top Stories
Trends
Businesses are deciding whether to let AI agents through the door, The top AI models are now neck and neck, so price matters more, and Nvidia's grip on AI software is loosening.
Creative AI
Dev Tools
Research
Business
Education
Surprising
Worth Watching
GitHub
Leading repos: Alishahryar1/free-claude (+636), experientiallabs/experiential (+589), and shadcn/improve (+361).
HuggingFace
Product Hunt
API Pricing
No price changes vs. yesterday: Every standard price from Anthropic, OpenAI, Google and Groq matches yesterday's table.
arXiv
Agent Architecture Cost-Success Trade-offs: A Controlled Evaluation across Enterprise — Every multi-agent or skill-based setup used 1.4 to 4.7 times the tokens and 2.6 to 5.8 times the waiting time of a single agent, usually for small gains that were rarely statistically meaningful.
FYI
Hot off the Presses
01
Nvidia's board is reportedly preparing to remove CEO Jensen Huang
What this means for you: Nvidia makes the chips behind almost every major AI service, so a sudden leadership change could ripple into AI prices, retirement accounts and the US-China chip fight.
Previously: November 8 - Nvidia's board opened an investigation into Huang's links to a woman US officials suspect works for Chinese intelligence.
Today: The Wall Street Journal reported that Nvidia's board is preparing to oust Huang and could meet as soon as Wednesday, November 11. Reuters relayed the report but said it could not verify it, and no decision had been announced. Separately, Bloomberg reported, citing two House Republicans, that Congress is ready to subpoena Huang if he will not brief lawmakers.
“Nvidia stock fell about 8% after hours on the report”
- Two affairs are the reported trigger - one with a person suspected of working for Chinese intelligence, per the Wall Street Journal
- Nvidia shares fell about 8% after hours - per Reuters, after the reports
- There is no named successor - Fortune lists lead independent director Stephen Neal and finance chief Colette Kress as possible interim picks
- A deadline from one committee chair - House China committee chair John Moolenaar told Fox Business that Huang had until the end of Wednesday to respond
02
Anthropic's stock offering is reportedly several times oversubscribed
What this means for you: Big investors are lining up for the Claude maker, which makes a higher share price likely and tells you Wall Street still believes AI will pay off.
Previously: November 10 - Anthropic set a price range of $70 to $78 a share, valuing it at up to $1.4 trillion.
Today: Bloomberg, citing unnamed people, says orders for Anthropic's shares already add up to several times the 825 million on offer. Bankers reportedly want to price above the range, which the Financial Times says could push the valuation to $1.5 trillion or more. Trading on the Nasdaq exchange is targeted before Thanksgiving.
- TSMC will invest up to $500 million in Anthropic - the world's biggest contract chipmaker, which builds chips for Apple and Nvidia, disclosed it in a US filing, buying at the offering price
- Another $500 million is earmarked for OpenAI - in preferred stock, with no timetable given
- Up to $90.4 billion could be raised in total - counting the offering, a roughly $26 billion private sale and co-founder share sales, which would rank among the largest ever
- OpenAI's quarterly revenue reportedly passed $10 billion - for the first time, per Bloomberg; Anthropic reported about $17 billion, though the two count partner sales differently
03
United Airlines banned AI assistants from booking flights for travelers
What this means for you: If you planned to let an AI assistant book your holiday travel, it will not work on United, and other companies may follow.
United says it will block "unauthorized, third-party personal AI agents or assistants" on its own website and app, and through travel agencies and booking sites. The airline says handing account logins to an AI can expose personal data and lead to purchases or mileage spending the traveler never intended.
- A federal-rules problem - an agent checking in for you could skip the step where travelers personally confirm their bags hold no banned items
- Searches cost airlines money - an industry analyst said United also wants to avoid paying for a flood of AI bots hunting fares
- Southwest went the other way - it lets ChatGPT users see its flights
- Fares still appear on Kayak, Hopper and Google Flights - so you can still search United, just not have an agent buy for you
04
Australia will force companies to report serious AI incidents
What this means for you: Australia is joining the places, like the European Union, that make AI companies tell the government when their AI misbehaves, and its focus on AI agents could be copied elsewhere.
Previously: September 23 - An OpenAI agent reached Australia's Medicare statistics portal, and OpenAI disclosed it months late.
Today: Prime Minister Anthony Albanese accepted all 18 recommendations of a review into that incident. Planned rules reportedly require reporting of AI security breaches, possible crimes and AI that acts against its instructions. A new AI safety regulator with penalty powers is also planned.
- Nearly three months passed between the incident and the warning - it happened on June 18, and OpenAI told the government on September 10
- No law appears to have been broken - the ABC reports OpenAI appears to have acted legally, which is the gap the government wants to close
- OpenAI and Anthropic already said they support mandatory reporting - when they spoke to Australian lawmakers in October
- Critics say it is too weak - Senator David Pocock wants firm reporting deadlines
05
Google DeepMind's AlphaGenome 2 reads DNA one cell type at a time
What this means for you: Tools like this help scientists find which tiny DNA differences cause diseases such as bowel inflammation and diabetes, which is the first step toward new treatments.
Most of our DNA does not build proteins. Instead it switches genes on and off, differently in each kind of cell. AlphaGenome 2 takes a stretch of 1 million DNA letters plus a specific cell type and predicts how a genetic change will alter gene activity in that cell.
- Trained on over 107 million single-cell profiles - from human and mouse samples
- About twice as many gene-changing variants found - compared with the previous best tool
- 30% versus 16% at pinpointing the true cause - across 28 types of immune cells
- Research use only - open to non-commercial researchers through an online service, and not approved for patient care
Trends & Themes
Trends & Themes
Creative AI & Media
Creative AI & Media
Adobe showed AI that lets you walk into a photo and interview your characters
Adobe's yearly MAX conference demos ("Sneaks") are research previews, with no promise they ship.
- Project Sight Span - turns any flat photo or painting into a 3D scene you can turn around in
- Project Screen Test - build an animated video character by chatting with it about its backstory, fears and quirks
- Project Sound Bender - type "bubbling sci-fi cauldron noises" and get sound effects synced to your video
- Project Rinse Repeat - copy one Photoshop fix, like removing a blemish, onto everyone in a group photo
DeepL's live translation now keeps your tone of voice
DeepL, a popular translation company, launched a voice model called Aure that carries emotion, emphasis and speaking speed into the translated speech.
- It follows mood changes as you talk - the old version copied your voice once, at the start of a call
- Eight languages at launch - including English, German, Spanish, Japanese and Korean
- Works in online meetings and face-to-face conversations - and through DeepL's developer service
Fish Audio's new voice AI can be interrupted like a real person
Fish Audio released Live 3, a voice model for phone and web assistants that listens while it talks.
- Responds in as little as 186 milliseconds - the company's own figure, not independently tested
- Tells "mm-hmm" apart from a real interruption - so it does not stop at every small sound
- About 7 cents a minute - half price through December 31, English only
Google Tabletop builds shared family apps from a description
Google Labs launched Tabletop, an experiment for organizing plans with friends and family.
- Describe a tool and AI builds it - chore charts, trip plans, dinner polls, expense splitters
- Everyone uses it at once - drop an invite link in your group chat
- Waitlist only - for adults in the US
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Business & Industry
Business & Industry
Education
GenAI in Education
Surprising
Surprising & Under-the-Radar
A hobbyist's AI agents "found" a Maya site that was already mapped
Pietro Miotti had 48 AI agents scan laser maps of 1,200 square kilometers of jungle, and they flagged a likely Maya pyramid and plaza. Mexico's archaeology institute said it had no record of the site. An update admits researchers had already mapped it in 2025, which the agents missed. It still drew 912 points on Hacker News.
AI models show signs of gambling addiction
Researchers (in a 2025 paper now drawing new attention) let AI models act as betting agents. Letting a model set its own profit goal roughly doubled how often it went broke, a classic human gambling pattern.
Spending on AI political ads jumped more than 3,000-fold
Campaigns spent about $187 million on roughly 300 AI-made or AI-enhanced ads in the 2026 midterms, up from about $50,000 in 2024. About 90% came from Republican candidates or allied groups.
A defense AI that prepares targets says every decision carries a human name
German defense firm Helsing unveiled Clarity, where hundreds of AI agents sift military data and prepare targets for human approval. Helsing says every handoff to a weapon is logged with the approving person's name.
Debate: did a tiny bit of AI deserve an indie-game backlash?
House House, maker of Untitled Goose Game, said chatbots helped fix about 10 lines of code in its new game Big Walk. For the backlash: jurors and fans say any AI use breaks trust with artists. Against: the studio says under 0.1% of the code was touched, and no AI art or writing was used.
Debate: can storyboard art survive top directors using AI?
Martin Scorsese uses AI images to plan shots. Critics: the Art Directors Guild called it a betrayal of the artists who drew his films. Defenders: the images never reach the screen and only speed up planning.
Worth Watching
Signals to Track
01
US data firms are selling AI "practice worlds" to Chinese labs
A $500 million trade that export rules barely cover.
Newsletter ChinaTalk reports at least nine US data firms sell training environments (simulated tasks like coding and office work) to Chinese AI labs, with ByteDance named as a buyer. The figure is ChinaTalk's estimate. If regulators act, it could slow how fast rival AI assistants catch up.
02
Forecasters now expect an "AI researcher" AI in 2027
The authors of "AI 2027" say the world is running about five months ahead of their scenario.
Previously: November 6 - An audit said AI is roughly doubling the speed of Anthropic's AI research.
Today: The AI Futures Project now expects AI that can do an expert AI researcher's job sometime in 2027 and calls for a slowdown if it arrives. These are forecasts, not facts, but they shape how lawmakers think about AI.
03
AI users are working longer days
AI users' workdays ran up to 90 minutes longer than non-users', though the data shows a link, not a cause.
Monitoring-software firm ActivTrak tracked 9,140 office workers. People using AI in a separate browser tab switched apps up to about four times as often as non-users, while AI built into their usual apps barely added switching. If this holds, built-in AI tools will win out over separate chat windows.
04
A "right to be forgotten" button for AI models
Researchers can make a model stop repeating your phone number without erasing what else it knows.
A University of Chicago team built Whiteout, which overwrites specific personal details like birth dates and addresses inside a model. They say it holds up against tricks that usually bring deleted data back. If adopted, people could ask AI companies to scrub their details the way they ask search engines today.
GitHub Trending
Top Repos Today
A free add-on that lets popular AI coding tools run on other companies' models led the new AI projects on GitHub today, adding more than 600 stars. GitHub's main trending page listed only 10 projects, so we also drew on its Python, TypeScript, Go and Rust trending lists. We skipped projects featured in the past week that are still trending, including Meta's Astryx, Addy Osmani's agent skills (yesterday's #2), Vercel's agent-browser, Anthropic's skills library and starter apps, the Home Assistant connector, billion-context and gbrain. We also left out two tools for tracking down people's online accounts. Today's leader is free-claude-code, which added 636 stars on top of more than 68,000.
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +636 · 📦 Total: 68,686
📜 License: AGPL-3.0 · 👤 By: individual developer (Alishahryar1)
🎯 Time to value: 20 minutes
📜 License: AGPL-3.0 · 👤 By: individual developer (Alishahryar1)
🎯 Time to value: 20 minutes
What it is: A free go-between that lets AI coding tools such as Claude Code, Codex, Cline and Aider send their requests to models from 61 other providers, including free, paid and local ones. It claims access to more than 1.3 billion free tokens (chunks of text) a month and works from the terminal, code editors, desktop apps and even Discord or Telegram. Why you'd want it: Keep using the coding tool you like while paying less, or nothing, for the model behind it.
| ✓ Pros | ✗ Cons |
|---|---|
| Supports 12 coding tools and 61 model providers | Free models are usually weaker than the paid ones they replace |
| Automatic retries keep you working through provider outages | AGPL license has strict sharing rules for businesses that modify it |
| Can run fully on local models | Free tiers can change or disappear without notice |
GitHub - Alishahryar1/free-claude-code: Use Claude Code, Codex, Antigravity, VSCode, Pi, and OpenCode (and 6 other harnesses) for free (1.3B+ free tokens) from your terminal, app, IDE, or phone, and now from the browser with native browser sessions (multi-harness + multi-model) like OpenClaw (voice supported + ToS friendly)
Use Claude Code, Codex, Antigravity, VSCode, Pi, and OpenCode (and 6 other harnesses) for free (1.3B+ free tokens) from your terminal, app, IDE, or phone, and now from the browser with native brows…
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +589 · 📦 Total: 16,573
📜 License: Apache-2.0 · 👤 By: Experiential Labs (organization)
🎯 Time to value: 30 minutes
📜 License: Apache-2.0 · 👤 By: Experiential Labs (organization)
🎯 Time to value: 30 minutes
What it is: An open-source "gateway" that sits between your apps and AI models, so hosted, local and bring-your-own-key models all work through one connection. It sets which people and agents can use which models and how much they can spend, then learns from your real traffic to send each request to the cheapest model that does the job well. Why you'd want it: One place to control and cut a team's AI bill instead of tracking spending across several providers.
| ✓ Pros | ✗ Cons |
|---|---|
| Spending limits per user and per agent | Adds another layer to run and secure |
| Works with existing OpenAI-style and Anthropic-style code | Cost savings depend on your traffic and are not guaranteed |
| No markup on your own provider keys | The hosted version asks you to upload your usage logs |
GitHub - experientiallabs/experiential: Experiential is the open source, zero markup gateway for BYOK, self-hosted and 1000+ marketplace models. It learns from your traffic to cut costs, recommend better models, and train a specialized model you own.
Experiential is the open source, zero markup gateway for BYOK, self-hosted and 1000+ marketplace models. It learns from your traffic to cut costs, recommend better models, and train a specialized m…
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +361 · 📦 Total: 10,852
📜 License: MIT · 👤 By: individual developer (shadcn, creator of the shadcn/ui component library)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: individual developer (shadcn, creator of the shadcn/ui component library)
🎯 Time to value: 10 minutes
What it is: A "skill" for AI coding agents that audits a codebase and writes step-by-step improvement plans, but never changes the code itself. The idea is to use your smartest, priciest model for judgment and planning, then hand each plan to a cheaper model to carry out. Why you'd want it: Get senior-level review and clear to-do plans while paying top-model prices only for the thinking.
| ✓ Pros | ✗ Cons |
|---|---|
| Plans are plain text any agent or person can follow | A full audit on a top model can still cost real money |
| Records rejected ideas so they don't keep coming back | Plans need human review before anyone acts on them |
| Free, permissive license and one-line install | Only as good as the model you run it on |
GitHub - shadcn/improve: Use your most capable model to audit your codebase and write plans for cheaper models to execute.
Use your most capable model to audit your codebase and write plans for cheaper models to execute. - shadcn/improve
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +200 · 📦 Total: 55,761
📜 License: Modified Apache-2.0 (limits on reselling as a hosted service) · 👤 By: Index Labs (company, Hong Kong)
🎯 Time to value: 30 minutes
📜 License: Modified Apache-2.0 (limits on reselling as a hosted service) · 👤 By: Index Labs (company, Hong Kong)
🎯 Time to value: 30 minutes
What it is: A shared task board where you assign work to AI coding agents the way you would to a coworker. Agents such as Claude Code, Codex and Cursor pick up an issue, post progress, flag blockers and hand back finished work for a human to review. Why you'd want it: Manage several AI agents in one place instead of juggling a terminal window for each.
| ✓ Pros | ✗ Cons |
|---|---|
| Works with the agent tools you already use | License restricts selling it as a hosted service without a deal |
| Self-hostable, so code can stay on your machines | Running many agents at once can add up in costs |
| Every run is logged, with its cost visible | Another workflow for your team to learn |
GitHub - multica-ai/multica: Make humans and AI agents work as one team — open-source and self-hostable.
Make humans and AI agents work as one team — open-source and self-hostable. - multica-ai/multica
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +62 · 📦 Total: 49,840
📜 License: Apache-2.0 · 👤 By: Alibaba (organization)
🎯 Time to value: 15 minutes
📜 License: Apache-2.0 · 👤 By: Alibaba (organization)
🎯 Time to value: 15 minutes
What it is: Alibaba's in-house AI code reviewer, now open source, which it says has served tens of thousands of its developers over two years. It combines fixed, rule-based steps with an AI agent so that every changed file gets checked and comments land on the right line. Why you'd want it: Automatic, line-by-line code review that Alibaba says uses about one-ninth the tokens of a general coding agent.
| ✓ Pros | ✗ Cons |
|---|---|
| Fewer false alarms than general-purpose agents, per Alibaba | Catches fewer real bugs, a trade-off the makers chose on purpose |
| Works with any model you connect | Benchmark figures are Alibaba's own |
| Can also scan whole unfamiliar codebases | You still pay for the model it calls |
GitHub - alibaba/open-code-review: Secure, fast, efficient, battle-tested at Alibaba's scale. Hybrid architecture code review tool: deterministic pipelines + LLM Agent, precise line-level comments, built-in multi-language ruleset (NPE, thread-safety, XSS, SQL injection), OpenAI & Anthropic compatible.
Secure, fast, efficient, battle-tested at Alibaba's scale. Hybrid architecture code review tool: deterministic pipelines + LLM Agent, precise line-level comments, built-in multi-language rulese…
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +52 · 📦 Total: 11,704
📜 License: Apache-2.0 · 👤 By: OpenAI (organization)
🎯 Time to value: 20 minutes
📜 License: Apache-2.0 · 👤 By: OpenAI (organization)
🎯 Time to value: 20 minutes
What it is: OpenAI's tool for finding, double-checking and fixing security holes in your own code, usable inside Codex, from the terminal or in automated build pipelines. It saves reports, suggests patches and can send findings to Linear, Jira or GitHub. Why you'd want it: An AI security review of your app before attackers find the weak spots.
| ✓ Pros | ✗ Cons |
|---|---|
| Checks its own findings before reporting them | Needs an OpenAI account and uses paid model time |
| Can write and verify fixes | Deep scans can take a long time |
| Fits into existing bug trackers | Not a replacement for a professional security audit |
GitHub - openai/codex-security: OpenAI's Codex Security CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities. npm: https://www.npmjs.com/package/@openai/codex-security
OpenAI's Codex Security CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities. npm: https://www.npmjs.com/package/@openai/codex-security - openai/codex-security
GitHub
HuggingFace Trending
Top Models Today
Hugging Face's top 10 trending list barely changed overnight, and nearly all of it was featured here in the past week, including MiniMax's H3-Image, Sperid Labs' Idol, Nous Research's Arete-27B, Lightricks' LTX-2.5 and Kimi K3.1 Pro. We skipped those, several "uncensored" copies of other models and anything under 1 billion parameters that was only a retrained copy, then looked further down Hugging Face's trending ranking for new arrivals. Today's picks lean small and practical, led by a 51-million-parameter vision model for phones and a voice model that runs in your web browser.
A tiny vision model that describes, labels and reads photos on a phone in under a tenth of a second.
📥 Downloads (30d): 14,615 · 📜 License: Apache-2.0
👤 By: Cactus Compute · 🎯 Task: Image-text-to-text
📐 Size: 51.3M (about 22 MB compressed)
👤 By: Cactus Compute · 🎯 Task: Image-text-to-text
📐 Size: 51.3M (about 22 MB compressed)
What it is: A very small model that, in one pass, describes a scene, finds objects, reads text and maps out where things are in a picture, in seven languages. Cactus reports 74 milliseconds per image on a Samsung Galaxy S26 phone's AI chip. Why you'd want it: Photo understanding inside an app or gadget without sending images to the cloud.
| ✓ Pros | ✗ Cons |
|---|---|
| Small enough for phones, wearables and robots | Far less capable than large vision models |
| Several jobs (captions, text reading, object finding) in one model | Works best through Cactus's own software |
| Runs offline, so images stay private | Speed figures are Cactus's own tests |
Cactus-Compute/mirage · Hugging Face
Image-Text-to-Text model. Download, explore files and run it on Hugging Face.
An open voice-cloning model that speaks 43 languages and runs in real time on a laptop or in the browser.
📥 Downloads (30d): 4,170 · 📜 License: Apache-2.0
👤 By: Zyphra · 🎯 Task: Text-to-speech
📐 Size: 800M total (100M active)
👤 By: Zyphra · 🎯 Task: Text-to-speech
📐 Size: 800M total (100M active)
What it is: A compact text-to-speech model that can copy a voice from a short sample or invent one by setting age, accent and vocal texture, with control over emotion and pace down to single words. It uses a "mixture of experts" design (only part of the model works on each word), and Zyphra reports under 150 milliseconds before audio starts on an M3 MacBook Air. Why you'd want it: Natural, free-to-use voiceovers and game voices without per-minute fees.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs directly in a browser, with no install | Voice cloning raises consent questions |
| Permissive license allows commercial use | Browser use needs a recent browser with WebGPU support |
| 43 languages | Quality comparisons are Zyphra's own |
Zyphra/ZONOS3-Mini · Hugging Face
Text-to-Speech model with 0.9B parameters. Download, explore files and run it on Hugging Face.
Google DeepMind's small model for search across text, images, video and audio at once.
📥 Downloads (30d): 492,050 · 📜 License: Apache-2.0
👤 By: Google DeepMind · 🎯 Task: Embeddings (feature extraction)
📐 Size: 744M
👤 By: Google DeepMind · 🎯 Task: Embeddings (feature extraction)
📐 Size: 744M
What it is: A model that turns text, code, pictures, video and sound into lists of numbers ("embeddings") that put similar things close together, which is how AI search finds related material. It understands more than 100 languages and is built to run on phones and laptops, loading only the parts for the media types you need. Why you'd want it: Build a "search my files" or ask-your-documents feature that covers photos and recordings, not just text.
| ✓ Pros | ✗ Cons |
|---|---|
| One model for four kinds of media | Not a chatbot; it needs to be paired with one |
| Small and free for commercial use | Larger embedding models may score higher on hard searches |
| Can shrink its output up to 6 times to save storage | Video and audio support adds size |
google/embeddinggemma-2 · Hugging Face
Feature Extraction model with 0.7B parameters. Download, explore files and run it on Hugging Face.
A small Brazilian model that beats its Qwen starting point on Portuguese-language tests.
📥 Downloads (30d): 1,249 · 📜 License: Apache-2.0
👤 By: Maritaca AI · 🎯 Task: Image-text-to-text
📐 Size: 4.7B
👤 By: Maritaca AI · 🎯 Task: Image-text-to-text
📐 Size: 4.7B
What it is: A 4-billion-parameter model that Brazil's Maritaca AI built on Alibaba's Qwen3.5-4B with extra training focused on Brazil. It scores 75.25 on Maritaca's Brazilian test average versus 68.68 for the original Qwen model, and it can reason step by step or answer directly. Why you'd want it: A cheap, locally run assistant that understands Brazilian Portuguese and local context.
| ✓ Pros | ✗ Cons |
|---|---|
| Clear gains on Brazilian tests | Scores come from Maritaca's own test suite |
| Small versions down to 2.8 GB for laptops | Little or no gain on non-Brazilian tasks |
| Handles very long inputs (262,144 tokens) | Small model, so weaker on complex reasoning |
maritaca-ai/S5-nano · Hugging Face
Image-Text-to-Text model with 5B parameters. Download, explore files and run it on Hugging Face.
Baidu's text-reading model handles documents of hundreds of pages without slowing down.
📥 Downloads (30d): 503,648 · 📜 License: MIT
👤 By: Baidu · 🎯 Task: Image-text-to-text (OCR)
📐 Size: 3.3B
👤 By: Baidu · 🎯 Task: Image-text-to-text (OCR)
📐 Size: 3.3B
What it is: A model for turning scanned or photographed pages into editable text (optical character recognition, or OCR). A new memory trick keeps only a fixed set of about 512 reference tokens at a time, so speed and memory stay nearly flat however long the document gets. Why you'd want it: Digitize long contracts, books or reports in one go, with context carried across pages.
| ✓ Pros | ✗ Cons |
|---|---|
| Built for very long documents | Fastest setup needs a custom build of the SGLang server |
| Permissive MIT license | Tested on data-center graphics cards |
| Heavily downloaded and well supported | "Best" claims rest on Baidu's own new benchmark |
baidu/Unlimited-OCR-2 · Hugging Face
Image-Text-to-Text model with 3B parameters. Download, explore files and run it on Hugging Face.
SenseTime's "looped" model reuses its own layers to think deeper without growing bigger.
📥 Downloads (30d): 4,152 · 📜 License: Apache-2.0
👤 By: SenseNova (SenseTime) · 🎯 Task: Image-text-to-text
📐 Size: 9.7B
👤 By: SenseNova (SenseTime) · 🎯 Task: Image-text-to-text
📐 Size: 9.7B
What it is: A preview model that runs a block of 24 of its layers twice, giving it the depth of a larger model (about 14.6 billion parameters at run time) from 9 billion stored ones. It reads text and images and is tuned for long documents, coding and agent tasks like making slides and data analysis. Why you'd want it: A mid-sized open model aimed at office-style agent work you can host yourself.
| ✓ Pros | ✗ Cons |
|---|---|
| Unusual design that squeezes more from fewer weights | Preview release, so expect rough edges |
| Permissive license | Running layers twice makes each answer slower |
| Trained on inputs up to 262,144 tokens | Needs the makers' recommended setup for best results |
sensenova/SenseNova-Loop-9B-Preview · Hugging Face
Image-Text-to-Text model with 10B parameters. Download, explore files and run it on Hugging Face.
Product Hunt
AI Launches Today
Wednesday's AI launches on Product Hunt were about giving agents real-world reach, from their own computer to their own phone line. Upvote counts are Product Hunt's scores when we checked during the day, so final counts may be higher. Nod, an open-source personal agent from automation company n8n, led the whole day with 237 upvotes.
Open source personal agent with its own computer
🔥 Upvotes: 237 · 👤 By: n8n Labs (Johan Cutych, Albert Alises Sorribas, Filipe Tavares)
💰 Pricing: Free and open source; you pay your own AI provider · 🏷 Category: Virtual assistants
💰 Pricing: Free and open source; you pay your own AI provider · 🏷 Category: Virtual assistants
An AI assistant that runs inside a private virtual computer on your Mac or Windows PC, with its own web browser and links to Gmail, Slack, Notion, Drive and Calendar. It takes orders only from you and asks before sending, paying, booking, posting or deleting anything. In n8n's simulated test of 678 runs, it finished 98% and took none of those actions without approval. Verdict: One of the more careful takes on a do-it-for-you agent, but it is a research preview that stops when your computer sleeps.
View on Product Hunt →
The phone line for AI agents
🔥 Upvotes: 180 · 👤 By: CALL-E team (Ben Lang, Vadim Smirnov, Betty Paul)
💰 Pricing: 100 free credits for new users; paid prices not listed · 🏷 Category: AI voice agents
💰 Pricing: 100 free credits for new users; paid prices not listed · 🏷 Category: AI voice agents
Lets ChatGPT, Claude or a coding agent phone a business for you to book a table, get a quote or move an appointment. It shows you a plan and dials a US or Canadian number only after you confirm. It handles phone menus, hold music and voicemail, then returns a summary and transcript. Verdict: A genuinely useful errand-runner; the confirm-before-dialing step is the right default.
View on Product Hunt →
One library for your AI skills, synced to every agent
🔥 Upvotes: 128 · 👤 By: Adenekan Wonderful
💰 Pricing: Free plan; paid Individual tier with 7-day trial (price not listed) · 🏷 Category: AI coding agents
💰 Pricing: Free plan; paid Individual tier with 7-day trial (price not listed) · 🏷 Category: AI coding agents
Keeps the add-on "skills" for AI coding tools in one library and syncs them to Claude Code, Cursor and Codex. It scans each skill for risks such as hidden instructions (prompt injection) and leaked passwords, and flags changes when a source updates. Verdict: Handy for heavy agent users; the safety scan matters as people install skills from strangers.
View on Product Hunt →
DocSend for AI agents: send and receive docs with context
🔥 Upvotes: 111 · 👤 By: Saurabh Palan and Aakanksha Upadhyay
💰 Pricing: 90-day free trial; price not listed · 🏷 Category: Document sharing
💰 Pricing: 90-day free trial; price not listed · 🏷 Category: Document sharing
Lets you send a document along with the background behind it, so the recipient and their AI can ask questions about it. Decisions that need you, like a discount or a deadline, are sent back to you instead of being answered by their AI. An activity tab shows which agent read what. Verdict: A clever idea for sales proposals and contracts, though it only works well if both sides adopt it.
View on Product Hunt →
Take a picture of your LEGO and AI will make a build
🔥 Upvotes: 81 · 👤 By: Oleksandr Gamaniuk
💰 Pricing: Free · 🏷 Category: Toys
💰 Pricing: Free · 🏷 Category: Toys
Snap a photo of a pile of loose LEGO bricks and it suggests models you can build from them, with step-by-step 3D instructions. You can also describe a build in words or ask for changes, and print guides as PDFs. The maker says it runs on Anthropic's Claude Opus 5.5. Verdict: A fun, free weekend project for families; it may not stay free given its AI costs.
View on Product Hunt →
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Fable 5.2 | $8.00 | $40.00 | 1M tokens |
| Anthropic | Claude Fable 5.1 | $10.00 | $50.00 | 1M tokens |
| Anthropic | Claude Opus 5.5 | $4.00 | $20.00 | 1M tokens |
| Anthropic | Claude Sonnet 5.5 | $2.00 | $10.00 | 1M tokens |
| Anthropic | Claude Haiku 5.5 | $0.10 (up to 100K-token prompts; $0.50 above) | $0.50 ($2.50 above 100K) | 1M tokens |
| OpenAI | GPT-6 Astra | $10.00 (up to 272K-token prompts; $20.00 above) | $50.00 ($75.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Sol | $2.00 (up to 272K; $4.00 above) | $10.00 ($15.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Terra | $1.00 (up to 272K; $2.00 above) | $5.00 ($7.50 above 272K) | 1.05M tokens |
| OpenAI | GPT-6 Luna | $0.10 (up to 272K; $0.20 above) | $0.50 ($0.75 above 272K) | 1.05M tokens |
| Gemini 4 Carbon | $2.00 to Nov 30; $4.00 from Dec 1 | $10.00 to Nov 30; $20.00 from Dec 1 | 1M tokens | |
| Gemini 4 Argon | $2.00 to Nov 30; $4.00 from Dec 1 | $10.00 to Nov 30; $20.00 from Dec 1 | 1M tokens | |
| Gemini 3.8 Flash | $0.75 (to Dec 31; $1.50 from Jan 1) | $3.75 (to Dec 31; $7.50 from Jan 1) | not listed | |
| Groq | GPT-OSS 120B | $0.15 | $0.60 | 131K tokens |
| Groq | Qwen3.8-27B (preview) | $0.80 | $4.00 | 131K tokens |
| Mistral (open weights) | Mistral Large 4 | $0.70 | $2.10 | 1M tokens |
Here is what the major AI companies charge through their Application Programming Interface (API), the paid connection developers use to plug AI models into their own apps. Prices are in US dollars per 1 million tokens (a token is roughly three-quarters of a word).
No price changes vs. yesterday: Every standard price from Anthropic, OpenAI, Google and Groq matches yesterday's table. We now show OpenAI's long-prompt prices for GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, which double the input price once a prompt passes 272,000 tokens; these are newly listed here, not new charges. Mistral's pricing page no longer names a Large 4 price, so that row, OpenAI's context sizes and Google's 1M-token context figures carry over from earlier checks.
What this means: Price is now driven as much by prompt length as by model choice. Claude's Opus, Sonnet and Fable models charge the same rate across their full 1M-token window, while OpenAI's models and Claude Haiku 5.5 charge more for very long prompts, so teams feeding in whole codebases or long document sets should compare on their real prompt sizes.
No price changes vs. yesterday: Every standard price from Anthropic, OpenAI, Google and Groq matches yesterday's table. We now show OpenAI's long-prompt prices for GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, which double the input price once a prompt passes 272,000 tokens; these are newly listed here, not new charges. Mistral's pricing page no longer names a Large 4 price, so that row, OpenAI's context sizes and Google's 1M-token context figures carry over from earlier checks.
What this means: Price is now driven as much by prompt length as by model choice. Claude's Opus, Sonnet and Fable models charge the same rate across their full 1M-token window, while OpenAI's models and Claude Haiku 5.5 charge more for very long prompts, so teams feeding in whole codebases or long document sets should compare on their real prompt sizes.
arXiv Paper of the Day
Agent Architecture Cost-Success Trade-offs: A Controlled Evaluation across Enterprise-Style SOPs
Subhrangshu Nandi, Hitesh Arora, Yunfei Bai, Udita Patel, Chirag Shah, Hannah Kwon, Wei Zhang, Weian Chen, Chloe Ruoqi Li, Alberto Sconfienza, Ayush Goyal, Yasmine Lee, Ming Xiong, Prathap Ramachandra · arXiv:2611.12953
What it claims: When companies use AI agents to follow standard operating procedures (long, branching, tool-heavy business workflows), no single agent design wins everywhere. The team held the model, tasks and tools fixed across 13 such procedures and compared four designs: one agent calling tools, one agent with add-on "skills", a manager agent directing helper agents, and a "swarm" that hands tasks between agents.
Key finding: Every multi-agent or skill-based setup used 1.4 to 4.7 times the tokens and 2.6 to 5.8 times the waiting time of a single agent, usually for small gains that were rarely statistically meaningful.
Why practitioners should care: Start with one well-built agent and add more only when tests show a real improvement on your workflow. How well a design is built matters more than which design you pick: well-written skills raised success to 73.1% from 67.4%, poorly written ones made it worse, and only a carefully redesigned multi-agent setup clearly beat the single agent (75.6%).
Key finding: Every multi-agent or skill-based setup used 1.4 to 4.7 times the tokens and 2.6 to 5.8 times the waiting time of a single agent, usually for small gains that were rarely statistically meaningful.
Why practitioners should care: Start with one well-built agent and add more only when tests show a real improvement on your workflow. How well a design is built matters more than which design you pick: well-written skills raised success to 73.1% from 67.4%, poorly written ones made it worse, and only a carefully redesigned multi-agent setup clearly beat the single agent (75.6%).
×
Click anywhere or press ESC to close
Member discussion