Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
1
trillion parameters in total, with about 52
Mistral releases its 1-trillion-parameter Large 4 as a free
Top Story
One Thing to Tell Your Friends
One Thing to Tell Your Friends
When researchers traced 1,285 statistics quoted by AI chatbots back to their sources, only about 4 in 10 led to a real study that actually said that number.
Summary
TL;DR
Top Stories
Trends
Creative AI
Gemini can now write and edit documents, spreadsheets and slides inside the chat, Gemini adds zoomable "immersive explorations" for learning, and Meta quietly launched Style Sketch, an AI stylist for clothes you already own.
Dev Tools
OpenAI now gives ChatGPT subscribers free API credits, Hugging Face releases free speech and decision tools for Apple apps, and IBM's Granite Switch 4.1.1 upgrades an AI model like a software patch.
Research
Business
Education
Surprising
50 AI agents built a Minecraft society, and it got bureaucratic fast, OpenAI explains its model names: Luna, Terra, Sol and Astra, and Debate: can you spot "LLM Voice," and should you ever let AI write your drafts?.
Worth Watching
GitHub
HuggingFace
Leading models: deepseek-ai/DeepSeek-V4.1-Flash (131,062), Nanbeige/Nanbeige4.3 (9,655), and mistralai/Mistral-Large-4-1T (278).
Product Hunt
API Pricing
No price changes vs. yesterday - Anthropic's, OpenAI's and Groq's official pricing pages are word-for-word the same as yesterday, and every model price on Google's page is unchanged.
arXiv
Does More Thinking Pay Off? Reasoning Cost and Correctness of LLMs in Code Generation and Bug Fixing — On one common code-writing test, turning reasoning on raised the Qwen3-32B model's success rate only from 68.1% to 69.5%, a gain too small to be statistically meaningful, even though it spent a median of more than 2,000 thinking tokens per task.
FYI
Hot off the Presses
01
Mistral releases its 1-trillion-parameter Large 4 as a free download
What this means for you: Companies that want a powerful AI they fully own and run themselves now have a strong option from Europe, at a fraction of the price of top US models.
Previously: October 6 - Mistral released Mistral Large 4 as a preview and said the free download would follow after safety testing.
Today: Mistral, a French AI company, released the model's weights (the files anyone can download to run it). It spent about three more weeks training the model after the preview. Mistral nicknames it "Le Chonk" because of its size. It reads text and images, works in more than 160 languages and can handle about 1 million tokens (word pieces) at once.
Independent tests put its coding scores below Mistral's own claims, according to VentureBeat. Mistral also says it is tightening guardrails because the model is strong at computer-security tasks.
“63% cheaper per task than the preview.”
- About 1 trillion parameters in total, with about 52 billion active at a time - a design called Mixture of Experts (MoE), where only a small part of the model works on each request
- Independent firm Artificial Analysis scores it 26 on its intelligence index - the best downloadable model made outside China, but six points behind the top Chinese downloadable models
- The cost of each test task fell 63% - from $5.54 for the preview to $2.06, mostly because it now writes far fewer words while thinking
- Its API price is $0.70 per million input tokens and $2.10 per million output tokens - API means the paid connection developers use to plug a model into their apps
- The license is free for most uses - only companies that resell it as a hosted AI service need a separate commercial deal
02
Alphabet has its worst day ever as investors fear AI is eating Google Search
What this means for you: The ads you see in Google's AI answers are about to multiply. Google is under heavy pressure to make its AI search earn as much as classic search.
Previously: October 28 - Alphabet missed revenue and profit forecasts and its shares fell more than 10% after hours.
Today: Shares of Alphabet, Google's parent company, fell about 13% on Thursday. It was the biggest one-day drop since Google went public in 2004. Search ads grew 12% to $63.35 billion but still missed forecasts.
That was the verdict of Ben Barringer, an analyst at investment firm Quilter Cheviot. Another analyst, Scott Devitt of Rosenblatt Securities, argued Google's roughly 4 billion search users still give it a head start in AI.
“A fairly messy set of results and a poorly handled earnings call.”
- About $500 billion in stock-market value disappeared in one day - per CNBC
- Free cash flow was negative $9.07 billion for the quarter - meaning Google spent far more than it brought in, mainly on AI data centers
- Google's business chief blamed weaker ad earnings inside AI Mode - Google's chat-style search, for part of the search slowdown
- AI Mode ads are now expanding to 10 more countries - including Australia, Canada and India, with Brazil, Mexico and the UK planned before year end
- In the US, ads can now appear in follow-up questions - not just after your first search, and they stay labeled "Sponsored"
03
AI tutors matched human tutors in a big UK classroom trial, but test scores didn't move
What this means for you: If your child's school adds an AI tutor, expect it to work about as well as a human tutor on practice problems. Don't expect it to raise standardized test scores on its own, at least not yet.
Eedi, a London education startup, ran the trial with Google DeepMind, Google's AI research lab. About 1,179 students aged 12 to 15 in 10 UK state schools practiced math on Eedi once a week for 10 weeks. When a student got a question wrong, they got either a standard on-screen hint, a human tutor by text chat, or an AI tutor built on Google's Gemini, checked by a human before each message was sent.
Last year's smaller study had the AI slightly ahead of human tutors. That edge disappeared at scale. Google says it will follow students for a full school year.
- Any tutoring raised the share of follow-up questions answered correctly to 56% to 58% - versus 43% with the standard hint, and AI and human tutors did about equally well
- Neither kind of tutoring beat plain practice on an independent math test - possibly because tutored students worked through fewer problems, about 14 per session versus about 19
- Human reviewers approved 94.7% of the AI's messages unchanged - but 7 messages containing false math still reached students, per The 74
- The AI guided students instead of giving answers in nearly 99% of sessions - according to Google
- A larger trial with more than 15,000 students began in September - this time the AI works alone and humans step in only when needed
04
Most statistics that AI chatbots quote can't be traced to a real study
What this means for you: When a chatbot gives you a percentage, treat it as a lead to check, not a fact. Six times out of ten, the trail goes cold or leads somewhere else.
Trakkr, a firm that tracks how brands appear in AI answers, asked ChatGPT, Perplexity, Gemini, Claude and Google's AI Overviews 200 questions that called for numbers. Researchers then followed 1,285 of the quoted figures up to five links deep to find where they really came from. They checked many trails by hand because an AI judge alone got the verdicts badly wrong.
That claim is a zombie stat. A 2025 MIT study found 95% of organizations saw no return from generative AI. Chatbots routinely retold it as 95% of AI pilots failing, which the study does not say.
“95% of pilots fail.”
- About 40% of statistics led back to a study that said the same thing - another 8% reached the study, but the number or its meaning had changed
- 40% hit a dead end - a locked, deleted or silent page, and 11% led to no study at all
- ChatGPT was the best sourced at 57% - while Claude quoted the most figures, about 6.8 per answer, with 38% sourced
- Agreement between chatbots did not mean a number was real - four figures quoted by all five assistants had no study behind them
- More than half of the pages chatbots cited were company blogs - only 21% were original research
05
An OpenAI-backed company is buying home-security giant ADT for $15.3 billion
What this means for you: Expect more familiar companies, from security to accounting, to be bought and rebuilt around AI. Your next home alarm may come with an AI layer you didn't ask for.
Thrive Holdings was spun out of Thrive Capital, the investment firm run by Josh Kushner. It buys traditional companies, keeps them for the long term and rebuilds them with AI. OpenAI is both an investor and a partner, supplying models, technology and engineers who work inside the companies Thrive owns.
- Thrive will pay $10 per ADT share - valuing ADT's stock at about $7.6 billion, or $15.3 billion including debt
- Thrive Holdings already owns more than 75 companies - and raised $2 billion in August to bring AI to businesses
- ADT will become the foundation of Thrive's smart-home security business - though no specific AI products have been announced
- Home security is getting crowded with AI - Apple and Amazon have both pushed into the market this year
Trends & Themes
Trends & Themes
Wall Street is punishing companies that look exposed to AI
Why this matters to you: Retirement accounts and index funds hold these companies. AI is now moving their stock prices more than their actual results are.
Alphabet's record drop (see Top Stories) fits the same pattern. Investors are rewarding only companies that look like AI winners and selling anything that merely meets expectations. ServiceNow sells workflow software to big companies, and investors fear AI agents will do that work instead.
- ServiceNow fell 27%, its worst day on record - even though sales grew more than 20% and its new AI contracts jumped 65%
- Reddit plans to shift to a mostly logged-in site - after visits sent from search engines dropped by more than 1 million a week in October
- Reddit's US daily users fell to 50.7 million - down from 53.2 million in the previous quarter
- Intel fell more than 20% in after-hours trading - its AI data-center sales rose 78% but its PC chip business shrank
AI money is still flowing, but it costs more and arrives later
Why this matters to you: When building AI gets more expensive, the free and cheap AI tools you use today may get pricier or slower to improve.
Nobody is pulling back on AI. But lenders are asking for more interest, and chip deliveries are slipping into 2027. Gartner, a research firm, found that companies are not yet saving money from AI.
- Microsoft cut its full-year spending forecast to $170 billion from $175 billion - blaming delays at outside providers that rent it graphics chips
- The riskiest slice of Apollo's $40 billion chip lease for SpaceX priced at 3.95 percentage points over government bonds - more than the roughly 3.5 points expected a week earlier
- Gartner says 97% of organizations can't cut costs through productivity gains alone - and 38% are already over their AI budgets
- Amazon Web Services still grew 39% - its fastest growth since 2021, showing demand for AI computing remains strong
Workers and voters are asking for AI rules before the midterms
Why this matters to you: AI protections at work and online are becoming an election issue, so expect candidates in both parties to court your vote on it.
Unions, voters and employees are pointing at the same worry from different angles. They want a say in how AI changes work before the changes are locked in. Dayforce calls the hidden cost "experience debt," meaning skills people never build because AI does the task.
- The AFL-CIO is drafting its own AI bill - the largest US federation of labor unions calls it its top priority, covering worker protections and privacy
- Ten proposed AI safety rules each won more than 70% support - among voters in Ohio and Iowa, in a poll by a firm run by a Republican pollster
- Six in 10 workers say they are losing skills because they rely on AI - per a survey by Dayforce, a payroll and HR software company
- 40% of workers expect AI-saved time to bring them more work - rather than better jobs or less burnout
Governments now treat AI companies as strategic assets
Why this matters to you: Who owns the platforms where AI models are shared, and which country they answer to, will shape which AI tools you can use and what they cost.
Governments are no longer just funding or regulating AI. They are buying stakes, reviewing deals and treating model outputs as something to protect. Copying a stronger model's answers to train your own is called distillation, and it is now a contest between rival nations.
- The US Justice Department is reviewing Nvidia's planned $12.93 billion purchase of Hugging Face - the biggest website for sharing AI models, according to Reuters
- Canada will take an ownership stake of up to $1.1 billion in Toronto AI company Cohere - according to Semafor and Reuters
- Chinese AI labs reportedly shifted to copying OpenAI's GPT-6 Astra - after Anthropic says its defenses cut what they could extract from Claude by 90% to 99%
- France's competition authority is investigating Nvidia separately - over the industry's dependence on its software and its stakes in cloud companies
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Business & Industry
Business & Industry
Education
GenAI in Education
Surprising
Surprising & Under-the-Radar
Worth Watching
Signals to Track
01
AI agents need flight recorders
Companies are letting AI agents act on their behalf without a reliable record of what those agents did.
The Partnership on AI, a nonprofit coalition of tech companies, academics and civil society groups, studied the four most widely used tools for building AI agents. None reliably logged basic facts such as which agent acted or which human approved a task. The authors want standard "what happened" records built in, much like a plane's black box. If this plays out, you could one day get a clear explanation when an AI agent denies your loan or changes your account.
02
A game studio spends 30 times its payroll on AI
One startup lets AI agents ship changes to its live game every hour, straight to the live game.
Spawn is a multiplayer game where players reshape the world by typing requests to an in-game AI. Its founder says a swarm of agents running Anthropic's Fable models builds most of it, for over $3 million a month in AI costs. An hour without shipping an update counts as the company's most serious kind of incident. If this works, more apps and games you use will change daily, built mostly by AI.
03
Amazon will sell its AI training chips outside its own cloud
Amazon plans to offer its Trainium chips outside its own cloud starting in late 2027.
Until now, Trainium has been available only inside Amazon Web Services. Google has already started offering its own AI chips outside its cloud. More competition for Nvidia's chips could lower what it costs companies to build AI, which eventually shapes what you pay for AI tools.
GitHub Trending
Top Repos Today
GitHub's main trending page listed eleven projects today, all of them AI-related. We took four of them and added three from the Python trending list. We left out a list of AI "jailbreak" prompts, JetBrains' Thinkrail coding editor (covered in an earlier issue) and, from the Python list, a tool for stripping the safety limits out of AI models and one project unrelated to AI. We also skipped five projects covered in recent issues to make room, including the AI tutor DeepTutor, which topped both lists again with no new release since yesterday, and the design tool impeccable, the day's biggest gainer. None of the seven picks was on yesterday's list, and two of them give AI agents a memory that lasts between sessions, led by Hindsight, the biggest gainer among today's picks.
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +1,352 · 📦 Total: 56,186
📜 License: MIT · 👤 By: organization (Vectorize)
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: organization (Vectorize)
🎯 Time to value: 15 minutes
What it is: A memory system for AI agents that aims to help them learn over time, not just recall old chats. It sorts what it stores into facts about the world, the agent's own experiences and "mental models" it builds from both, and your agent stores, searches and reflects on them with three simple commands. It works with more than 25 AI providers, including models running on your own computer, and plugs into coding agents such as Claude Code, Codex, Cursor and GitHub Copilot, where it builds a memory for each project from its history. Why you'd want it: Give an agent a long-term memory with a two-line code change, instead of building your own search over old conversations.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Every memory it saves or reflects on uses AI calls you pay for |
| Its benchmark scores were independently reproduced by Virginia Tech researchers and The Washington Post | Self-hosting means running a server and a database |
| Can use an existing Claude, ChatGPT, Cursor or Copilot subscription instead of an API key | The hosted version is billed by use, with no listed flat price |
GitHub - vectorize-io/hindsight: Hindsight: Agent Memory That Learns
Hindsight: Agent Memory That Learns. Contribute to vectorize-io/hindsight development by creating an account on GitHub.
GitHubRank yesterday: not listed - New entry 🆕
⭐ Stars today: +1,190 · 📦 Total: 7,430
📜 License: Apache-2.0 · 👤 By: individual developer (Milind Soni)
🎯 Time to value: 15 minutes
📜 License: Apache-2.0 · 👤 By: individual developer (Milind Soni)
🎯 Time to value: 15 minutes
What it is: A free desktop chat app where every contact in your sidebar is an AI bot with its own personality, model, computer and connected apps. The bots run on the Claude, Codex and Grok tools already installed on your machine, using your existing logins, and each can work on a cloud desktop, an isolated virtual machine or (after you opt in) your own computer. Risky steps such as running commands or editing files appear as cards you approve or deny. It shipped two releases today; in the first, a bot that hits its AI's usage limit now carries on with another AI you have set up and switches back when the limit resets. Why you'd want it: An open-source take on the "team of AI bots in a chat app" idea, without handing your data to one company's cloud.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on Mac, Windows and Ubuntu | Very young project (version 0.1) that changes daily |
| Uses the AI subscriptions you already pay for | Controlling your own computer is off on some Linux setups |
| Bots ask before they act | Its 500+ app connections run through a third-party service, Composio |
GitHub - milind-soni/OpenMausBot: Open-source Grok Bot alternative with a virtual machine that bots can use
Open-source Grok Bot alternative with a virtual machine that bots can use - milind-soni/OpenMausBot
GitHubRank yesterday: not listed - New entry 🆕
⭐ Stars today: +607 · 📦 Total: 75,587
📜 License: MIT · 👤 By: individual developer (Rohit Ghumare, a developer advocate based in the UK)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: individual developer (Rohit Ghumare, a developer advocate based in the UK)
🎯 Time to value: 10 minutes
What it is: A free course for learning to build AI systems by writing the pieces yourself: 542 lessons in 20 phases, from model basics to retrieval and agents, in Python, TypeScript, Rust and Julia. You can learn on its website, with a coding agent as your tutor, or by running the code locally, and the guide is translated into more than 30 languages. Today's release adds a "lesson screen" that opens a browser tab beside your terminal while an AI tutor teaches, plus starter manuals for the vLLM and Miles tools and 47 practice tasks. Why you'd want it: A structured, hands-on path into AI engineering that costs nothing and keeps your working code.
| ✓ Pros | ✗ Cons |
|---|---|
| Free and open source | 542 lessons is a long road |
| Choose a route: models, AI apps or agents | Learning with a coding-agent tutor uses paid AI time |
| Updated often, with a new edition today | Maintained by one person |
GitHub - rohitg00/ai-engineering-from-scratch: Learn AI Engineering! Learn it. Build it. Ship it for others.
Learn AI Engineering! Learn it. Build it. Ship it for others. - rohitg00/ai-engineering-from-scratch
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +538 · 📦 Total: 29,112
📜 License: MIT · 👤 By: individual developer (Donchitos, Melbourne, Australia)
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: individual developer (Donchitos, Melbourne, Australia)
🎯 Time to value: 15 minutes
What it is: A template that turns one Claude Code session into a full game studio of 49 AI agents, from a creative director and producer down to level designers, sound designers and testers, with 74 slash commands for each stage of making a game. It supports the Godot 4, Unity and Unreal Engine 5 game engines, checks work automatically on commits, pushes and file changes, and will not close a visual task until the game has been launched and a screenshot kept. Refreshingly, its own test found that the "standard" setting wrote 30 planning documents before any game code, and reviewers ranked that game last. Why you'd want it: Structure and a second opinion for solo game makers who build with AI, with a "minimal" mode when you just want to make the game.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Requires Claude Code |
| Works with the three most popular game engines | Its three directors are set to Anthropic's pricier Opus model |
| Honest about when more process does not help | No new release since October 19 |
GitHub - Donchitos/Claude-Code-Game-Studios: Turn Claude Code into a full game dev studio — 49 AI agents, 72 workflow skills, and a complete coordination system mirroring real studio hierarchy.
Turn Claude Code into a full game dev studio — 49 AI agents, 72 workflow skills, and a complete coordination system mirroring real studio hierarchy. - Donchitos/Claude-Code-Game-Studios
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +459 · 📦 Total: 104,726
📜 License: Apache-2.0 · 👤 By: individual developer (Alex Newman)
🎯 Time to value: 5 minutes
📜 License: Apache-2.0 · 👤 By: individual developer (Alex Newman)
🎯 Time to value: 5 minutes
What it is: A plugin that records what your coding agent does, has an AI compress it into short summaries and feeds the useful parts back into future sessions, so the agent remembers your project after you close it. It started with Claude Code and now also works with OpenCode, Codex and several other agent tools. It shipped two releases in under 24 hours: one now hides values stored under names like "token," "secret" and "api_key," and today's makes its "My Work" view remember your list view, filter and sort between visits. Why you'd want it: Stop re-explaining your project to your coding assistant every morning.
| ✓ Pros | ✗ Cons |
|---|---|
| One-command install | The installer steers you toward a hosted memory service with a 30-day free trial |
| Search past work in plain English | Summarizing uses AI calls from your plan or your own key |
| Works with many agent tools, not just Claude Code | Releases arrive almost daily |
GitHub - thedotmack/claude-mem: Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More
Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with…
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +318 · 📦 Total: 5,135
📜 License: MIT · 👤 By: individual developer (Beenu Arora, CEO and co-founder of security firm Cyble)
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: individual developer (Beenu Arora, CEO and co-founder of security firm Cyble)
🎯 Time to value: 30 minutes
What it is: A free, self-hosted AI security operations center (SOC), the team and tools a company uses to watch for and respond to cyberattacks. It collects alerts from your security tools, checks them against a library of 7,015 detection rules (2,530 of them ready to run), groups related alerts into incidents and has AI agents investigate each one, recording every prompt and step. It only proposes responses, and a person approves before anything is acted on. A small AI model (about 2 GB) comes built in, so it runs with no API keys. Why you'd want it: A way for small security teams to try AI-assisted alert triage on their own servers before buying a commercial product.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs with no API keys or cloud account | Needs 8 GB of memory and 20 GB of disk just for the core |
| Every AI decision is recorded for review | The built-in model is small, so verdicts need a human check |
| A smoke test proves the pipeline works end to end | Complex system with many moving parts |
GitHub - beenuar/AiSOC: Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.
Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API key…
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +240 · 📦 Total: 8,097
📜 License: MIT · 👤 By: organization (CopilotKit)
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: organization (CopilotKit)
🎯 Time to value: 30 minutes
What it is: A starting template for a company's own private AI assistant, where each AI "coworker" gets its own browser, files and only the tools you grant. Every action passes through one gateway that approves it before it happens and records it after, and you can watch a coworker's screen and take over when it reaches something it should not do alone. It runs on your own servers and works with agents built on many popular frameworks, through an open standard called AG-UI. Why you'd want it: A head start for teams that want ChatGPT-style coworkers they control, with an audit trail built in.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license | Beta, with rough edges by its makers' own account |
| Runs entirely on your own infrastructure | Needs a CopilotKit Intelligence license (a free plan exists) |
| Not tied to one agent framework or AI model | A template for developers to adapt, not a ready-made app |
GitHub - CopilotKit/OpenBot: Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent.
Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent. - CopilotKit/OpenBot
GitHubHuggingFace Trending
Top Models Today
DeepSeek's coding model held first place on Hugging Face's trending list again, and Nanbeige's agent model moved up to second. The day's big arrival is Mistral Large 4, a 1-trillion-parameter model that Mistral made free to download today. Two more models join our list: a small community coding model for 3D scenes and Liquid AI's first model that can see, listen and talk. We left out NVIDIA's new speech-to-text model, covered yesterday. To make room we dropped Recursive's flint, covered every day since October 22, plus models covered yesterday, and we skipped copies with safety limits removed.
Still first on the trending list, with 30-day downloads up 12% in a day.
📥 Downloads (30d): 131,062 · 📜 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 that improves reasoning, coding, agent and visual tasks while finishing tasks faster. 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 |
| A compressed community version is seventh on the list, with about 1.58 million downloads | Benchmarks are self-reported |
| Built for coding and agent tasks | No ready-made chat template file included |
deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Face
Image-Text-to-Text model with 763B parameters, built with Transformers. Download, explore files and run it on Hugging Face.
Up from third to second on the trending list, with downloads up 30% in a day.
📥 Downloads (30d): 9,655 · 📜 License: Apache-2.0
👤 By: Nanbeige · 🎯 Task: Text generation
📐 Size: 33.1B
👤 By: Nanbeige · 🎯 Task: Text generation
📐 Size: 33.1B
What it is: A mid-size model built for agent work such as coding and office tasks, which its makers also use as a "teacher" to train smaller models. Nanbeige reports it beats Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on most of its agent coding tests, including 65.5 on the DeepSWE test against 42.5 for Qwen3.8-27B. Why you'd want it: A capable mid-size agent model with a permissive license.
| ✓ Pros | ✗ Cons |
|---|---|
| Very permissive license | Benchmarks are the maker's own |
| Strong on coding and office-task tests | Trails Gemma4-31B on one science test |
| English and Chinese | Qwen3.8-27B still wins a few tests, including SWE-Bench Pro |
Nanbeige/Nanbeige4.3-33B · Hugging Face
Text Generation model with 33B parameters, built with Transformers. Download, explore files and run it on Hugging Face.
New to our list: fourteenth on the trending list on the day Mistral released the download.
📥 Downloads (30d): 278 · 📜 License: Mistral Ecosystem Open-Weight License (modified MIT)
👤 By: Mistral AI · 🎯 Task: Image-text-to-text
📐 Size: 1T (about 52B active)
👤 By: Mistral AI · 🎯 Task: Image-text-to-text
📐 Size: 1T (about 52B active)
What it is: Mistral's largest model, nicknamed "Le Chonk," trained from scratch in Mistral's own European data centers: it reads text and images, handles up to 1 million tokens and lets you pick low, medium or high reasoning effort per request. Mistral says it ties the leading downloadable model, GLM-5.3, on Artificial Analysis's coding-agent index (48%) while using 47% fewer tokens per task, and that it refused 96.6% of malicious hacking requests in its tests. Why you'd want it: The best downloadable model for coding, per Mistral, with a cheap API option if you cannot host it.
| ✓ Pros | ✗ Cons |
|---|---|
| Free for commercial use | Selling it as a hosted AI service needs a separate license |
| Fluent in more than 160 languages | Its setup examples assume a server with 8 data-center graphics cards |
| Also on Mistral's API at $0.70 / $2.10 per million tokens | Chinese downloadable models still score higher on an overall intelligence index |
mistralai/Mistral-Large-4-1T-A52B · Hugging Face
Image-Text-to-Text model, built with Transformers. Download, explore files and run it on Hugging Face.
New to our list: sixteenth on the trending list, a one-person experiment.
📥 Downloads (30d): 12,237 · 📜 License: Apache-2.0
👤 By: CCORE (independent developer, built on Qwen3.5-4B) · 🎯 Task: Coding (3D scenes)
📐 Size: 4.5B
👤 By: CCORE (independent developer, built on Qwen3.5-4B) · 🎯 Task: Coding (3D scenes)
📐 Size: 4.5B
What it is: A small coding model trained to build interactive 3D scenes for web pages with the popular Three.js library, such as temples, gardens and furniture, from a text prompt. Its maker calls it an early preview; today they published the full weights, the 4,146-example training dataset and a technical report. Why you'd want it: A free model small enough for a laptop that can sketch 3D web scenes.
| ✓ Pros | ✗ Cons |
|---|---|
| Small enough for an ordinary computer | Early preview that is still being trained |
| Training data and report are public | Not an official Qwen release |
| Permissive license | Some scenes still need a second try, per its maker |
ccore/Qwen3.5-4B-3D-Coder · Hugging Face
Machine learning model with 5B parameters. Download, explore files and run it on Hugging Face.
New to our list: nineteenth on the trending list, Liquid AI's first "see, hear and speak" model.
📥 Downloads (30d): 1,814 · 📜 License: LFM Open License v1.0
👤 By: Liquid AI · 🎯 Task: Any-to-any
📐 Size: 2.17B
👤 By: Liquid AI · 🎯 Task: Any-to-any
📐 Size: 2.17B
What it is: One small model that understands text, images, video and audio, and answers in text or speech, meant to replace a separate speech-to-text, image, language and voice model on phones and laptops. It can watch a video and hear its soundtrack at the same time, so it can answer questions that need both. Why you'd want it: A single voice-and-vision assistant model that runs on the device in your pocket.
| ✓ Pros | ✗ Cons |
|---|---|
| Small enough for phones and laptops | Commercial use is free only below $10 million a year in revenue |
| Speaks in 19 languages and locales | No "thinking" mode |
| Runs locally through llama.cpp, including a live speech server | Short 32K-token memory for a conversation |
LiquidAI/LFM3-Omni-2B · Hugging Face
Any-to-Any model with 2B parameters, built with Transformers. Download, explore files and run it on Hugging Face.
Product Hunt
AI Launches Today
Product Hunt's top of the day leaned toward everyday AI helpers: a voice assistant for small business owners, an AI email inbox and an app builder that lives inside ChatGPT. Upvote counts below are Product Hunt's "points" when we checked late in the day. We skipped Google's Gemini 4 Carbon, covered yesterday, and a cloud browser built partly to slip past websites' bot checks.
The voice-first assistant for small business owners
🔥 Upvotes: 378 · 👤 By: Empacer (founder and CEO Yerkebulan Akhmetzaki)
💰 Pricing: free options; paid plans, 60% off the first period for Product Hunt users · 🏷 Category: Productivity, AI workflow automation
💰 Pricing: free options; paid plans, 60% off the first period for Product Hunt users · 🏷 Category: Productivity, AI workflow automation
Built for businesses of 1 to 10 people, Empacer asks three questions about your business and then drafts your first week of work. You speak a request and it drafts customer replies, social posts or invoices in your voice, and nothing is sent until you approve it. It remembers your business details across tasks, connects to more than 1,000 tools, and runs in a web browser and on Android, with iPhone listed as coming soon. Verdict: The day's top launch and a sensible hands-off helper for busy owners; the approve-before-sending rule is the right default.
View on Product Hunt →
An intelligent email inbox that works for you
🔥 Upvotes: 283 · 👤 By: Scape (founder Melvin Hagberg; Y Combinator company)
💰 Pricing: freemium; code PH26 gives 30 days of Pro free · 🏷 Category: Email clients
💰 Pricing: freemium; code PH26 gives 30 days of Pro free · 🏷 Category: Email clients
A Mac email app that sorts your inbox, drafts replies in your tone using context from your emails, meetings and apps, and reminds you when nobody has answered. It can record meetings on your computer without sending a bot into the call, in 99 languages, and prepares follow-up emails for you to review. The maker says Scape meets the SOC 2, ISO 27001 and GDPR security and privacy standards and does not train AI on your data. Verdict: A polished AI inbox for Mac users drowning in email; the free plan is enough to test whether its drafts sound like you.
View on Product Hunt →
Build and publish real apps without leaving ChatGPT
🔥 Upvotes: 172 · 👤 By: Floot (Y Combinator company)
💰 Pricing: free options; AI use runs on your ChatGPT or Claude plan · 🏷 Category: AI coding agents, vibe coding tools
💰 Pricing: free options; AI use runs on your ChatGPT or Claude plan · 🏷 Category: AI coding agents, vibe coding tools
Floot now runs as an app inside ChatGPT's sidebar: describe an app, watch it build in a live preview inside the chat, then publish it at a real web address. A database, logins and hosting are built in, Floot does not charge for AI credits because your ChatGPT plan covers that, and the page says you can publish to the iPhone and Android app stores from the same chat. Verdict: One of the easiest ways yet for non-programmers to turn an idea into a working app, if you already pay for ChatGPT.
View on Product Hunt →
You don't need an AI SRE. Give your agent Radar.
🔥 Upvotes: 115 · 👤 By: Skyhook
💰 Pricing: free and open source; paid Radar Cloud · 🏷 Category: Observability tools, open source, developer tools
💰 Pricing: free and open source; paid Radar Cloud · 🏷 Category: Observability tools, open source, developer tools
Radar is a free dashboard for Kubernetes, the popular system for running apps across many servers, that also gives AI agents the same view of your system. Its "Investigate" button runs a coding agent such as Claude Code or Codex, read-only, against a failing app and returns the likely cause with evidence, and it scrubs secret-looking values before anything reaches the AI. The makers say it found the root cause in 44 of 50 real faults in their own test, the most of six tools. Verdict: A low-risk way for engineering teams to let AI help with outages without giving it the keys to change anything.
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, new today) | Mistral Large 4 | $0.70 (preview list was $1.36) | $2.10 (preview list was $4.18) | 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 - Anthropic's, OpenAI's and Groq's official pricing pages are word-for-word the same as yesterday, and every model price on Google's page is unchanged. Gemini 4 Carbon is still not on Google's pricing page. Google did add one limit: searching the web from inside Gemini 3 and newer models (Google calls it "grounding") is now capped at 1,500 requests per day per project on its developer service, on top of the existing $14 per 1,000 searches after 5,000 free ones a month. The one addition to our table is Mistral Large 4, released as a free download today and also sold on Mistral's own API at about half the list price of the preview version we last tracked on October 23, with reused (cached) input at $0.07.
What this means: Mistral Large 4 lists at about a third of Claude Sonnet 5.5's input price and about a fifth of its output price, and like most new models it lets you choose how hard it "thinks." That choice matters for your bill: Google's page notes that thinking tokens are billed like any other output, and today's arXiv paper below finds that extra thinking often buys little on everyday coding tasks. Gemini 4 Argon's introductory price still doubles in 33 days, on December 1.
No price changes vs. yesterday - Anthropic's, OpenAI's and Groq's official pricing pages are word-for-word the same as yesterday, and every model price on Google's page is unchanged. Gemini 4 Carbon is still not on Google's pricing page. Google did add one limit: searching the web from inside Gemini 3 and newer models (Google calls it "grounding") is now capped at 1,500 requests per day per project on its developer service, on top of the existing $14 per 1,000 searches after 5,000 free ones a month. The one addition to our table is Mistral Large 4, released as a free download today and also sold on Mistral's own API at about half the list price of the preview version we last tracked on October 23, with reused (cached) input at $0.07.
What this means: Mistral Large 4 lists at about a third of Claude Sonnet 5.5's input price and about a fifth of its output price, and like most new models it lets you choose how hard it "thinks." That choice matters for your bill: Google's page notes that thinking tokens are billed like any other output, and today's arXiv paper below finds that extra thinking often buys little on everyday coding tasks. Gemini 4 Argon's introductory price still doubles in 33 days, on December 1.
arXiv Paper of the Day
Does More Thinking Pay Off? Reasoning Cost and Correctness of LLMs in Code Generation and Bug Fixing
Yonghui (Andie) Huang, Ying Pang, Qian Zhang, Yihang Cheng and Lin Ma · arXiv:2610.36952
What it claims: Many AI models now "think out loud" in a hidden reasoning trace before they write code, and you pay for those extra tokens, but coding benchmarks usually report only whether the answer was right. Researchers from Massey University and the University of Otago in New Zealand and two Chinese universities measured what reasoning actually buys on 866 Python tasks, covering writing new code and fixing real bugs, under five model settings.
Key finding: On one common code-writing test, turning reasoning on raised the Qwen3-32B model's success rate only from 68.1% to 69.5%, a gain too small to be statistically meaningful, even though it spent a median of more than 2,000 thinking tokens per task. Reasoning helped more elsewhere (71.1% to 76.5% on a second code-writing test, and 3.5% to 7.9% on bug fixing), but across models a longer trace did not reliably mean more correct answers. When the researchers kept only the first half of each stored trace and had the model answer from that, accuracy did not drop significantly.
Why practitioners should care: Treat "reasoning effort" as a setting to test on your own workload, not something to leave on high, since thinking tokens are billed as output and often cost several times more than input. If your pipeline passes one model's reasoning to another, sending only the first half may cut input costs without hurting results.
Key finding: On one common code-writing test, turning reasoning on raised the Qwen3-32B model's success rate only from 68.1% to 69.5%, a gain too small to be statistically meaningful, even though it spent a median of more than 2,000 thinking tokens per task. Reasoning helped more elsewhere (71.1% to 76.5% on a second code-writing test, and 3.5% to 7.9% on bug fixing), but across models a longer trace did not reliably mean more correct answers. When the researchers kept only the first half of each stored trace and had the model answer from that, accuracy did not drop significantly.
Why practitioners should care: Treat "reasoning effort" as a setting to test on your own workload, not something to leave on high, since thinking tokens are billed as output and often cost several times more than input. If your pipeline passes one model's reasoning to another, sending only the first half may cut input costs without hurting results.
×
Click anywhere or press ESC to close
Member discussion