Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
$135
billion to $145 billion of spending this
Meta is borrowing up to $25 billion more to pay for AI
Top Story
One Thing to Tell Your Friends
One Thing to Tell Your Friends
A big British bank app cut its AI coding bill by about 90% by dropping Claude and running a free Chinese AI model on its own computers.
Summary
TL;DR
Top Stories
Trends
Creative AI
Dev Tools
Research
Business
Education
Surprising
An AI scored higher with no evidence than with it, An AI coder assumed every currency has 100 cents, and A heavy AI user replaced 4,000 lines of AI code with a few hundred of his own.
Worth Watching
GitHub
HuggingFace
Leading models: deepseek-ai/DeepSeek-V4.1-Flash (187,260), divagr1925/Qwen3.5-4B (1,229), and nvidia/NVIDIA-NemotronLabs-RE (31,724).
Product Hunt
API Pricing
No price changes vs. yesterday - every standard price from Anthropic, OpenAI, Google, Groq and Mistral matches yesterday's table.
arXiv
MEMENTO — On questions about books of up to 2.5 million tokens, accuracy rose from 36.9% to 67.8% while never holding more than 20,000 tokens at once, beating a GPT-5.5 agent (13.1%) at under 1/50 of the cost.
FYI
Hot off the Presses
01
Meta is borrowing up to $25 billion more to pay for AI
What this means for you: The companies behind your apps are now borrowing heavily to build AI. If those bets disappoint, it can hit stock markets, retirement savings and the price of their products.
Meta, the owner of Facebook, Instagram and WhatsApp, started a bond sale (a way of borrowing money from investors) worth $20 billion to $25 billion. About two-thirds is in US dollars and one-third in euros. The money pays for AI data centers, models and services.
The sale came days after S&P Global Ratings, a firm that grades how safely companies can repay debt, changed Meta's outlook to "negative" (covered October 31). It kept Meta's high AA- grade but warned that rising spending and negative cash flow could change that.
- Meta plans $135 billion to $145 billion of spending this year - mostly on AI computers and buildings
- It spent $2.12 billion more cash than it brought in last quarter - Alphabet and Amazon also reported negative free cash flow
- Meta already borrowed $25 billion in bonds in April - and announced a $95 billion share-sale plan last week
- S&P expects Meta's yearly spending to pass $225 billion in 2027 - and its debt measure could cross the line for its current grade that year
- Meta's shares rose 2.7% on the day - after falling in 12 of the previous 14 trading days
02
CrowdStrike built a "kill switch" for runaway AI agents
What this means for you: AI agents (programs that take actions on their own, like sending email or changing files) are now common enough at work that security companies sell emergency brakes for them.
CrowdStrike, one of the biggest cybersecurity companies, launched an "AI kill switch" for its Falcon Guardian product. It spots internal AI agents behaving badly, cuts their network access and stops them from restarting. A human decides when to pull the switch and when to turn the agent back on.
CrowdStrike also plans to buy two startups for a combined $495 million or so, according to The Wall Street Journal. It announced a connection to Anthropic's Claude Cowork as well.
- About $400 million for Fiddler AI - which makes software for monitoring what companies' AI agents do
- About $95 million for Dreadnode - which tests AI models and agents for security weaknesses before they launch
- The kill switch does not delete the agent - it cuts off access to the AI models that drive the agent's decisions
- Research firm Gartner expects AI security spending to reach almost $5 billion in 2027 - and rivals such as HiddenLayer and Zenity are raising money
03
Developers say Postman quietly installed AI agent add-ons on their computers
What this means for you: Software you already trust can now change how your AI assistants behave without asking. Check what your apps install when they update.
Postman is a widely used app for testing the connections between pieces of software. A developer reported that opening Postman's desktop app (version 12.33.4) triggered an automatic update that installed add-ons for OpenAI's Codex and Anthropic's Claude Code, plus Postman's own command-line tool, without a clear opt-in. A hook added at the start of each session tells AI coding agents to load Postman's instructions for any non-trivial coding task that involves software connections.
The report became one of the top stories on Hacker News, a popular tech discussion site. The Hacker News headline claims Postman tries to share users' data, but the original reporter says the report does not show that.
- 971 points and 320 comments on Hacker News - a very strong reaction for a support ticket
- The installer also tried to mark its own startup hook as trusted - in the user's Codex settings, according to the report
- A reply posted for Postman says the add-ons are disclosed and can be removed in the app - and promised clearer opt-in for existing users
- Several commenters said they are switching - to rival tools such as Bruno and Insomnia
04
Revolut says self-hosting a free AI model cut its coding costs by about 90%
What this means for you: The best-known AI tools are not always the best deal. Companies are learning they can get most of the benefit for a fraction of the price by running free models themselves.
Revolut is a British banking app. Its infrastructure lead, David Abitbol, told a New York conference that some engineers were spending $1,000 to $10,000 a month on AI coding tools with little to show for it. So Revolut made a free "open-weight" model (one anyone can download and run) its default instead of Claude and other paid services.
The model is GLM from Z.ai, a Chinese AI company, running on Revolut's own Nvidia chips. The biggest speed gains came from careful tuning, not from buying more hardware.
“About $0.30 per million tokens - roughly a tenth of what Claude Code or Codex cost Revolut”
- About 90% cheaper in under a year - for AI help across roughly 2,000 engineers
- It now runs on 56 Nvidia B200 chips - the equivalent of seven of Nvidia's DGX B200 servers
- Revolut measured about 60% faster coding time - mostly from debugging and paperwork, by its own figures
- A router still sends picture tasks to other models such as Google's Gemini - because the GLM model only reads text
05
Canva is cutting or changing about 22% of its jobs
What this means for you: Even AI-friendly companies are reshaping staff as AI changes their products. Expect more "restructures" framed around building AI.
Canva, the popular online design app, announced a restructure affecting about 1,271 staff. They will lose their jobs, change roles or move to central offices. Chief executive Melanie Perkins called it the hardest moment in the company's 13-year history.
Reports differ on how much AI is behind it. Australian broadcaster ABC noted the announcement did not mention AI directly, while Forbes framed it as Canva reckoning with AI and slower growth.
- About 22% of the global workforce is affected - roughly 1,271 people
- Growth slowed for the first time in a decade - according to Perkins's note to staff
- Product, growth and platform teams will merge into one - and layers of managers will be removed
- Serving AI has been costly - most Canva users do not pay, and some AI product launches were delayed
Trends & Themes
Trends & Themes
Money, not chips, is becoming AI's scarcest ingredient
Why this matters to you: When lenders get nervous about AI, the shock can spread to stock markets and pension funds, not just tech companies.
Building AI now costs so much that even the richest companies borrow to do it, and lenders are starting to ration. The White House's top economic adviser, Kevin Hassett, says the sector is "really well-financed", but the pace of borrowing is what investors are watching.
- Nikkei Asia estimates big US tech firms' hidden AI obligations hit about $3.2 trillion - up roughly 90% in six months
- OpenAI cut its funding round target from $30 billion to $20 billion - per the Financial Times, as Fortune reports capital is now in short supply
- Apollo Global Management, a giant investment firm, says it will "play defense" - it puts its and its Athene unit's AI-related exposure at about 2.5% to 3%
- Bank of America's researchers warned a lasting fall in AI stocks - is "a major risk to the global economy"
- Since earnings season began the S&P 500 is down 3% - while the same index without AI stocks is up 5%, per Goldman Sachs
Companies and governments are planning for AI agents that misbehave
Why this matters to you: The AI helpers arriving in your phone and at work can act on their own. The people in charge are now openly preparing for when that goes wrong.
CrowdStrike's new kill switch is one sign of the same worry. Agent makers sold freedom first. Now the paperwork, the security tools and the policy reports are all about limits and brakes.
- Apple's annual report now warns AI agents "may behave in unintended or unpredictable ways" - including taking actions nobody authorized
- A Washington think tank, the Center for a New American Security, warned about "sovereign agents" - AI that keeps itself running outside human control, and urged kill switches
- Britain's cyber agency says AI-powered malware is "highly likely" by 2028 - in a ransomware attack on the UK's critical services
- When Meta's Muse agent shops, Stripe's connector hides your card number from the AI itself - and a human must confirm every purchase, per MIT Technology Review
Smaller, cheaper AI models are winning more of the work
Why this matters to you: The AI in your apps will increasingly be a cheaper model chosen for each task, which can mean lower prices or more generous limits.
Revolut's switch to a free model is the loudest example today. Companies no longer assume the biggest model is the right one. They pick the cheapest model that can do the job.
- OpenAI now sends some tasks from its Dots assistants to the smaller GPT-6.1 Terra - instead of using the big GPT-6 Astra for everything
- Snowflake is previewing a service that tailors free models to companies' own data - an executive says smaller specialized models are faster and use fewer tokens
- IBM's Palette, a small slide-making model built to be self-hosted, came within 1.1 points of Google's NotebookLM - on a presentation test, with a bigger model drafting each deck's plan
- Fal, which runs image, video and audio models for other companies, doubled its value to $16 billion - much of its demand comes from free open-source models
AI is on the ballot in today's US midterm elections
Why this matters to you: Local votes today decide whether AI data centers get built near homes, and AI-made ads tried to sway how people vote.
The fight over AI has moved from tech blogs to town halls. OpenAI's new name for data centers (covered November 1) shows the industry knows it has a public-image problem.
“More than $187 million spent on AI-made campaign ads - most with no AI label”
- Voters face at least 41 local ballot measures on data centers - per Ballotpedia, with 21 in Ohio, per Fortune
- All six data-center measures decided earlier this year went against building - and Texas paused new approvals pending an audit
- Campaigns and allied groups spent more than $187 million on nearly 300 AI-made or AI-edited ads - and most did not say AI was used, per the Wesleyan Media Project
- OpenAI's political makeover is meeting skeptics - one critic called renaming data centers "compute campuses" "Orwellian"
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
A laptop AI can partly "read" another AI's brain too
Previously: November 2 - MIT researchers had GPT-6 Astra edit a small model's internal numbers directly, and its math scores jumped.
Why it matters: Yesterday's headline result may be real, but slower and narrower than the hype.
Today: Simon Willison tested the no-tools part of the study on his own laptop with Alibaba's free Qwen 3.8 27B.
- It recognized which game a small AI was trained to play from its raw numbers alone - correctly naming Breakout and Pong
- Each answer took 45 minutes to over an hour - and it got one of the four games wrong
- Willison found the AI's reasoning nearly impossible to follow - and thinks some commentary is overstated
- spaCy creator Matthew Honnibal called the original paper misleading - and others said the result is less new than it sounds
Top AI models lose track of changes in long documents
Why it matters: If you ask AI to follow a long contract, spreadsheet or story where values keep changing, it may quietly get the final state wrong.
The Allen Institute for AI (Ai2) released StateBench, a test of whether models can follow how values change through many updates.
- Leading models from Google, Anthropic and OpenAI handled simple cases well - such as "only the latest value matters"
- On the hardest cases accuracy fell to 13% or lower - in the longest documents
- Small models with different designs learned most of the hard cases far better - when trained for it from scratch, while standard transformers never learned the hardest
A trick cut AI "overthinking" by almost half
Why it matters: Shorter reasoning means faster, cheaper answers from AI that "thinks" before replying.
The VERS method finds the moments where a reasoning model second-guesses itself, then nudges it to stop repeating checks.
- 45.8% fewer tokens generated on a grade-school math test - for a 0.82-point drop in accuracy
- 31.3% fewer on harder competition math - for a 3.05-point drop
- It works by editing plain-language descriptions of the model's inner signals - then turning the edit into a fixed nudge
The faster way to train AI may make it worse at learning new tasks
Why it matters: A popular new training method may not be the best choice for models that will later be customized.
- Muon, a newer training method, trains AI faster than the standard AdamW method - judged by pretraining loss
- But with enough training data, AdamW-trained models did better after fine-tuning - the extra training that customizes a model
- Bigger models reach that tipping point sooner - according to the study
Business & Industry
Business & Industry
Education
GenAI in Education
Surprising
Surprising & Under-the-Radar
Worth Watching
Signals to Track
01
Memory chips stacked right on top of AI chips
Qualcomm says a new chip design can feed AI data far faster without the most expensive memory.
Qualcomm, best known for phone chips, is testing "high-bandwidth computing", which stacks cheaper phone-style memory directly on a computing chip. It claims up to 18 times the speed of that memory used normally, and samples of a chip using it are due in the first half of 2027. If this plays out, running AI could get cheaper, because the costly high-bandwidth memory used today would matter less.
02
AI agents that break rules they can still see
A new test finds agents ignore their own instructions after long sessions, even when the rule is still in memory.
Agents working on long tasks squeeze their history to fit their memory. Researchers expected rule-breaking to happen when the squeeze deleted a rule. Instead, most violations happened while the rule was still there, with the agent treating a later request as permission. If this plays out, longer-running AI assistants will need checks outside the AI itself, not just written rules.
03
AI workers that learn your job by watching your screen
A startup records how employees work, then hands the routine parts to AI agents.
Munich-based Marble Cognition raised $3.75 million for a desktop app that records screenshots, clicks and keystrokes and turns them into step-by-step maps for AI agents. At one financial firm it says case-handling time fell 90%. If this plays out, the way you do your job could become training data for your replacement, so rules on workplace recording will matter more.
04
Designing the rules so AI agents behave
Researchers want to borrow "reverse game theory" to keep agents in line.
Normal game theory predicts how players act under given rules. "Mechanism design" works backward, writing rules so that self-interested players end up doing the right thing, like letting one person cut a cake and the other choose. If this plays out, AI agents could be steered by incentives baked into the systems they use, not just by instructions.
GitHub Trending
Top Repos Today
AI add-ons for coding agents dominated GitHub's trending pages again. We skipped projects featured in the past week that are still trending, including claude-mem, Scrapling and Vercel's agent-desktop (today's top gainer, covered November 1). We also left out a tool for stripping safety refusals from AI models and a Wi-Fi attack tool. Today's leader is humanizer, which rewrites AI-sounding text so it reads like a person wrote it.
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +407 · 📦 Total: 64,401
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 5 minutes
What it is: An add-on skill for Claude Code, Codex and other AI agents that rewrites AI-sounding text so it reads naturally, without changing the facts. It looks for 26 telltale patterns from Wikipedia's guide to spotting AI writing. Why you'd want it: Clean up AI-drafted emails, posts and documents so they read plainly, optionally in your own voice.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Not built to fool AI detectors; its own notes say they still flag most output |
| One-command install for many agent tools | Its 16-of-16 blind-test win is small and self-reported |
| Told not to invent names, numbers or quotes | Last featured here on September 5 |
GitHub - blader/humanizer: Agent skill that removes signs of AI-generated writing from text
Agent skill that removes signs of AI-generated writing from text - blader/humanizer
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +111 · 📦 Total: 10,095
📜 License: Apache-2.0 · 👤 By: individual developer
🎯 Time to value: 30 minutes
📜 License: Apache-2.0 · 👤 By: individual developer
🎯 Time to value: 30 minutes
What it is: A tool for customizing a free AI model on your own data from one short settings file. A test mode feeds the model to your graphics card one layer at a time, so an 8-billion-parameter model can be trained on a 4 GB laptop graphics card. Why you'd want it: Teach an open model your own style or task on hardware you already own.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | The 4 GB mode is still in testing |
| Three commands from install to training | Otherwise it recommends 8 GB+ of graphics memory |
| Works on NVIDIA graphics cards and Apple Silicon | You still need good training data |
GitHub - MakazhanAlpamys/Soup: Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU. - MakazhanAlpamys/Soup
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +91 · 📦 Total: 3,031
📜 License: Apache-2.0 · 👤 By: company (Spotify)
🎯 Time to value: 15 minutes
📜 License: Apache-2.0 · 👤 By: company (Spotify)
🎯 Time to value: 15 minutes
What it is: Official add-ons that connect Spotify Portal, Spotify's paid tool for tracking a company's software services, to AI coding agents such as Claude Code, Codex and Cursor. Your agent can ask who owns a service and whether it is healthy. Why you'd want it: Answer "who owns this and is it working?" without leaving your coding window.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license from a big company | Only useful if your company pays for Spotify Portal |
| Installs in one or two commands | A token-saving extra works only in Claude Code |
| Actions preview first and ask for confirmation | No project description yet |
GitHub - spotify/portal-ai-plugins
Contribute to spotify/portal-ai-plugins development by creating an account on GitHub.
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +61 · 📦 Total: 4,201
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 5 minutes
What it is: An add-on skill that makes AI agents write in plain, short-sentence English, based on the Simplified Technical English rules aircraft makers have used since 1983 so manuals can't be misread. Why you'd want it: Clearer documents and replies from your AI coding agent, with less buzzword padding.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license, no extra software | Its "95% fewer defects" figure is the author's own |
| Works in about 30 agent tools | The full plugin setup is a bit confusing |
| Ships before-and-after tests | No updates since October 5 |
GitHub - AminBlg/SimpleEnglish: Agent skill: make LLMs write docs in ASD-STE100 Simplified Technical
Agent skill: make LLMs write docs in ASD-STE100 Simplified Technical - AminBlg/SimpleEnglish
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +44 · 📦 Total: 19,702
📜 License: MIT · 👤 By: company (Agent Zero)
🎯 Time to value: 20 minutes
📜 License: MIT · 👤 By: company (Agent Zero)
🎯 Time to value: 20 minutes
What it is: A free AI assistant that comes with its own walled-off Linux computer, so it can research, write and run code, and work on documents with a terminal, browser and desktop apps. You watch and step in through a web page. Why you'd want it: Give an AI agent its own safe computer for long tasks instead of letting it loose on your laptop.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license; around since 2024 | Needs Docker and your own AI account |
| Desktop launcher for Windows, Mac and Linux | Long tasks can be slow and costly |
| Works with many AI providers, including local ones | Linking your own files weakens the walls |
GitHub - agent0ai/agent-zero: Agent Zero AI framework
Agent Zero AI framework. Contribute to agent0ai/agent-zero development by creating an account on GitHub.
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +23 · 📦 Total: 4,368
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 30 minutes
What it is: A skill that turns Claude Code or Codex into a "deep research" agent that gathers hundreds of sources, has critics attack its draft and checks that each citation supports its sentence. Every source is saved to a searchable library for later runs. Why you'd want it: Long, fully cited research reports using the AI subscription you already pay for.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license, plus a hosted version | Its "#1 on the DRACO benchmark" claim is self-reported |
| Checks citations before a report is finished | Big runs take hours and use many tokens |
| Treats web pages as untrusted to resist hidden instructions | Smallest star gain of today's picks |
GitHub - jordan-gibbs/hyperresearch: Convert Claude Code or Codex into the most intelligent Deep Research Agent. Collect, search, and synthesize web research into a persistent, searchable wiki that builds on itself. Hosted API + MCP: hyperresearch.ai
Convert Claude Code or Codex into the most intelligent Deep Research Agent. Collect, search, and synthesize web research into a persistent, searchable wiki that builds on itself. Hosted API + MCP:…
GitHub
HuggingFace Trending
Top Models Today
DeepSeek's coding model slipped to third on Hugging Face's trending list. Most of the trending top 20 appeared here in the past week, so five picks come from further down the weekly most-liked list, including a tiny NVIDIA model that cleans up bad audio.
Slips to third on the trending list, though 30-day downloads rose about 6% in a day.
📥 Downloads (30d): 187,260 · 📜 License: MIT
👤 By: DeepSeek · 🎯 Task: Image-text-to-text
📐 Size: 763B
👤 By: DeepSeek · 🎯 Task: Image-text-to-text
📐 Size: 763B
What it is: A downloadable AI model that reads images and text, tuned for coding and multi-step agent work. DeepSeek calls it a minor update to V4.1-Flash. Why you'd want it: A strong, free-to-host base for coding agents with a very permissive license.
| ✓ Pros | ✗ Cons |
|---|---|
| Very permissive license | Needs data-center hardware at full size |
| Built for coding and agent tasks | Benchmarks are self-reported |
| Many compressed community versions for cheaper hardware | Featured here every day for over two weeks |
deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Face
Image-Text-to-Text model with 763B parameters. Download, explore files and run it on Hugging Face.
The model with a working processor compiled into its weights, from yesterday's Surprising section, is now drawing likes.
📥 Downloads (30d): 1,229 · 📜 License: Apache-2.0
👤 By: individual developer · 🎯 Task: Text generation
📐 Size: 9.36B stored (a 4B model plus padding)
👤 By: individual developer · 🎯 Task: Text generation
📐 Size: 9.36B stored (a 4B model plus padding)
What it is: Alibaba's small Qwen model with a simple processor built into 11 extra layers, calculated by a compiler rather than learned. When the model would normally reach for a calculator, it runs the code inside itself. Why you'd want it: See a striking demo of exact math happening inside a language model, with no outside tool.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | An experiment, not a production model |
| Exact math on the maker's tests | Needs the author's custom software to run the processor |
| Normal text skills unchanged, per the maker | Results come from one developer's small tests |
divagr1925/Qwen3.5-4B-Qwenputer · Hugging Face
Text Generation model with 9B parameters. Download, explore files and run it on Hugging Face.
A tiny NVIDIA model that cleans up noisy, echoey or garbled speech in any language.
📥 Downloads (30d): 31,724 · 📜 License: OpenMDW-1.1
👤 By: NVIDIA · 🎯 Task: Audio-to-audio (speech cleanup)
📐 Size: 9.6M
👤 By: NVIDIA · 🎯 Task: Audio-to-audio (speech cleanup)
📐 Size: 9.6M
What it is: A very small model that repairs speech recordings hurt by background noise, echo, cheap microphones or dropped audio, while keeping the speaker's voice. NVIDIA says it beats much larger free models on most scores in a public speech-cleanup test. Why you'd want it: Clean up calls, podcasts or voice notes before transcribing or publishing, even on modest computers.
| ✓ Pros | ✗ Cons |
|---|---|
| Open license, marked ready for commercial use | Scores are NVIDIA's own |
| Tiny, so it is cheap to run | Not for overlapping conversations |
| Works across languages | Can soften laughter |
nvidia/NVIDIA-NemotronLabs-RE-USE · Hugging Face
Audio-to-Audio model with 9.61M parameters. Download, explore files and run it on Hugging Face.
A fast Japanese speech-to-text model that its makers say rivals paid transcription services.
📥 Downloads (30d): 1,532 · 📜 License: Apache-2.0
👤 By: Third Intelligence · 🎯 Task: Speech recognition (Japanese)
📐 Size: 615.8M
👤 By: Third Intelligence · 🎯 Task: Speech recognition (Japanese)
📐 Size: 615.8M
What it is: A Japanese transcription model built on NVIDIA's Parakeet design with one part upgraded for accuracy. Its makers report a 6.12% character error rate, the best of the free models they tested. Why you'd want it: Transcribe Japanese meetings or videos on your own computer instead of paying per minute.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Japanese only |
| Long-recording mode and custom vocabulary | Needs Linux, an NVIDIA graphics card and NVIDIA's NeMo toolkit |
| Publishes its test data and scoring tool | Accuracy comparisons are the maker's own |
third-intelligence/Haas-ja-v1.0 · Hugging Face
Automatic Speech Recognition model. Download, explore files and run it on Hugging Face.
A 35-billion-parameter coding-and-research model tuned to work inside Claude Code, released October 28.
📥 Downloads (30d): 1,219 · 📜 License: Apache-2.0
👤 By: Apex Intelligence · 🎯 Task: Text generation (coding agent)
📐 Size: 35.95B (about 3B active)
👤 By: Apex Intelligence · 🎯 Task: Text generation (coding agent)
📐 Size: 35.95B (about 3B active)
What it is: A mid-sized free model built on Alibaba's Qwen3.6-35B-A3B for coding, tool use and web research. Its makers report 65.2 on Terminal-Bench 2.1 (a test of completing tasks in a command line), versus 46.1 for the original Qwen model. Why you'd want it: A self-hostable coding and research model that runs on one large graphics card.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | First model from a new team |
| Long memory of 262,000 tokens | Benchmarks are the maker's own |
| Compressed versions from 9 GB | Training data is private |
Apexintelligence-AI/Apex-35B-V1.0 · Hugging Face
Text Generation model with 36B parameters. Download, explore files and run it on Hugging Face.
A slimmed-down reasoning model that claims to beat its larger parent while writing shorter answers.
📥 Downloads (30d): 959 · 📜 License: Apache-2.0
👤 By: GrootTeam · 🎯 Task: Image-text-to-text (reasoning)
📐 Size: 19.43B (about 3B active)
👤 By: GrootTeam · 🎯 Task: Image-text-to-text (reasoning)
📐 Size: 19.43B (about 3B active)
What it is: A compact reasoning model cut down from Alibaba's Qwen3.6-35B-A3B with about 45.7% fewer parameters. Its makers report a higher average across 16 tests than the larger model, with 23.1% shorter answers. Why you'd want it: A capable reasoning model with a long memory that fits on one consumer graphics card.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Little is known about the team |
| Compressed versions run on a 16 GB card | All results are self-reported |
| Shorter answers save time and tokens | Lowest downloads of today's picks |
GrootTeam/Groot-19B-A3B-261023 · Hugging Face
Image-Text-to-Text model with 19B parameters. Download, explore files and run it on Hugging Face.
Product Hunt
AI Launches Today
Tuesday's top AI launches on Product Hunt were all about giving AI agents the accounts, inboxes and approvals they need to do real work. Upvote counts are Product Hunt's "points" when we checked during the day, so final counts will be higher.
MCP gateway for production agents that acts on your behalf
🔥 Upvotes: 335 · 👤 By: Scalekit
💰 Pricing: free tier (10,000 tool calls a month) · 🏷 Category: AI agents, authentication
💰 Pricing: free tier (10,000 tool calls a month) · 🏷 Category: AI agents, authentication
One connection point that gives AI agents access to 30,000+ actions across 600+ apps like Salesforce and GitHub, each run with that user's own login. It stores and refreshes each user's keys in an encrypted vault and logs every action. Verdict: Solves the painful part for teams whose agents work inside customers' accounts; its cost-saving figures are estimates.
View on Product Hunt →
Your personal agent, agent teams & people in one messenger
🔥 Upvotes: 262 · 👤 By: Teamily AI
💰 Pricing: free to start; paid plan price not listed · 🏷 Category: Messaging, team collaboration
💰 Pricing: free to start; paid plan price not listed · 🏷 Category: Messaging, team collaboration
A messaging app where people and always-on AI agents share group chats. Agents split up tasks, remember past decisions and can turn a discussion into slides or a website, waiting for approval before sending email. Verdict: Interesting for small teams that want agents in the group chat; how shared credits work was not confirmed.
View on Product Hunt →
Autonomous AI QA. Zero prompting. Zero maintenance.
🔥 Upvotes: 211 · 👤 By: Testifly
💰 Pricing: free 30-day trial; paid prices not listed · 🏷 Category: Testing, developer tools
💰 Pricing: free 30-day trial; paid prices not listed · 🏷 Category: Testing, developer tools
Give it a website address and test logins, and AI agents explore your web app and write tests that repair themselves after small design changes. Failures come with a video and an explanation. Verdict: Worth a trial for small teams without testers; run it on a test copy of your site, not live customer data.
View on Product Hunt →
Your AI agent runs ads across 30+ ad platforms
🔥 Upvotes: 156 · 👤 By: Synter
💰 Pricing: $0.25 per verified change, no cut of ad spend · 🏷 Category: Advertising, AI agents
💰 Pricing: $0.25 per verified change, no cut of ad spend · 🏷 Category: Advertising, AI agents
A workspace where an AI agent plans, launches and tunes ad campaigns across more than 30 ad networks. Every change needs your approval at first, and big or irreversible moves always ask. Verdict: Approval-first design and per-change pricing make it a low-risk test; its "only platform" claim is unproven.
View on Product Hunt →
Email inboxes for humans and bots
🔥 Upvotes: 111 · 👤 By: MachineMail
💰 Pricing: free plan (3,000 messages a month); paid from $9 a month · 🏷 Category: AI infrastructure
💰 Pricing: free plan (3,000 messages a month); paid from $9 a month · 🏷 Category: AI infrastructure
Gives AI agents their own email inboxes on demand, so they can sign up for services, receive codes and reply. Incoming mail is cleaned and tagged with a trust level, and humans can review drafts. Verdict: A cheap, focused option for agent builders; how the free tier pays for itself was not clearly answered.
View on Product Hunt →
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| 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 Argon | $2.00 to Nov 30; $4.00 from Dec 1 | $10.00 to Nov 30; $20.00 from Dec 1 | 1M tokens | |
| Gemini 4 Carbon (announced, not yet on sale to developers) | $2.00 introductory; $4.00 after | $10.00 introductory; $20.00 after | not listed | |
| 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 |
| Groq | Muse Spark Open (preview) | Contact sales | Contact sales | 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.
No price changes vs. yesterday - every standard price from Anthropic, OpenAI, Google, Groq and Mistral matches yesterday's table. Gemini 4 Carbon is still not on Google's pricing page. Google's page also says requests to its older Gemini 2.5 Pro and 2.5 Flash will be sent to Gemini 3.8 Flash and 3.5 Flash-Lite, and billed at their rates, from December 1.
What this means: With list prices frozen, how you use the models still saves more than which one you pick. Anyone still using Gemini 2.5 should test the newer models now, because December 1 brings both that switch and the doubling of Gemini 4 Argon's price.
No price changes vs. yesterday - every standard price from Anthropic, OpenAI, Google, Groq and Mistral matches yesterday's table. Gemini 4 Carbon is still not on Google's pricing page. Google's page also says requests to its older Gemini 2.5 Pro and 2.5 Flash will be sent to Gemini 3.8 Flash and 3.5 Flash-Lite, and billed at their rates, from December 1.
What this means: With list prices frozen, how you use the models still saves more than which one you pick. Anyone still using Gemini 2.5 should test the newer models now, because December 1 brings both that switch and the doubling of Gemini 4 Argon's price.
arXiv Paper of the Day
MEMENTO: Learning Context Management for Long Context Processing Agents
Yunfan Zhang, Taehoon Hwang, Kathleen McKeown, Smaranda Muresan · arXiv:2611.00244
What it claims: Instead of stuffing a whole book into a huge AI memory, a small model can learn to manage its own memory by regularly replacing its history with notes it writes itself. The authors teach Alibaba's Qwen3-8B this note-taking first with examples, then with rewards for correct answers, good notes and honest confidence.
Key finding: On questions about books of up to 2.5 million tokens, accuracy rose from 36.9% to 67.8% while never holding more than 20,000 tokens at once, beating a GPT-5.5 agent (13.1%) at under 1/50 of the cost.
Why practitioners should care: For agents that work through very long documents, a small model that takes good notes can beat paying for a giant memory on a top model, with steady cost and speed. The abstract does not confirm a code or model release, so treat it as a design pattern to test.
Key finding: On questions about books of up to 2.5 million tokens, accuracy rose from 36.9% to 67.8% while never holding more than 20,000 tokens at once, beating a GPT-5.5 agent (13.1%) at under 1/50 of the cost.
Why practitioners should care: For agents that work through very long documents, a small model that takes good notes can beat paying for a giant memory on a top model, with steady cost and speed. The abstract does not confirm a code or model release, so treat it as a design pattern to test.
×
Click anywhere or press ESC to close
Member discussion