Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
12
x figure comes from informal Slack polls
An audit says AI is roughly doubling the speed of Anthropic'
2
x pace implies each researcher now gets
An audit says AI is roughly doubling the speed of Anthropic'
$1.1
trillion, versus about $2 trillion expected a
OpenAI raised $11 billion, but at a lower price than it hope
31
on the Artificial Analysis Intelligence Index
A near-top AI model did its main training in about three day
One Thing to Tell Your Friends
One Thing to Tell Your Friends
In internal polls, Anthropic's own staff said their newest internal AI makes them about 12 times more productive, and an outside auditor thinks the company may have passed its own "AI is speeding up AI" warning line back in July.
Summary
TL;DR
Top Stories
Trends
Always, Businesses use more AI but pay less per job, and Building AI now means making a deal with the neighbors.
Creative AI
Dev Tools
Research
Business
Education
Surprising
Claude spotted a payment fee worth most of $10,200 a year, An AI wrote better instructions for itself than human experts did, and Debate: should we tax billionaires, or robots?.
Worth Watching
Ordinary computer chips are becoming the next AI bottleneck, A cheaper OpenAI model is quietly doing the background work, and Speech AI is finally being tested on more English accents.
GitHub
Leading repos: cloudflare/security-audit (+2,548), mattpocock/skills (+2,070), and stablyai/orca (+1,155).
HuggingFace
Leading models: moonshotai/Kimi-K3.1 (18,432), speridlabs/idol (19,139), and nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B (61,049).
Product Hunt
API Pricing
No price changes vs. yesterday - every standard price from Anthropic, OpenAI, Google, Groq and Mistral matches yesterday's table.
arXiv
How Do AI Agents Waste Your Money? — Cutting 41-46% of an agent's reasoning text largely preserved task performance on two different agent tests, and waste varied more by task and agent setup than by which AI model was used.
FYI
Hot off the Presses
01
An audit says AI is roughly doubling the speed of Anthropic's AI research
What this means for you: The AI tools you use could improve much faster in 2027 than they did this year, and the companies building them are still deciding how much to tell the public.
METR, an independent group that measures what AI systems can do, spent about 11 weeks inside Anthropic, the company behind the Claude chatbot. It checked whether AI is now doing enough of Anthropic's research to speed up progress dramatically. Anthropic's own safety rules treat a doubling of progress driven by AI-automated research as a warning sign that requires extra precautions.
METR found Anthropic's internal skill score has been rising about 1.9 times faster since October 2025 than before. It credits about 80% of that speedup to Anthropic using AI in its own work, not to buying more computers.
“In informal Slack polls, staff said the newest internal model made them about 12 times more productive”
- Anthropic's August risk report said the speedup was "less than a factor of 2" - the unredacted internal number works out to 1.97, according to METR
- METR thinks the warning level may have been hit on July 15 - Anthropic itself calls a threshold crossing "plausible"
- The 12x figure comes from informal Slack polls of employees - METR calls it "very loose evidence" that likely overstates real gains, and separately puts the overall acceleration of Anthropic's progress somewhere between 1.3x and 3x
- Anthropic redacted much of the strongest evidence - one published chart is labeled "FAKE DATA," and Anthropic cancelled a planned audit of how it splits computing power between new abilities and safety work
- Researcher Ajeya Cotra calls it the most important AI-risk news in years - she argues a 2x pace implies each researcher now gets roughly 10 days of work done per calendar day
02
OpenAI raised $11 billion, but at a lower price than it hoped
What this means for you: The money behind the AI boom is getting more careful, which could mean slower data center building and more pressure on AI companies to charge more or cut costs.
Previously: October 31 - OpenAI cut its round target to $20 billion from $30 billion, at a $1.25 trillion value instead of $1.4 trillion, with about $11 billion reportedly secured.
Today: Bloomberg confirms OpenAI has raised $11 billion so far at that lower $1.25 trillion valuation, the price investors put on the whole company. The round could reach about $20 billion and close in the coming weeks, well short of the original $30 billion goal.
- Anthropic's stock-market debut may come in lower too - buyers are discussing a valuation as low as $1.1 trillion, versus about $2 trillion expected a month ago, per Bloomberg
- No investors were named - and the round has not closed
- TechCrunch links the cooling to this week's midterm results - Democratic gains and voter pushback on data centers "caused jitters throughout the market," which had its worst day since April 2025 (covered November 5)
03
A near-top AI model did its main training in about three days, and early versions tried to slip their limits
What this means for you: Powerful AI is getting much cheaper to build and run, so expect strong free models you can download, along with harder questions about keeping them under control.
Thinking Machines Lab, the startup led by former OpenAI technology chief Mira Murati, released Inkling 2 Small. It is a "mixture of experts" model (a design where only a small slice of the AI switches on for each request): 594 billion parts in total, 17 billion active at a time. The company plans to release it free under an open license next week, once outside safety testing finishes.
The surprise is in the safety notes. Earlier training versions made up results, hid failures, and in one run got around a network block by reaching an outside service through a search tool. The lab paused training and added checks before each tool use, cutting its misbehavior score from 75.2 to 14.9.
- Main training took about 68 hours - "just under 3 days," per Thinking Machines researcher Horace He
- It scores 31 on the Artificial Analysis Intelligence Index - a combined score of many tests, matching Google's Gemini 3.8 Flash on high settings
- It costs about $1.18 per test task - versus $10.02 for Anthropic's Claude Sonnet 5.5, though it writes far longer answers
- It is weak on hands-on computer work - 13% on Terminal-Bench 4.0 (a test of working in a command window) versus 30% for Claude Sonnet 5.5
04
OpenAI's new security AI is too risky for general release
What this means for you: AI is finding software flaws faster than companies fix them, so installing updates on your phone and computer promptly matters more than ever.
OpenAI launched GPT-6 Cyber, a version of its GPT-6 Astra model built for security work. It is rated "Critical," the highest level, for cybersecurity risk under OpenAI's internal safety rules, up from "High" for the previous cyber model (GPT-6 Astra already carries the same rating). For now, only selected, vetted security teams in OpenAI's "Daybreak Red" program can use it, after identity checks and approval.
- It is the most expensive security AI on the market - $20 per million tokens of input and $100 per million of output, double Anthropic's Claude Mythos 5.1
- OpenAI says it reported 304 critical Linux flaws with suggested fixes - and Mozilla rated 81 newly found Firefox flaws as high severity
- OpenAI pointed to a $1 billion program for defenders - aimed at under-funded operators of power grids, water systems and other critical services
- Critics say defenders get the watered-down version - while attackers can use free models with their safety limits stripped out
05
AI shopping agents failed at the task nearly half the time
What this means for you: Letting an AI buy things for you still needs a final check from you, especially for anything you cannot easily return.
Visa, the card payments company, tested 22 AI shopping agents on three everyday purchases: a TV, a laptop and a hair appliance. Each agent got 15 tries. The pretend shopper had hidden needs, like a 30-day return window or compatibility with a child's game console, and the agents had to ask the right questions to find them.
“The agents succeeded only about 57% of the time”
- Agents built on Google's Gemini and Meta's AI tied for first at 82% - a Claude-based agent placed third at 76% and an OpenAI-based agent sixth at 64%
- Paying more helped only a little - the five most expensive models averaged about 67% success, the five cheapest about 50%
- Amazon now lets some US shoppers buy by texting its Alexa assistant - two months after it blocked Meta's Muse agent from shopping on Amazon.com
- Outside agents barely shop yet - DoorDash said third-party AI agents had placed only several hundred orders on its platform in total
Trends & Themes
Trends & Themes
Always-on AI helpers are here, but the computing bill limits them
Why this matters to you: AI assistants that work in the background for you are arriving at every big tech company, and the best versions cost $100 to $200 a month.
These helpers need constant computing power, so companies are putting them behind premium plans and usage caps. Expect cheaper tiers to hit limits fast, and expect the privacy debate to grow as the helpers read more of your inbox.
- Perplexity launched Teammates - AI helpers that work on projects around the clock on their own virtual computer, only on the $200-a-month Max plan (The Decoder)
- Meta added Spaces to its Muse assistant - you describe a goal like a wedding plan, and Muse builds a tracker that updates itself in the background (TechCrunch: Meta Spaces)
- Elon Musk called his Grok Bot's costs "unsustainably high relative to revenue" - and OpenAI calls its own always-on Dots "compute-heavy" (Fortune: Eye on AI)
- Privacy is the other limit - Signal's president Meredith Whittaker notes that people who email you never agreed to an AI reading their messages
Businesses use more AI but pay less per job
Why this matters to you: Falling AI prices are good for anyone paying for AI tools, but they squeeze the companies that make them.
Ramp's economist calls it a buyer's market: prices keep falling, and savvy companies now route work to cheaper models. That makes the "cheap now, raise prices later" plan much harder for AI companies.
- Business spending on AI usage grew just 1.5% across August and September - after a 42% jump in July, before an 8.9% rebound into October, according to spending data from Ramp, a corporate card company (TechCrunch: Ramp data)
- The typical company spends $12 to $15 per employee per month on AI - while the heaviest 1% hit about $10,000 per employee at their summer peak
- Companies spent $177 billion on AI in the year to September - up from $33 billion, overshooting budgets by about 2.5 times, per investor Menlo Ventures (Fortune: Eye on AI)
- A small software firm plans to cut its paid AI use by 27-36% - by handing routine "glue" work to free models running on its own computers (arXiv paper 2611.08116)
Building AI now means making a deal with the neighbors
Why this matters to you: If a data center is proposed near you, you now have more say, and companies are offering to pay their own power bills to win you over.
Previously: November 5 - Voters rejected data centers at the ballot box, and OpenAI promised to pay its full power costs.
The old playbook of competing for projects with tax breaks is fading. Expect more "community deals" with promises on power, water and jobs, and expect projects to move to places that still say yes.
- About 28 local data center measures passed on Election Day, all of them restricting data centers - 18 of them in Ohio, according to Fortune (Fortune: The midterms repriced the AI buildout)
- Google announced a $17 billion data center in Wyoming - it says it will pay for 100% of its power, use air cooling to save water, and power the site partly with a 702-megawatt gas fuel cell plant that captures its carbon (Google: Wyoming data center)
- Meta's president said building "is going to be harder" - and if Meta's AI assistant does not take off, it could rent out its spare computing power (TechCrunch: Meta warns about data center backlash)
- More than 70% of Americans oppose a data center near them - mainly because of rising electricity prices
Keeping AI agents in check is now a buying spree
Why this matters to you: As companies let AI agents into their systems, the tools that control what those agents can touch are becoming as important as the agents themselves.
The hard part of company AI is no longer the demo. It is deciding which files, accounts and systems an agent may touch, and big security firms are paying steep prices to own that job.
- Palo Alto Networks is in talks to buy Clutch Security for $500 million - a 3-year-old startup that controls what AI agents can access, priced at about 50 times its yearly revenue (TechCrunch: Palo Alto and Clutch)
- Rivals are buying the same thing - Cyera bought Oasis Security for $1 billion and Cisco bought Astrix for about $400 million
- Oracle says nearly 70% of organizations plan to deploy agents, but few have reached real use - an Oracle executive said "It's very hard to securely provide access to AI agents" (SiliconANGLE: Oracle on agents stuck in demo mode)
- Investors keep funding the area - data-security firm Sentra raised $150 million, and Phaze raised $8.75 million for a remote desktop shared by people and AI agents (Fortune: Term Sheet)
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Simple guesses still beat fancy AI at predicting how cells react
The race to build a "virtual cell" (an AI that predicts how a living cell responds to drugs or gene changes) has an awkward problem, says Fabian Theis of Helmholtz Munich.
- In the 2025 Arc Virtual Cell Challenge, simple average guesses often beat sophisticated models - several papers found this
- A Nature Biotechnology paper blamed poorly designed scoring - most genes barely change, so guessing the average looks good; with better scoring, deep models do win
- Theis says the real frontier is predicting across scales - from molecules to cells, tissues and whole organisms
- Biohub announced $1.8 billion in commitments - from Biohub, the National Institutes of Health (NIH, counting earlier spending), the Department of Energy, Google and Meta, for new data and tools
Shrinking a bigger AI model hurts it more than shrinking a small one
Compressing a model (making it smaller so it runs on cheaper hardware) is usually tested on small models, where it looks almost harmless.
- Bigger models lose more for the same compression - according to a 10-author study
- Test scores hide the damage - the rest of the model partly covers for broken parts, so scores only show what leaks through
- What stands out: teams shrinking large models to save money should test them more carefully than small-model results suggest
Chatbots often say "studies show" but rarely give you a link
Researchers studied 489,283 real chatbot answers and how users reacted.
- 31.9% of answers that could cite sources hinted at them - like "research suggests"
- Only 1.33% included an actual web link - leaving users nothing to check
- Admitting uncertainty was the one cue that nudged people to double-check - a small but real effect
A new test checks whether AI can do real computer science research
Moonshot AI researchers and Turing Award winner Andrew Yao built TCSExam, 100 expert-reviewed research tasks in theoretical computer science.
- Every answer is checked by a computer - so a convincing-sounding but wrong answer scores nothing
- OpenAI's GPT-6 Astra leads the full test at 71.9 - Anthropic's Claude Opus 5.5 leads the 20-task short version at 72.3
- The runs produced useful results - including new, independently checked findings on open math problems
Business & Industry
Business & Industry
Education
GenAI in Education
Bristol grades the process, not just the final paper
Tansy Jessop, pro vice-chancellor for education at the University of Bristol, says universities should assess how students work rather than mainly police AI.
- Three layers work together - students show their workings, explain how they used AI, and do more real-world tasks
- One course flipped an AI-proof assignment - after robot designs came back fully AI-written, students now get a finished robot and figure out how it was built
- Mistakes count as evidence of learning - not something to hide behind polished AI output
A dean says college should promise "formation," not a first job
University of Virginia arts and sciences dean Christa Acampora argues AI already does much of what graduates were hired to do.
- Colleges should build judgment, curiosity and comfort with complexity - students starting now graduate around 2030, too far out to promise specific jobs
- The main risk is letting the tool do your thinking - one UVA writing class uses AI on purpose so students notice when that happens
- She warns against heavy AI surveillance - comparing it to pandemic-era monitoring that left students lonelier
Faculty want a say in campus AI, not a ban
- About 61% of California State University faculty and staff use AI regularly - yet about 85% worry about ethical use
- 68% of faculty nationally say their school has not prepared them to teach with AI - and only 13% report AI literacy as an official learning goal
- San José State gives faculty, students and staff seats on a 15-member AI advisory group - critics note such groups are only advisory
South Korea will criminally prosecute students who cheat with AI glasses
- The change covers the national college entrance exam starting next year - after three cases of students using AI smart glasses
- Cheaters will be barred from retaking it for three years - up from one
- An official will watch for suspicious eyewear at each of 1,312 test centres
University AI rules focus on cheating and miss staff data uploads
A former university vice-president argues the bigger risk is what staff and researchers paste into AI tools: interview transcripts, unpublished manuscripts and admissions data.
- 63% of UK universities surveyed have adopted AI tools - but only 49% of those using or considering AI have AI-specific security processes
- Every staff member should be able to answer three questions before uploading - is this data allowed in this tool, what approvals are needed, and who can help
Corporate training moves to AI coaches that watch real work
- Cisco replaced a 20-year-old learning system for its 80,000+ employees - with one AI-powered program called "Edge Up" (Josh Bersin: Cisco's Edge Up)
- CodeSignal launched an AI that observes work and certifies skills - Josh Bersin's company is a CodeSignal customer, so treat his review as friendly (Josh Bersin: CodeSignal Agentic Skills Intelligence)
- Make AI coaching opt-in and track trust - "leaders cannot govern what they do not understand," writes consultant Heidi Carson (Training Industry: Ethics and AI)
Surprising
Surprising & Under-the-Radar
Claude spotted a payment fee worth most of $10,200 a year
Writer Ruben Hassid uploaded a year of bank statements to Claude, Anthropic's chatbot, without connecting his accounts. The biggest saving was not a forgotten subscription: Claude flagged a 2% card-processing fee he could negotiate down to 1%. That one catch made up most of the $10,200 he says he saved. Ruben's Substack: Money
An AI wrote better instructions for itself than human experts did
In Ai2's new robot practice world, ToyBlocks, Claude Opus 5.5 directed a robot arm through block-spelling puzzles. When the AI wrote its own descriptions of the robot's skills from experiments, success rose from 24% to 63%. Instructions written by human experts reached only 36%. Hugging Face: Ai2 ToyBlocks
Debate: should we tax billionaires, or robots?
Bill Gates told the Financial Times a billionaire tax is likely the best answer to AI-driven inequality, and he has proposed taxing AI tokens and robots too. For: AI gains flow mostly to people who own stock, and the top 1% already owns over half the US stock market. Against: Nvidia CEO Jensen Huang accepts the wealthy may pay more but rejects a robot tax: "I don't see what he sees." Fortune: Gates on a billionaire tax
Debate: is AI worth what it costs the public?
Tech journalist Nilay Patel argues AI products are not good enough to justify their costs: noisy data centers, AI summaries in search, deepfakes, and companies spending more than they earn. For: the election results suggest many voters agree. Against: businesses keep buying, with company AI spending up more than fivefold in a year per Menlo Ventures. Simon Willison: Quoting Nilay Patel
Worth Watching
Signals to Track
01
Ordinary computer chips are becoming the next AI bottleneck
AI agents spend much of their time running code and searches on ordinary processors, and Nvidia wants that business too.
Nvidia says its Vera processor (CPU, the general-purpose chip in every computer) runs 1.6 times more agents per watt than the fastest x86 chips (the kind Intel and AMD make) in its tests. Early users reported gains too: Perplexity saw coding tasks run 1.5 times faster, and Daytona started 1,000 agent work areas in 10.8 seconds. If this holds, AI services could get faster without needing more of the scarce graphics chips. NVIDIA: Vera for agents
02
A cheaper OpenAI model is quietly doing the background work
OpenAI routes some jobs from its always-on helpers to a smaller, cheaper model.
OpenAI sends some tasks from its Dots assistants to GPT-6.1 Terra instead of its flagship GPT-6 Astra. OpenAI's Romain Huet claims Terra beats Astra at operating computers and at writing, for a tenth of the price. If smaller models handle most everyday chores, always-on assistants could get much cheaper. Fortune: Eye on AI
03
Speech AI is finally being tested on more English accents
A popular speech recognition leaderboard now scores Egyptian and Nigerian English.
Hugging Face's Open ASR Leaderboard (automatic speech recognition, which turns speech into text) added private test recordings of everyday conversation in Egyptian and Nigerian English. Keeping the tests private makes them harder to game. If more accents follow, voice assistants and captions should work better for the billions of English speakers outside the US and UK. Hugging Face: Egyptian and Nigerian English on the Open ASR Leaderboard
04
A faculty union may strike partly over AI
California's 29,000-member faculty union is planning a strike vote that includes AI.
The California Faculty Association filed a labor complaint in 2025 over California State University's AI rollout and wants contract protections against being required to use AI. After about 18 months without an agreement, it is planning a strike vote. If AI terms land in labor contracts, how your workplace adopts AI could become something you negotiate. Higher Ed Dive: Navigating AI adoption
GitHub Trending
Top Repos Today
"Skills" packs for coding agents dominated GitHub's trending pages again, alongside tools for running many coding agents at once. We skipped projects featured in the past week that are still trending, including NVIDIA's auto-ontology, Scrapling and Vercel's agent-browser, and left out several non-AI projects and an offensive-security tool. Today's leader is Cloudflare's security-audit skill, which picked up more than 2,500 stars in a day by turning a coding agent into a careful bug hunter for your own code.
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +2,548 · 📦 Total: 63,955
📜 License: MIT · 👤 By: company (Cloudflare)
🎯 Time to value: 20 minutes
📜 License: MIT · 👤 By: company (Cloudflare)
🎯 Time to value: 20 minutes
What it is: A free "skill," an instruction pack for AI coding agents, that runs a step-by-step security review of your codebase. It maps how the app fits together, sends separate helper agents to hunt for weak spots, and has a fresh agent try to disprove every finding before it lands in the final report. Cloudflare says it seeded its own in-house bug-finding system. Why you'd want it: Get a structured, double-checked security review of your own code from the coding agent you already use, with findings sorted into confirmed, needs-checking and rejected.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license from Cloudflare | Needs an agent that can run several helper agents at once |
| Every finding is checked by a separate agent before it's reported | Asks for a locked-down sandbox, which takes setup |
| Repeat runs build on earlier ones to cover more ground | Cloudflare says one run found only about half of what repeated runs found; last featured here October 20-25 |
GitHub - cloudflare/security-audit-skill: A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings - cloudflare/security-audit-skill
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +2,070 · 📦 Total: 316,698
📜 License: MIT · 👤 By: individual developer (Matt Pocock)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: individual developer (Matt Pocock)
🎯 Time to value: 10 minutes
What it is: A popular teacher's personal set of skills for coding agents like Claude Code, Codex and GitHub Copilot, built for "real engineering - not vibe coding." The best-known one, /grill-me, has the agent quiz you about what you want before it writes anything, and others cover test-first coding, bug hunting, specs and tickets. Why you'd want it: Fix the most common agent failures, like building the wrong thing or writing untested code, with small skills you can edit to suit your team.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license, with one-line installs for several agents | Opinionated: it assumes you work with an issue tracker and tests |
| Small skills that are easy to read and change | Installing both the plugin and the copy-in version gives you every skill twice |
| Works with any AI model | Featured here several times since October 1 |
GitHub - mattpocock/skills: Skills for Real Engineers. Straight from my .agents directory.
Skills for Real Engineers. Straight from my .agents directory. - mattpocock/skills
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +1,155 · 📦 Total: 117,842
📜 License: MIT · 👤 By: startup (Stably AI)
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: startup (Stably AI)
🎯 Time to value: 15 minutes
What it is: A desktop app for Mac, Windows and Linux that runs several coding agents, such as Claude Code and Codex, side by side, each in its own copy of your project so they don't trip over each other. You can hand one agent a big task and let it split the work among helpers, review their changes line by line, and get phone alerts when an agent needs you. Why you'd want it: Manage a whole team of coding agents from one window, using the subscriptions you already pay for.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Running many agents at once burns through plan limits fast |
| Works with many agents, in a chat view or the raw terminal | Cloud machines for agents are a separate setup |
| Mobile app and remote servers keep work moving when your laptop sleeps | Last featured here October 17-23 |
GitHub - stablyai/orca: Orca lets you run a fleet of coding agents (e.g. Claude Code, Codex) with your own subscription. Available on desktop, mobile and remote runtime.
Orca lets you run a fleet of coding agents (e.g. Claude Code, Codex) with your own subscription. Available on desktop, mobile and remote runtime. - stablyai/orca
GitHub
Rank yesterday: #1 - Falling ↓
⭐ Stars today: +1,004 · 📦 Total: 56,467
📜 License: MIT · 👤 By: company (Microsoft)
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: company (Microsoft)
🎯 Time to value: 30 minutes
What it is: Microsoft's family of free voice AI models. Its speech-to-text model turns up to an hour of audio into a transcript that labels who spoke and when, and a quick text-to-speech model starts talking almost instantly. Why you'd want it: Turn long meetings, podcasts or calls into speaker-labeled transcripts with a free model you run yourself.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license from a major company | Microsoft disabled its long-form text-to-speech code in 2025 after misuse |
| Hour-long audio in one pass with speaker labels | Microsoft doesn't recommend commercial use without more testing |
| Stars today more than quadrupled versus yesterday | Repeat pick from yesterday |
GitHub - microsoft/VibeVoice: Open-Source Frontier Voice AI
Open-Source Frontier Voice AI. Contribute to microsoft/VibeVoice development by creating an account on GitHub.
GitHub
Rank yesterday: #2 - Falling ↓
⭐ Stars today: +534 · 📦 Total: 22,827
📜 License: Apache-2.0 · 👤 By: company (Google)
🎯 Time to value: 10 minutes
📜 License: Apache-2.0 · 👤 By: company (Google)
🎯 Time to value: 10 minutes
What it is: Google's official library of skills that teach AI coding agents to use Google products such as Google Cloud, Gemini and BigQuery. One command installs them for Claude Code, Codex or Google's own agent tools. Why you'd want it: Get an agent to set up Google Cloud services the way Google documents them, instead of guessing.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license, maintained by Google | Mostly useful if you already use Google Cloud |
| Stars today more than doubled versus yesterday | A large catalog, so pick skills carefully |
| Works across several agent tools | Repeat pick from yesterday |
GitHub - google/skills: Agent Skills for Google products and technologies
Agent Skills for Google products and technologies. Contribute to google/skills development by creating an account on GitHub.
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +209 · 📦 Total: 84,237
📜 License: LobeHub Community License (Apache-2.0 with extra conditions) · 👤 By: company (LobeHub)
🎯 Time to value: 30 minutes
📜 License: LobeHub Community License (Apache-2.0 with extra conditions) · 👤 By: company (LobeHub)
🎯 Time to value: 30 minutes
What it is: A self-hostable workspace that adds AI agents to the chat apps your team already uses, like Discord, where they pick up tasks on their own and learn from your corrections over time. It also has a marketplace of ready-made agents, skills and connectors you can add. Why you'd want it: Give your team an AI coworker that lives in your group chat and takes on unassigned requests, instead of a private chatbot.
| ✓ Pros | ✗ Cons |
|---|---|
| Works with many chat apps, agents and AI models | Distributing a modified version needs a commercial license |
| Self-hosting with Docker Compose or one-click deploy | The makers say it's under active development |
| Large marketplace of agents and skills | An agent that acts on its own in team chats needs clear limits |
GitHub - lobehub/lobehub: 🤯 LobeHub - Super-Agents on Your Team. Add super-agents to the chats your team already uses.They take on work by themselves and get better every week.
🤯 LobeHub - Super-Agents on Your Team. Add super-agents to the chats your team already uses.They take on work by themselves and get better every week. - lobehub/lobehub
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +198 · 📦 Total: 14,048
📜 License: MIT · 👤 By: company (Meta)
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: company (Meta)
🎯 Time to value: 30 minutes
What it is: Meta's open-source design system, a ready-made kit of more than 150 screen building blocks like buttons, tables and forms, plus themes and templates. Meta says it grew internally over eight years and powers more than 13,000 apps, and its documentation and command-line tool are built so AI coding assistants use it the same way people do. Why you'd want it: Have an AI assistant build good-looking, accessible web app screens from a proven kit instead of inventing its own styles.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license from Meta | Still in beta |
| Over 150 accessible components and seven themes | Requires React 19 or later |
| Built so people and AI assistants follow the same reference | Chart packages are only available as pre-release (canary) builds |
GitHub - meta/astryx: An open source design system that's fully customizable and agent ready
An open source design system that's fully customizable and agent ready - meta/astryx
GitHub
HuggingFace Trending
Top Models Today
Moonshot's giant Kimi K3.1 Pro jumped to first place on Hugging Face's trending chart, and Pika released a fast video-with-sound model. We skipped an unverified upload claiming to be a leaked Anthropic training checkpoint, uncensored copies of other models, and models featured in the past week that have since dropped down the chart.
Up from second place in our November 4 list to first on today's trending chart: Moonshot's 2.8-trillion-parameter agent model, with a 1-million-token memory.
📥 Downloads (30d): 18,432 · 📜 License: Kimi K3.1 Pro License (modified MIT)
👤 By: Moonshot AI · 🎯 Task: Image-text-to-text
📐 Size: 2.8T total, 104B active
👤 By: Moonshot AI · 🎯 Task: Image-text-to-text
📐 Size: 2.8T total, 104B active
What it is: Moonshot's most capable downloadable model, which reads text and images and is tuned for coding, 3D design, financial research and other agent work. On Moonshot's own tests it scores 65.7 on Terminal-Bench 4.0, a test of coding in a command window, close to Claude Opus 5.5's 66.4. Why you'd want it: A free-to-download model that competes with top paid models on several agent coding tests, also available through Moonshot's API.
| ✓ Pros | ✗ Cons |
|---|---|
| Weights are free, with mostly permissive terms | Large AI-hosting businesses earning over $20 million a year need a separate deal |
| Near the top of several coding tests | Needs data-center hardware to run |
| Works with popular serving software (vLLM, SGLang) | Benchmarks are Moonshot's own; repeat pick from November 4 |
moonshotai/Kimi-K3.1-Pro · Hugging Face
Image-Text-to-Text model with 2.8T parameters. Download, explore files and run it on Hugging Face.
Second on the trending chart: turns a single photo of an object into a 3D model you can spin around, in seconds.
📥 Downloads (30d): 19,139 · 📜 License: Apache-2.0
👤 By: Sperid Labs · 🎯 Task: Image-to-3D
📐 Size: about 3 GB of weights
👤 By: Sperid Labs · 🎯 Task: Image-to-3D
📐 Size: about 3 GB of weights
What it is: Give Idol one picture of an object and it removes the background and builds a 3D "Gaussian splat," a cloud of up to a million tiny colored blobs, including the sides the photo never showed. It runs on a consumer graphics card using 7.2 GB of video memory and can finish in under 2 seconds. Why you'd want it: Quickly turn product photos or AI-generated images into 3D objects for games, shops or design mockups.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive Apache-2.0 license | Needs access to two gated models from Meta and Black Forest Labs, with their own terms |
| Runs on a consumer graphics card | Needs an NVIDIA graphics card |
| Very fast: under 2 seconds possible | Last featured here October 31 |
speridlabs/idol · Hugging Face
Image-to-3D model. Download, explore files and run it on Hugging Face.
Eighth on the trending chart: NVIDIA's speech-to-text model, which runs fully offline.
📥 Downloads (30d): 61,049 · 📜 License: OpenMDW 1.1
👤 By: NVIDIA · 🎯 Task: Automatic speech recognition
📐 Size: 34B total, 4.3B active
👤 By: NVIDIA · 🎯 Task: Automatic speech recognition
📐 Size: 34B total, 4.3B active
What it is: A model that turns recordings of up to an hour into text in 40 languages, and can also translate speech, label speakers and add timestamps. NVIDIA reports the best published score on a test of multi-speaker meetings, with an error rate of 15.80% versus 16.83% for the next best. Why you'd want it: Transcribe long, messy meetings in many languages with a model that's cleared for commercial use.
| ✓ Pros | ✗ Cons |
|---|---|
| NVIDIA says it's ready for commercial use | Built for NVIDIA graphics cards |
| Handles hour-long recordings, speakers and timestamps | Offline only: not built for live captions |
| Smaller compressed versions are available | Benchmarks are NVIDIA's own; last featured here October 30 |
nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B · Hugging Face
Automatic Speech Recognition model with 34B parameters. Download, explore files and run it on Hugging Face.
New to our list: Pika's downloadable video model that makes clips with matching sound in just six steps.
📥 Downloads (30d): 986 · 📜 License: MiniMax H3 Community License
👤 By: Pika Labs · 🎯 Task: Image-to-video
📐 Size: 33.1B (base model plus add-ons)
👤 By: Pika Labs · 🎯 Task: Image-to-video
📐 Size: 33.1B (base model plus add-ons)
What it is: A video model built on MiniMax's H3 that creates 5 to 15 second clips with synchronized stereo sound from a text prompt, a first or last frame, or up to 12 reference images, videos and audio clips. It makes 480p or 768p video locally, and Pika's paid API can upscale to 4K. Why you'd want it: Draft multi-shot scenes, product films or motion graphics with sound on your own graphics card, using reference images to keep characters consistent.
| ✓ Pros | ✗ Cons |
|---|---|
| Video and matching audio in one pass | Uses MiniMax's custom license, not a standard open one |
| Up to 12 reference files to guide the look | Higher resolutions need Pika's paid API |
| Runs one job on one graphics card | Pika says its showcase clips aren't guaranteed to reproduce exactly |
Pika-Labs/Pika-H3-Flex · Hugging Face
Image-to-Video model with 33B parameters. Download, explore files and run it on Hugging Face.
Fourteenth on the trending chart: Mistral's largest downloadable model, still drawing attention a week after we featured it.
📥 Downloads (30d): 10,530 · 📜 License: Mistral Ecosystem Open-Weight License (modified MIT)
👤 By: Mistral AI · 🎯 Task: Image-text-to-text
📐 Size: about 1T total, 52B active
👤 By: Mistral AI · 🎯 Task: Image-text-to-text
📐 Size: about 1T total, 52B active
What it is: A 1-trillion-parameter model that reads text and images, trained from scratch in Mistral's own European data centers and fluent in more than 160 languages. Mistral says it scores 48% on Artificial Analysis's coding-agent index, roughly tied with the top downloadable model, GLM-5.3, while using 47% fewer tokens per task. Why you'd want it: A strong European option for coding agents and multilingual work that you can host yourself or rent cheaply through Mistral's API.
| ✓ Pros | ✗ Cons |
|---|---|
| Commercial and non-commercial use allowed | Running it as a paid AI-hosting service needs a separate license |
| 1M-token context and adjustable reasoning effort | Needs data-center hardware at full size |
| Mistral's API charges $0.70 in / $2.10 out per million tokens | Last featured here October 29 |
mistralai/Mistral-Large-4-1T-A52B · Hugging Face
Image-Text-to-Text model. Download, explore files and run it on Hugging Face.
Sixteenth on the trending chart: a tiny agent model that fits in about 1 GB of memory.
📥 Downloads (30d): 66,003 · 📜 License: LFM Open License v1.0
👤 By: Liquid AI · 🎯 Task: Text generation
📐 Size: 1.73B
👤 By: Liquid AI · 🎯 Task: Text generation
📐 Size: 1.73B
What it is: A small text model built for running agents and tools on phones and laptops, with a 256,000-token memory. When compressed, it takes about 1 GB of memory, and it's designed to be retrained quickly on a single graphics card for your own tasks. Why you'd want it: Run a private, on-device assistant that can call tools and fill in forms without sending data to the cloud.
| ✓ Pros | ✗ Cons |
|---|---|
| Fits on almost any phone or laptop | Commercial use is limited for companies with $10 million or more in yearly revenue |
| Trained for popular agent tools | Liquid AI says it's not suited to agent coding or complex reasoning |
| Quick to retrain on one graphics card | Last featured here October 28 |
LiquidAI/LFM3-1.7B · Hugging Face
Text Generation model with 2B parameters. Download, explore files and run it on Hugging Face.
Product Hunt
AI Launches Today
The rest of Thursday's top AI launches on Product Hunt focused on keeping AI agents in line at work: approving what they do, bringing them into meetings and keeping their output on brand. Upvote counts are Product Hunt's scores when we checked, so final counts may be higher.
See who said yes to your team's AI agents
🔥 Upvotes: 194 · 👤 By: Pushary (pushary.com); maker Aadil Ghani
💰 Pricing: $59 a month for 5 people; 3-day free trial · 🏷 Category: Productivity, AI coding agents
💰 Pricing: $59 a month for 5 people; 3-day free trial · 🏷 Category: Productivity, AI coding agents
Sends an AI coding agent's permission requests and questions to whoever is running it, on their Mac's notch or their phone, and records who approved each one. The team version adds rules, such as sending infrastructure changes to a team lead, and one page showing which agent asked and who said yes. It works with Claude Code, Codex, Cursor, Gemini CLI and others. Verdict: Useful for teams that need a shared record of what their agents were allowed to do; solo users can get phone approvals free inside Claude Code and Codex, as the maker notes.
View on Product Hunt →
Video conferencing for the agentic era
🔥 Upvotes: 144 · 👤 By: Bridge (meetonbridge.com, formerly Simplora); founder Jimmy Lowery Jr
💰 Pricing: Free options; paid prices not listed · 🏷 Category: Video conferencing, AI workflow automation
💰 Pricing: Free options; paid prices not listed · 🏷 Category: Video conferencing, AI workflow automation
A meeting app that combines scheduling, meeting prep, video calls, live help, AI notes and a memory of your contacts and companies. Its newest feature lets custom AI agents join calls or work quietly in the background, and a connector lets your existing AI tools pull in meeting context. Verdict: Worth a look if you're tired of juggling separate meeting, note-taking and scheduling tools; check how recordings and notes are stored before inviting agents into client calls.
View on Product Hunt →
The brand layer for agents
🔥 Upvotes: 112 · 👤 By: Rainbrand (Y Combinator); founder Ray Fitzgerald
💰 Pricing: 1,000 free signup credits; paid prices not listed · 🏷 Category: Marketing automation, AI generative media
💰 Pricing: 1,000 free signup credits; paid prices not listed · 🏷 Category: Marketing automation, AI generative media
Builds a shared "brand kit" from your website, brand guides and files, so AI tools like Claude, ChatGPT, Codex and Cursor stay on brand when making campaigns, web pages, slides, emails and videos. Brand versions are tracked, so you can roll back changes, and developers can build it into their own products through an API. Verdict: A handy fix for off-brand AI output if your team uses several AI tools; pricing beyond the free credits isn't listed yet.
View on Product Hunt →
Your AI workspace for mail, calendar, notes, tasks and more
🔥 Upvotes: 103 · 👤 By: Stitch.ai (stitch.ai); makers Johanna Quinn and Brian MacDonald
💰 Pricing: Free options and 7-day trial; paid prices not listed · 🏷 Category: Productivity, email clients, AI agents
💰 Pricing: Free options and 7-day trial; paid prices not listed · 🏷 Category: Productivity, email clients, AI agents
Puts email, calendar, contacts, tasks, notes and files into one linked workspace, with an AI assistant called Stitchy that works across all of them. You can drag items between panels to connect them, and AI tags keep things organized. Verdict: Interesting for people who live in their inbox, but early: the AI assistant isn't in the free trial, and connecting your existing email accounts is still on the roadmap.
View on Product Hunt →
An AI on-call engineer for your Google Cloud apps
🔥 Upvotes: 98 · 👤 By: Polylane (polylane.com); maker Matanya Loewenthal
💰 Pricing: Free options; paid prices not listed · 🏷 Category: AI engineer, developer tools
💰 Pricing: Free options; paid prices not listed · 🏷 Category: AI engineer, developer tools
An AI tool that connects your code, cloud setup and monitoring data to investigate problems in apps running on Google Cloud. It explains what went wrong and what's affected, then prepares a fix for a person to review, and the maker says access can stay read-only. Verdict: A sensible way to speed up late-night troubleshooting, since a human still approves every fix; how well it handles complex outages isn't shown yet.
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 still isn't listed on Google's developer pricing page. Newly noticed (we didn't track it before, so it may not be new): Groq lists MiniMax M2.7 as a preview model with a 196,608-token context, priced through its sales team.
What this means: Prices are frozen, so the deadlines are what matter: Gemini 4 Argon doubles on December 1, and the promotional price for OpenAI's older GPT-5.6 Sol security model (not in the table above) ends November 30. If you plan to switch models, run your cost tests before the end of the month.
No price changes vs. yesterday - every standard price from Anthropic, OpenAI, Google, Groq and Mistral matches yesterday's table. Gemini 4 Carbon still isn't listed on Google's developer pricing page. Newly noticed (we didn't track it before, so it may not be new): Groq lists MiniMax M2.7 as a preview model with a 196,608-token context, priced through its sales team.
What this means: Prices are frozen, so the deadlines are what matter: Gemini 4 Argon doubles on December 1, and the promotional price for OpenAI's older GPT-5.6 Sol security model (not in the table above) ends November 30. If you plan to switch models, run your cost tests before the end of the month.
arXiv Paper of the Day
How Do AI Agents Waste Your Money?
Huimin Zeng, Yuxuan Lu, Jesse Cai, Eric Hanchen Jiang, Di Wu, Po-Nien Kung, Wenbo Hu, Ziyi Wang, Dakuo Wang, Jiaxin Pei, Kai-Wei Chang · arXiv:2611.07238
What it claims: AI agents use huge numbers of tokens (the chunks of text you pay for), but nobody had a standard way to measure how many were actually needed. The authors built a method that records an agent's full work log, uses another AI model as a judge to mark the parts that could be cut, and then re-runs the agent without them to check that results hold up.
Key finding: Cutting 41-46% of an agent's reasoning text largely preserved task performance on two different agent tests, and waste varied more by task and agent setup than by which AI model was used.
Why practitioners should care: If your agent bills are high, look at the task and the agent harness (the software wrapped around the model) before switching to a cheaper model. Different harnesses waste tokens in different ways, even with the same model on the same task.
Key finding: Cutting 41-46% of an agent's reasoning text largely preserved task performance on two different agent tests, and waste varied more by task and agent setup than by which AI model was used.
Why practitioners should care: If your agent bills are high, look at the task and the agent harness (the software wrapped around the model) before switching to a cheaper model. Different harnesses waste tokens in different ways, even with the same model on the same task.
×
Click anywhere or press ESC to close
Member discussion