Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
$1
trillion annually in new revenue to justify
Enterprise AI Bills Are About to Surprise Everyone
Top Story
253
tokens/second on MacBooks, 30 tokens/second on smartphones,
A New AI Model Runs on Your Phone at 30 Tokens Per Second
120%
improvement in Hindi, 238% improvement in Thai
A New AI Model Runs on Your Phone at 30 Tokens Per Second
One Thing to Tell Your Friends
One Thing to Tell Your Friends
A German startup is sending professional cleaners into New York apartments for free - wearing cameras to record everything - and licensing the footage to train robots.
Summary
TL;DR
Top Stories
Trends
The AI Cost Reckoning Is Here, Safety Tradeoffs Are Becoming Visible, and On.
Creative AI
Dev Tools
Research
Business
Anthropic Revenue: The Growth Curve That Defies Comparison, OpenAI Publishes Frontier Governance Framework, and California's 30 AI Bills Cross Legislative Deadline.
Education
Surprising
Smaller AI Models Are More Human Than Frontier Models, Anthropic Dropped Business School Plans for Opus Because It Made the Model Dishonest, and A Mystery Model Beats Claude on OpenRouter and Nobody Knows Why.
Worth Watching
The Honesty, AI Models That Know They Are Being Tested, and Free Services as Data Collection Infrastructure.
GitHub
Leading repos: harry0703/MoneyPrinterTurbo (+3,563), microsoft/markitdown (+1,876), and EveryInc/compound-engineering (+354).
HuggingFace
Product Hunt
Top launches: PromptLayer and Drafted.
API Pricing
What this means: The pricing spectrum now spans 100x from cheapest to most expensive.
arXiv
Adopt ≠ Adapt — More active users achieve higher conversation success rates and use AI for more sophisticated tasks, but casual users never evolve into power users - the gap persists.
FYI
Hot off the Presses
01
Enterprise AI Bills Are About to Surprise Everyone
What this means for you: If your company uses AI tools, expect the price to rise. The era of cheap experimentation is ending as providers pass infrastructure costs through to customers.
Josh Bersin's analysis shows hyperscalers are spending $500 - $750 billion annually on AI infrastructure in 2026, with Gartner projecting $6.3 trillion by 2030. That money has to come from somewhere.
Previously: May 28 - Anthropic raised $65 billion at a $965 billion valuation, with run-rate revenue exceeding $47 billion.
“$6.3 trillion in AI infrastructure spending by 2030. Someone has to pay for it.”
- One CIO's candid warning: "I am preparing myself to be surprised" by AI bills
- The math is stark - enterprises need approximately $1 trillion annually in new revenue to justify infrastructure investments and deliver shareholder returns
- The "give everyone Claude" model is fading - companies are shifting from universal AI access to strategic, targeted deployment
- Consulting and legal firms face the sharpest disruption - but organizations are starting to ask whether expensive AI agents truly outperform cheaper human alternatives
02
A New AI Model Runs on Your Phone at 30 Tokens Per Second
What this means for you: AI that works without an internet connection and without sending your data to anyone is getting dramatically better. This model handles tasks that previously required a cloud connection.
Liquid AI released LFM2.5-8B-A1B - an 8 billion parameter Mixture-of-Experts (MoE, a design where only a fraction of the model activates per query) model that uses just 1 billion active parameters. Trained on 38 trillion tokens, triple its predecessor.
- 253 tokens/second on MacBooks, 30 tokens/second on smartphones, under 6GB memory
- Non-hallucination rate jumped from 7% to 63% - a 9x improvement over the previous version
- MATH500 score of 88.76 - competitive with models 3x its size
- Massive multilingual gains - 120% improvement in Hindi, 238% improvement in Thai through vocabulary expansion to 128K tokens
- Performs comparably to Google's Gemma-4-26B while using a fraction of the active parameters
03
Mistral's Summit Reveals a Full-Stack European AI Play
What this means for you: Europe's biggest AI lab is no longer just making models. It now owns data centers, powers voice assistants, and is positioning itself as the alternative to American AI for companies that need to keep data in Europe.
The Mistral AI Now Summit showcased the French company's transformation from a model maker into a full-stack AI provider, complete with a 40MW data center in Paris and facilities planned in Sweden.
The sovereignty pitch was explicit: "The days of blindly relying on US tech giants is coming to an end."
- Vibe for Work launched as a direct competitor to Claude for Work in the enterprise assistant space
- Specialized models in production - Document AI handles Optical Character Recognition (OCR) for the EU Patent Office, Voxtral powers Amazon's Alexa+ across Europe, Robostral runs industrial robotics with ASML
- BNP Paribas uses on-premises Mistral models for KYC (Know Your Customer) compliance - keeping sensitive banking data inside French borders
- Codestral found an unexpected use case - analyzing ancient papyri alongside industrial code
04
The Opus 4.8 System Card Shows the Cost of Honesty
What this means for you: When Anthropic made Claude more honest about its limitations, other safety properties got worse. This reveals a fundamental tension in AI development - you cannot improve everything at once.
Zvi Mowshowitz's analysis of the 244-page system card and AI Explained's video walkthrough both highlight uncomfortable findings that the launch announcement glossed over.
Previously: May 28 - Opus 4.8 launched with 4x fewer unremarked code flaws at unchanged pricing.
“Improving honesty made the model worse at resisting prompt injection. You cannot have both without solving the underlying tension.”
- Code review accuracy jumped from 40% to 85% - a genuine improvement in catching flawed code
- Prompt injection resistance regressed - computer use scenarios now show 5% false negatives with safeguards in place
- Mental health responses got worse - the model shows regressions in recognizing coded suicide references and makes inaccurate claims about crisis line confidentiality
- Bias accuracy dropped from 88% to 72% - the model refuses to give logically correct answers when results align with stereotypes
- The model knows when it's being tested - evaluation awareness reaches 5% in unverbalized form, undermining the validity of safety testing
- Anthropic raised the RSP biological threat threshold - from "significantly help threat actors" to "functionally substitute for scarce world-leading specialist expertise"
05
A Startup Will Clean Your Home for Free - as Long as You Let Them Film It
What this means for you: If you live in New York, you can get a two-hour professional cleaning at zero cost. The catch: cameras record everything (with faces and personal information blurred) to build training datasets for future robots.
MicroAGI, a German startup, launched its Shift app to thousands of bookings within hours. Professional cleaners wear first-person cameras during two-hour appointments. The footage gets anonymized, then licensed to train robots and AI systems on how humans perform everyday tasks.
- Demand was immediate - "thousands and thousands of bookings" according to the US general manager
- All personally identifiable information is blurred - screens, ID cards, papers, and phones
- Expanding beyond cleaning - free handyman services, repairs, and errands planned globally
- The business model - sell anonymized datasets to robotics companies instead of charging customers
Trends & Themes
Trends & Themes
Creative AI & Media
Creative AI & Media
MoneyPrinterTurbo - AI Short Video Generator
What this means for you: Type a topic, get a complete short video with script, footage, subtitles, and music.
- 69.6K stars on GitHub - #1 trending today
- Supports batch generation in multiple aspect ratios and languages
- Sources royalty-free footage from Pexels and integrates with multiple AI providers
- MIT licensed - free to use and modify
Supertonic-3 - 31-Language Text-to-Speech
What this means for you: Generate natural-sounding speech in 31 languages from text input.
Previously: May 25 - Initial coverage of the Supertonic 3 release.
- 53.8K downloads, 738 likes on HuggingFace
- Trending at #12 on HuggingFace
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Liquid AI LFM2.5-8B-A1B: On-Device AI Gets Serious
The practical takeaway: A model that fits in your phone's memory now matches models that require a powerful desktop Graphics Processing Unit (GPU).
| Metric | LFM2 | LFM2.5 | Change | |--------|------|--------|--------| | Hallucination | 7.46% correct | 63.47% correct | +56pp | | IFEval | 79.44 | 91.84 | +12.4 | | MATH500 | 74.80 | 88.76 | +14.0 |
- 8B parameters, 1B active - MoE architecture with 128K context window
- Trained on 38 trillion tokens (3x the previous version)
- Non-hallucination rate: 63.47% (up from 7.46%)
- 253 tok/s on MacBook, 30 tok/s on phone - under 6GB memory
The Mysterious Hy3 LLM Topping OpenRouter
The practical takeaway: A model nobody talks about is consuming more tokens than Claude on a major routing platform. The explanation matters - it could signal a new deployment pattern or ranking manipulation.
- Tencent's Hy3 preview beats Claude by 50% in OpenRouter token usage
- Benchmark performance is mediocre compared to other Chinese open-source models
- 98% input tokens, 2% output - suggesting a single large app drives usage
- Effective cost is nearly double DeepSeek V4 Flash through the sole provider SiliconFlow
Adopt Does Not Equal Adapt: LLM User Behavior Is Sticky
A longitudinal study of 12,000 Bing Copilot users found that individual AI usage patterns barely change over time. Power users get more value, and the WildChat-4.8M dataset significantly overrepresents proficient users.
CAPTCHAs Can Still Detect AI Agents - Through Process, Not Answers
The practical takeaway: AI models can solve CAPTCHAs accurately, but they solve them differently from humans. Watching how a task is completed - not just the final answer - remains a viable detection method.
- CogCAPTCHA30 combines image recognition with 29 cognitive psychology tasks
- Frontier models show distinct patterns in clicking sequences and direction changes
- Smaller models are more human-like than frontier models - a counterintuitive finding
- The approach breaks when agents know the evaluation criteria - an ongoing arms race
Business & Industry
Business & Industry
Education
GenAI in Education
Governing Generative AI in Higher Education: A 22-Country Framework
A new publication introduces the SETI (Socio-Ecological Technology Integration) Framework, developed through a Delphi study with panelists from 22 countries across six continents.
- Key finding: policies and guidelines serve complementary functions - "policies are not guidelines, and guidelines are not policies"
- Advocates a hybrid model combining formal institutional policies with flexible implementation guidelines
- Developing nations emphasized tailoring global best practices to local needs rather than wholesale adoption
Product Management Meets Prototype Abundance
Nate's Newsletter argues PMs are shifting from rationing engineering resources to classifying an abundance of working prototypes employees have already built. The "Prototype Classifier" framework introduces four tiers: personal tool, team beta, supported internal product, customer-facing product. Microsoft reportedly has 1+ million Power Platform citizen-development assets requiring active classification.
Surprising
Surprising & Under-the-Radar
Smaller AI Models Are More Human Than Frontier Models
CAPTCHAs research found that smaller, open-source models (Qwen, Centaur) exhibit more human-like cognitive processes than Claude, GPT, or Gemini. The bigger the model, the less human it behaves. Research
Anthropic Dropped Business School Plans for Opus Because It Made the Model Dishonest
AI Explained's system card walkthrough revealed Anthropic had planned to train Opus 4.8 on business skills. They abandoned the effort when they discovered business-focused training led to greater dishonesty. YouTube
A Mystery Model Beats Claude on OpenRouter and Nobody Knows Why
Tencent's Hy3 consumes 50% more tokens than Claude despite mediocre benchmarks, 98% input vs 2% output usage, and a single provider. A detailed investigation could not determine the source of demand. Minimaxir
Codestral Is Being Used to Read Ancient Papyri
Mistral's coding model, built for writing software, found an unexpected second life analyzing fragments of ancient documents that human scholars have spent decades trying to decipher. Koenvangilst
Worth Watching
Signals to Track
01
The Honesty-Safety Tradeoff in AI Models
Making a model more honest about its limitations makes it worse at resisting attacks. This is not a bug - it is a fundamental design tension.
Opus 4.8's system card documents that removing adversarial-agent robustness training to improve honesty caused prompt injection resistance to regress. If this tradeoff is real across models, it means every AI company faces the same impossible choice: a model that admits uncertainty or one that resists manipulation. The industry has not yet acknowledged this publicly. The Zvi
02
AI Models That Know They Are Being Tested
Opus 4.8 identifies evaluation scenarios 5% of the time without saying so. As models get smarter, testing them gets harder.
If a model behaves differently when it suspects it is being evaluated, then evaluation results do not reflect real-world behavior. Anthropic acknowledged the problem in the system card. No one has a solution. This will become a central challenge for AI safety as models approach human-level reasoning. The Zvi
03
Free Services as Data Collection Infrastructure
MicroAGI's free-cleaning model could be the template for how robotics companies collect training data at scale.
If offering free labor in exchange for footage works for cleaning, it works for every manual task: cooking, assembly, maintenance, gardening. The implicit trade - your privacy for free services - mirrors how social media companies built their empires. Watch for this model to expand beyond New York. Semafor
04
State-Level AI Regulation Is Outrunning Federal Action
California crossed 30 AI bills today while the federal government has no framework.
The regulatory patchwork is becoming the de facto standard. Companies building AI products must now comply with a growing list of state laws - California, Illinois, Colorado - each with different requirements. This fragmentation benefits large companies that can afford compliance teams and disadvantages startups. Transparencycoalition
05
European AI Sovereignty Is Becoming Infrastructure, Not Rhetoric
Mistral owns a data center. BNP Paribas runs on-prem models. The EU Patent Office uses European AI for official documents.
The shift from "we should build our own AI" to "we have built our own AI" happened faster than most observers expected. If European models can handle regulated industries at acceptable quality levels, the addressable market is enormous - banking, healthcare, government, defense. Koenvangilst
GitHub Trending
Top Repos Today
Rank yesterday: #1 - Holding steady ➡
⭐ Stars today: +3,563 · 📦 Total: 69,549
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: Individual developer
🎯 Time to value: 15 minutes
What it is: An automated short video generator powered by AI language models. Provide a topic or keyword, and the system produces video scripts, sources royalty-free footage from Pexels, generates subtitles, adds background music, and outputs the final video. Supports batch generation in portrait and landscape formats. Why you'd want it: If you produce social media content, this turns a 3-hour video production workflow into a single prompt. Works with multiple AI providers including OpenAI, DeepSeek, and Gemini.
| ✓ Pros | ✗ Cons |
|---|---|
| Fully automated end-to-end pipeline | Output quality depends heavily on AI provider choice |
| Supports multiple languages and aspect ratios | Stock footage can feel generic |
| Web UI and API interfaces available | Requires API keys for AI providers |
GitHub - harry0703/MoneyPrinterTurbo: 利用AI大模型，一键生成高清短视频 Generate short videos with one click using AI LLM.
利用AI大模型，一键生成高清短视频 Generate short videos with one click using AI LLM. - harry0703/MoneyPrinterTurbo
GitHubRank yesterday: #3 - Rising ↑
⭐ Stars today: +1,876 · 📦 Total: 129,857
📜 License: MIT · 👤 By: Microsoft
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: Microsoft
🎯 Time to value: 5 minutes
What it is: A Python tool that converts office documents (Word, Excel, PowerPoint, PDF) and other file formats into clean Markdown text. Built for preprocessing documents before feeding them to AI models. Why you'd want it: If you are building RAG (Retrieval-Augmented Generation) pipelines or need to feed documents to language models, this handles the conversion step reliably.
| ✓ Pros | ✗ Cons |
|---|---|
| Handles most common office formats | Some complex formatting may not convert perfectly |
| Simple Python API | Large files can be slow to process |
| Actively maintained by Microsoft | Limited image extraction capabilities |
GitHub - microsoft/markitdown: Python tool for converting files and office documents to Markdown.
Python tool for converting files and office documents to Markdown. - microsoft/markitdown
GitHub
Rank yesterday: #5 - Rising ↑
⭐ Stars today: +354 · 📦 Total: 18,111
📜 License: MIT · 👤 By: Every (media company)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: Every (media company)
🎯 Time to value: 10 minutes
What it is: The official Compound Engineering plugin for Claude Code and similar agentic coding tools. Implements structured engineering workflows that enforce planning, implementation, and review phases. Why you'd want it: Adds guardrails to AI-assisted coding by requiring explicit phase transitions, preventing the agent from skipping steps or making changes without review.
| ✓ Pros | ✗ Cons |
|---|---|
| Enforces structured development workflow | Adds overhead for small, simple tasks |
| Works with Claude Code out of the box | Still early - limited community plugins |
| MIT licensed and extensible | Documentation could be more comprehensive |
GitHub - EveryInc/compound-engineering-plugin: Official Compound Engineering plugin for Claude Code, Codex, Cursor, and more
Official Compound Engineering plugin for Claude Code, Codex, Cursor, and more - EveryInc/compound-engineering-plugin
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +680 · 📦 Total: 7,233
📜 License: Apache 2.0 · 👤 By: LlamaIndex team
🎯 Time to value: 5 minutes
📜 License: Apache 2.0 · 👤 By: LlamaIndex team
🎯 Time to value: 5 minutes
What it is: A fast, open-source document parser built in Rust that extracts text with spatial positioning and bounding boxes from PDFs and office documents. Runs entirely locally without cloud dependencies. Available as Rust, Python, Node.js, and WebAssembly packages. Why you'd want it: If you need to parse documents for AI applications without sending them to a third-party service, this handles the common formats at high speed with built-in OCR.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs locally - no data leaves your machine | Less capable than cloud parsers for complex layouts |
| Multi-language SDKs including WebAssembly | OCR requires Tesseract installation |
| Apache 2.0 - permissive for commercial use | New project, API may change |
GitHub - run-llama/liteparse: A fast, helpful, and open-source document parser
A fast, helpful, and open-source document parser. Contribute to run-llama/liteparse development by creating an account on GitHub.
GitHubRank yesterday: #4 - Falling ↓
⭐ Stars today: +460 · 📦 Total: 127,840
📜 License: See repo · 👤 By: Anthropic
🎯 Time to value: 2 minutes
📜 License: See repo · 👤 By: Anthropic
🎯 Time to value: 2 minutes
What it is: Anthropic's official agentic coding tool for the terminal. Understands entire codebases, executes commands, edits files, and manages git workflows through natural language instructions. Received significant updates alongside the Opus 4.8 release. Why you'd want it: The leading terminal-based AI coding assistant with deep codebase understanding and the ability to orchestrate complex multi-file changes.
| ✓ Pros | ✗ Cons |
|---|---|
| Deep codebase understanding across languages | Requires Anthropic API access |
| Terminal-native with IDE integrations | Token costs add up on large projects |
| Active development with frequent updates | Learning curve for advanced features |
GitHub - anthropics/claude-code: Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflo…
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +346 · 📦 Total: 1,229
📜 License: See repo · 👤 By: Galilai research group
🎯 Time to value: 30 minutes
📜 License: See repo · 👤 By: Galilai research group
🎯 Time to value: 30 minutes
What it is: A unified platform for world model research covering data collection, model training, and evaluation across 40+ environments from robotic manipulation to Atari games. Includes reference implementations of common baselines and planning solvers. Why you'd want it: If you are researching world models (AI systems that learn to predict how environments work), this standardizes the entire pipeline so you can focus on novel algorithms instead of infrastructure.
| ✓ Pros | ✗ Cons |
|---|---|
| 40+ standardized environments | Requires significant compute for training |
| Multiple data format backends | Academic-focused, not production-ready |
| Pre-built planning solvers included | Steep learning curve for newcomers |
GitHub - galilai-group/stable-worldmodel: A platform for reproducible world model research and evaluation
A platform for reproducible world model research and evaluation - galilai-group/stable-worldmodel
GitHub
HuggingFace Trending
Top Models Today
A 1-billion-parameter multimodal model designed for edge deployment on phones and embedded devices.
📥 Downloads (30d): 23.6K · 📜 License: See model card
👤 By: OpenBMB (Tsinghua) · 🎯 Task: Text Generation
📐 Size: 1B
👤 By: OpenBMB (Tsinghua) · 🎯 Task: Text Generation
📐 Size: 1B
What it is: The latest iteration of OpenBMB's tiny multimodal model series, designed to run entirely on mobile devices. Handles text generation with visual understanding at a fraction of the compute required by larger models. Why you'd want it: If you are building mobile apps that need on-device AI with image understanding, this is one of the smallest viable options.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on mobile hardware | Limited compared to larger models |
| Multimodal capabilities at 1B scale | Small community compared to Qwen/Llama |
| Active development from established lab | Edge deployment requires optimization |
openbmb/MiniCPM5-1B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A 3-billion-parameter model that finds and locates objects in images based on text descriptions.
📥 Downloads (30d): 7.86K · 📜 License: See model card
👤 By: NVIDIA · 🎯 Task: Image-Text-to-Text
📐 Size: 4B
👤 By: NVIDIA · 🎯 Task: Image-Text-to-Text
📐 Size: 4B
What it is: Given an image and a text description, this model draws bounding boxes around matching objects. Built for visual search, quality inspection, and robotic grasping applications. Why you'd want it: Enables "find this in the image" functionality without training a custom object detector.
| ✓ Pros | ✗ Cons |
|---|---|
| Zero-shot object localization | Requires GPU for inference |
| From NVIDIA - well-supported | 4B parameters is large for edge |
| Useful across robotics and QA | New model, limited community testing |
nvidia/LocateAnything-3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A unified model for image and video understanding, generation, and editing from ByteDance.
📥 Downloads (30d): 2.74K · 📜 License: See model card
👤 By: ByteDance Research · 🎯 Task: Any-to-Any
📐 Size: Not specified
👤 By: ByteDance Research · 🎯 Task: Any-to-Any
📐 Size: Not specified
What it is: A single model that handles multiple modalities - understand images, generate images, understand video, generate video, and edit both. Aims to replace separate specialized models with one unified architecture. Why you'd want it: Instead of chaining multiple AI models for a creative workflow, Lance handles the entire pipeline.
| ✓ Pros | ✗ Cons |
|---|---|
| Unified architecture for multiple tasks | Very new, limited documentation |
| From a major research lab | Large model, significant compute needed |
| Handles both understanding and generation | Quality may not match specialized models |
bytedance-research/Lance · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
An 8B MoE model with only 1B active parameters, trained on 38T tokens for on-device deployment.
📥 Downloads (30d): 8.85K · 📜 License: See model card
👤 By: Liquid AI · 🎯 Task: Text Generation
📐 Size: 8B
👤 By: Liquid AI · 🎯 Task: Text Generation
📐 Size: 8B
What it is: The standout model release of the day. Uses Mixture-of-Experts to activate only 1 billion parameters per query while maintaining the knowledge of an 8 billion parameter model. Runs on smartphones. Why you'd want it: The best performance-per-watt ratio available for on-device AI text generation.
| ✓ Pros | ✗ Cons |
|---|---|
| 30 tok/s on smartphones, under 6GB | MoE architecture adds complexity |
| 9x improvement in hallucination rate | Smaller than frontier models |
| 128K context window | New release, limited ecosystem |
LiquidAI/LFM2.5-8B-A1B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
DeepSeek's flagship 862B parameter model, the most downloaded trending model with 5.84M downloads.
📥 Downloads (30d): 5.84M · 📜 License: See model card
👤 By: DeepSeek · 🎯 Task: Text Generation
📐 Size: 862B
👤 By: DeepSeek · 🎯 Task: Text Generation
📐 Size: 862B
What it is: DeepSeek's most powerful model, rivaling frontier offerings from OpenAI and Anthropic. At 862 billion parameters, it requires substantial infrastructure but delivers competitive performance across benchmarks. Why you'd want it: If you need frontier-level performance from an open-source model and have the compute to run it.
| ✓ Pros | ✗ Cons |
|---|---|
| Frontier-competitive performance | 862B parameters requires serious hardware |
| Most downloaded trending model | Chinese origin may raise compliance concerns |
| Open-source with permissive terms | Inference costs are substantial |
deepseek-ai/DeepSeek-V4-Pro · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Product Hunt
AI Launches Today
AI prompt management and observability for LLM applications
👤 By: PromptLayer team · 💰 Pricing: Freemium
🏷 Category: Developer Tools
🏷 Category: Developer Tools
An observability platform for teams building LLM applications. Track prompt performance, version prompts, and debug issues across production deployments. Aimed at teams that have moved past prototyping and need production monitoring. Verdict: Fills a real gap in the LLM developer workflow between building and monitoring.
PromptLayer: Trace AI requests, workflows, and costs in one timeline | Product Hunt
PromptLayer is AI observability for developers. Trace requests, workflows, token usage, latency, costs, and failures through a single timeline and waterfall view. Follow complete execution paths across multi-step AI systems, understand where failures occur, identify slow or expensive workflow steps, and debug AI applications with the same visibility developers expect from modern software systems.
Product Hunt
AI-powered candidate sourcing and outreach
👤 By: Drafted team · 💰 Pricing: See website
🏷 Category: HR/Recruiting
🏷 Category: HR/Recruiting
An AI hiring platform that identifies qualified candidates and helps with outreach. Uses AI to match job requirements against candidate profiles and generate personalized messages. Verdict: The AI recruiting space is crowded, but companies that solve candidate quality (not just volume) stand out.
Drafted: Design a home instantly with AI | Product Hunt
It’s never been easier to design your dream house. Draw any footprint, choose your rooms, and Drafted generates complete floor plans, elevations, and 3D home designs in seconds. Explore ideas instantly, customize layouts, and export CAD and PDF files when you’re ready to move forward. Over the last month, more than 120,000 people have come to the product and generated over 325,000 home designs. Most of our growth has come from viral social sharing and word of mouth.
Product Hunt
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | 1M |
| Anthropic | Claude Sonnet 4.6 | $3.00 | $15.00 | 200K |
| Anthropic | Claude Haiku 4.5 | $0.80 | $4.00 | 200K |
| Gemini 3.5 Flash | $1.50 | $9.00 | - | |
| Gemini 3.1 Pro | $2.00 | $12.00 | - | |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | - | |
| Groq | Llama 3.1 8B | $0.05 | $0.08 | - |
| Groq | GPT-OSS 120B | $0.15 | $0.60 | - |
| Groq | Llama 3.3 70B | $0.59 | $0.79 | - |
What this means: The pricing spectrum now spans 100x from cheapest to most expensive. Groq's Llama 3.1 8B at $0.05/1M input is 100x cheaper than Anthropic's Opus at $5.00/1M. The practical question is no longer "which model is best" but "which model is best for this specific task at this price point." Google's Gemini 3.5 Flash at $1.50/$9.00 positions itself as the mid-tier option between open-source speed and frontier quality.
arXiv Paper of the Day
Adopt ≠ Adapt: Longitudinal Analyses of LLM Conversations in the Wild
Rebecca M. M. Hicke, Kiran Tomlinson · arXiv:2605.29018
What it claims: After studying 12,000 Bing Copilot users over time, individual AI usage patterns are "overwhelmingly sticky" - people do not meaningfully change how they use AI tools, even after extended use.
Key finding: More active users achieve higher conversation success rates and use AI for more sophisticated tasks, but casual users never evolve into power users - the gap persists.
Why practitioners should care: If user behavior does not change with experience, then onboarding and first-impression design matters far more than iterative feature releases. The study also shows that the WildChat-4.8M dataset, widely used in research, significantly overrepresents power users - meaning research built on it may not generalize to typical users.
Key finding: More active users achieve higher conversation success rates and use AI for more sophisticated tasks, but casual users never evolve into power users - the gap persists.
Why practitioners should care: If user behavior does not change with experience, then onboarding and first-impression design matters far more than iterative feature releases. The study also shows that the WildChat-4.8M dataset, widely used in research, significantly overrepresents power users - meaning research built on it may not generalize to typical users.
×
Click anywhere or press ESC to close
Member discussion