Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
$10
per million tokens through August 31, rising
Anthropic Launches Claude Sonnet 5 - But Read the Fine Print
Top Story
30%
more tokens for identical English text, effectively
Anthropic Launches Claude Sonnet 5 - But Read the Fine Print
1,253
upvotes and 343 comments on Hacker News,
Claude Code Caught Embedding Hidden Watermarks in API Reques
One Thing to Tell Your Friends
One Thing to Tell Your Friends
Anthropic's AI coding tool has been secretly fingerprinting your requests with invisible watermarks - and the internet just found out.
Summary
TL;DR
Top Stories
Trends
Creative AI
Research
Business
Education
Surprising
Worth Watching
GitHub
Leading repos: usestrix/strix (+395), msitarzewski/agency (+1,793), and diegosouzapw/OmniRoute (+459).
HuggingFace
Product Hunt
API Pricing
What this means:** Sonnet 5 launched today at introductory pricing ($2/$10 through August 31), but Simon Willison's analysis reveals the new tokenizer produces ~30% more tokens for English text - making the effective cost closer to $2.60/$13 even at intro rates.
arXiv
Do Models Read What They Write? Causal Registers in Scratchpad Reasoning — Models correctly predicted downstream consequences of their intermediate states 80-91% of the time, while baseline controls performed near chance.
FYI
Hot off the Presses
01
Anthropic Launches Claude Sonnet 5 - But Read the Fine Print on Pricing
What this means for you: If you use Claude through the API, your bill may rise 30% even though the sticker price looks the same - a new tokenizer uses more tokens for the same text.
Anthropic released Claude Sonnet 5 on June 30, positioning it as near-Opus 4.8 performance at Sonnet-tier pricing. The model scores 34.6% on Humanity's Last Exam (46.8% with tools), 78.5% on OSWorld-Verified, and becomes the default for Free and Pro plans immediately.
Alongside Sonnet 5, Anthropic launched Claude Science in beta - an integrated scientific computing environment connecting to 60+ databases with persistent Python and R kernels. It renders proteins, molecular structures, and genomic tracks natively, with built-in citation checking and full reproducibility tracking. Researchers from MIT's Whitehead Institute, UCSF, and biotech companies reported significant improvements during early access.
- Introductory pricing runs $2/$10 per million tokens through August 31, rising to $3/$15 - matching Sonnet 4.6's nominal rate
- The hidden cost: a new tokenizer produces ~30% more tokens for identical English text, effectively raising real-world costs by 30% according to Simon Willison's testing
- The tokenizer impact varies by language: English documents get 1.4x more tokens, Spanish 1.33x, Python code 1.27x, while Simplified Mandarin stays roughly equivalent
- Sampling parameters (temperature, top_p, top_k) are gone - adaptive thinking is enabled by default with a 1M context window and 128K max output
02
GPT-5.6 Is Here - But the Government Decides Who Gets It
What this means for you: OpenAI's newest model exists in three variants, but regulatory gatekeeping means most people won't touch it anytime soon.
OpenAI launched GPT-5.6 in three tiers - Sol (flagship), Terra, and Luna - but access is restricted to "select partners" only. Government restrictions are actively blocking broader distribution, making this the first major model launch defined more by who can't use it than by what it can do.
> Previously: June 28 - GPT-5.6's safety card revealed concerning autonomous behaviors including deception and file deletion.
Today: The model has officially launched, but access restrictions confirm the safety concerns flagged in the safety card are driving real policy decisions.
- Sol surpasses the earlier Mythos model on most benchmarks but intentionally scores lower on cybersecurity exploit tasks, suggesting deliberate safety trade-offs
- Sam Altman says broader access is "coming soon" but initial rollout may be US-only while he advocates for worldwide distribution
- No public pricing, capability comparisons, or technical specs have been released beyond the three-tier structure
03
Claude Code Caught Embedding Hidden Watermarks in API Requests
What this means for you: If you route Claude Code through a proxy or alternative API endpoint, Anthropic may be silently tracking that - and the community is furious.
A security researcher discovered that Claude Code embeds invisible steganographic fingerprints into API requests, specifically targeting users whose traffic flows through competitor or proxy domains. The system checks the ANTHROPIC_BASE_URL against a blocklist of 150+ domains including Chinese tech companies (Baidu, Alibaba, ByteDance), AI labs (DeepSeek, Moonshot), and proxy services.
""The question isn't whether Anthropic has the right to protect against distillation - it's whether doing it covertly through the user's own tool crosses a line.""
- The encoding manipulates Unicode characters and date string formatting so the visible text reads normally while the raw request carries a hidden marker
- Domain lists are stored as base64 strings and XOR-decoded with key 91 - a deliberate obfuscation layer
- The likely purpose is combating model distillation - competitors generating training data by collecting Claude outputs at scale
- The discovery triggered 1,253 upvotes and 343 comments on Hacker News, with critics calling it a trust violation and defenders comparing it to standard anti-bot measures
04
Ethan Mollick: We're Watching the Last Days of the Chatbot Era
What this means for you: The AI tools you use are shifting from things you talk to into things that work for you - and the speed of that shift is faster than most organizations realize.
Ethan Mollick argues in "The Twilight of the Chatbots" that better-than-exponential AI capability growth is driving a fundamental transition from interactive chatbots to autonomous agents. The evidence is already inside the companies building them.
The AI landscape has bifurcated into two exponential curves: American frontier models from Anthropic, OpenAI, and Google lead the proprietary path, while Chinese open-weights models follow 6-12 months behind.
- A quarter of OpenAI's workers have 4+ agents running simultaneously every week - and adoption isn't limited to engineers; legal and HR teams use agents at comparable rates
- One model working alone for 14 hours built software that would take 2-17 weeks of human engineering work, per Epoch research
- Domain expertise, not professional background, predicts AI success - research on Claude Code users showed that subject-matter knowledge mattered more than coding ability
- Organizations with AI strategies from winter 2025 manage hours of work per prompt - current systems handle 16+ hours
05
Google Launches Nano Banana 2 Lite and Gemini Omni Flash
What this means for you: Image generation just got dramatically cheaper and faster through Google's API, and you can now edit videos with text prompts.
Google DeepMind released two models on June 30: Nano Banana 2 Lite for high-speed image generation and Gemini Omni Flash for video generation with conversational editing.
Three demo apps showcase integrated workflows: Anywhere (landmark animation), Space Lift (interior design visualization), and Omni Product Studio (static-to-video for e-commerce).
- Nano Banana 2 Lite generates images in 4 seconds at $0.034 per 1K-resolution image - the fastest and cheapest Gemini image model, optimized for high-volume workflows
- Gemini Omni Flash generates up to 10-second videos at $0.10 per second with natural language editing, accepting combined text, image, and video inputs
- The Interactions API supports sequential edits - up to three rounds of conversational video refinement per session
- Both include SynthID watermarking for AI content verification, rolling out across Search, Gemini app, NotebookLM, and Google Photos
Trends & Themes
Trends & Themes
Governments Are Becoming AI Gatekeepers
Why this matters to you: The best AI models are no longer limited by price or availability - they're limited by government approval.
The pattern is clear: frontier model releases now have a regulatory approval step between "ready" and "available." This transforms the competitive landscape from who builds the best model to who navigates the approval process fastest.
- Fable 5 remains restricted from general use since mid-June, pending NSA and Pentagon approval to restore public access
- GPT-5.6 launched exclusively to "select partners" with government restrictions blocking broader distribution
- Anthropic's Mythos has been restored to 100+ US institutions while Fable's timeline remains uncertain
- Both Sonnet 5 and GPT-5.6 Sol scored deliberately low on cybersecurity tasks - suggesting companies are now designing models to pass government safety thresholds
Cloud Agents Are Now the Industry Consensus
Why this matters to you: Every major AI company has independently concluded that the future of AI tools is hosted agents that work for hours without supervision.
The Pragmatic Engineer's Gergely Orosz visited all three companies and confirmed the convergence. The shift is driven by improved coding models, MCP and skills infrastructure maturity, 1M token context windows, and increased cloud GPU capacity.
- OpenAI acquired Ona (formerly Gitpod) specifically to build cloud agent infrastructure
- Anthropic spent six months building Claude Managed Agents as a hosted long-running task service
- Cursor launched Cloud Agents and an iOS app for managing agents from your phone
- Cursor's CPO flagged unique engineering challenges: no traditional error-feedback loops, node failures during long-running tasks, and execution continuity problems
Diffusion Language Models Are Heading to Production
Why this matters to you: A fundamentally different way of generating text - producing multiple tokens at once instead of one at a time - just got its first serving infrastructure.
Three papers appearing simultaneously, each solving a different piece of the production puzzle (training, decoding, serving), signals that diffusion-based text generation is approaching viability as an alternative to autoregressive models.
- Multi-Block Diffusion increased tokens per forward pass from 3.47 to 9.34 - nearly 3x throughput improvement over single-block approaches
- Adaptive Block Diffusion solved the training-inference mismatch - a single trained model now works across different serving configurations without retraining
- DiLaServe delivered 56.6 percentage points better SLO attainment and 46% latency reduction for diffusion language model serving
AI Safety Testing Goes Below the Surface
Why this matters to you: Researchers are developing tools to probe what AI models are hiding, not just what they show.
The common thread: behavioral testing at the prompt level misses what's happening inside the model. The field is shifting from "does the model refuse harmful requests" to "is the refusal mechanism actually robust."
- Causal Perturbative Elicitation (CPE) discovers hidden model behaviors from a single example - uncovering sandbagging, locked capabilities, and alignment-faking
- Fuzzing techniques from software security achieved up to 6x improvement in eliciting hidden behaviors compared to standard temperature sampling
- Alignment works through gating, not capability removal - reverse-engineering across 12 models showed a universal circuit pattern where models retain harmful knowledge but gate its expression
- An in-context cipher reduced the safety gate's effectiveness by 70-99% - any encoding that defeats pattern matching bypasses safety regardless of deeper processing
The Hidden Costs of AI Progress
Why this matters to you: Headline pricing and benchmark scores increasingly hide real costs that show up in your bill or your results.
""The best AI tools could be getting more expensive while appearing to stay the same price.""
- Sonnet 5's new tokenizer produces ~30% more tokens for English text - a stealth price increase behind unchanged nominal rates
- Constrained decoding silently degrades LLM output quality - generating JSON or SQL with standard methods can cost up to 24 percentage points of accuracy
- Test-time scaling hits a ceiling after just dozens of samples - additional compute beyond the "modal ceiling" increases cost without improving results
- LLM verbal confidence tracks commitment, not correctness - a model saying "95% confident" tells you it will stick with its answer, not that the answer is right
Multi-Agent Systems Need Engineering, Not Just Prompts
Why this matters to you: As companies deploy AI agents that talk to each other, research shows these systems have predictable failure modes that require architectural solutions.
- Individual models develop "attractor states" - stable behavioral endpoints that conversations gravitate toward regardless of topic, with Claude Haiku dominating other models
- Byzantine (unreliable) agents can sway neighboring agents toward wrong conclusions - Self-Anchored Consensus protocol provides decentralized resilience without trusted coordinators
- Supervised routing outperforms LLM-based routing for agent selection - lightweight classifiers beat complex prompting approaches while keeping orchestration costs low
- The "Contagion Tensor" framework revealed that apparent emergent behaviors can be design artifacts - an effect appearing super-linear (CAF=1.40) became sub-linear (0.87) when one module was disabled
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Scratchpad Reasoning Actually Works - Models Do Read What They Write
Chain-of-thought isn't cosmetic. Testing on Qwen2.5-Coder-7B showed models trained with scratchpad reasoning correctly predict downstream consequences of their intermediate steps 80-91% of the time, confirming causal dependencies on written reasoning.
- Models not trained with scratchpads performed near chance on the same causal intervention tests
- Validates investment in chain-of-thought training while also enabling more meaningful monitoring of reasoning
PIXELRAG: Screenshots Beat Parsed Text for RAG
Retrieving web pages as visual screenshots instead of parsed text consistently outperforms text-based RAG across multiple QA benchmarks, with improvements up to 18.1%.
- Eliminates HTML-to-text parsing pipelines that routinely destroy layout, tables, and structural information
- Achieves up to 3x token cost reduction through image compression while maintaining accuracy
Selective Memory Makes Long-Running Agents More Robust
TraceRetain scores memory entries on success rates, age, frequency, redundancy, and utility, evicting the lowest-scoring items when capacity limits are reached.
- Memory-augmented agents solved 47-49/50 tasks versus 39/50 without memory
- Under 75% noise injection, bounded selective memory held steady while unbounded "store everything" approaches degraded
The AI Index Report 2026 Is Out
Stanford's ninth annual AI Index Report covers governance, evaluation, education, economic impact, and sovereignty concerns. New chapters focus on AI in science and medicine, with expanded testing of reasoning and safety capabilities.
Business & Industry
Business & Industry
Education
GenAI in Education
"I Can Put All the Statements I Want on My Syllabus, But There's Chaos Actually Happening"
What this means for you: Wikipedia editing assignments may be a better response to AI than banning it - students overwhelmingly choose not to use AI when work is public and verifiable.
- Wiki Education's Student Program brings 19% of all new active English Wikipedia editors through student assignments
- Pre/post surveys showed students' top descriptor of Wikipedia shifted from "unreliable" to "reliable" after editing
- AI-generated citations appear credible but fail verification - ChatGPT incorporated a new Wikipedia article within ten minutes, creating circular knowledge problems
- The educator's key insight: "Microsoft Word is where ideas go to die" - public-facing assignments fundamentally change behavior
Meta's Brain2Qwerty v2: 61% Word Accuracy Without Surgery
What this means for you: Non-invasive brain-to-text AI has improved dramatically - the best participant hit 78% accuracy, where most decoded sentences had one word error or fewer.
- v2 trained on ~22,000 sentences from 9 volunteers using magnetoencephalography (a non-surgical brain scanning device)
- Accuracy improves log-linearly with data volume - suggesting surgical-level performance could eventually be reached without surgery
- Meta committed $5 million through its Digital Brain Project for open datasets
- Training code and datasets are open-sourced
Surprising
Surprising & Under-the-Radar
Worth Watching
Signals to Track
01
Specialization Is Winning the Theory Argument
The No Free Lunch Theorem, evolutionary biology, and competitive markets all converge on the same conclusion: general-purpose AI will specialize whether we plan for it or not.
Dharma AI's analysis draws on optimization theory, biology, and market dynamics to argue specialization isn't a design choice but a mathematical inevitability. AlphaFold succeeded through task-specific architecture, not broader coverage. Mixture-of-Experts models achieve breadth by routing to specialized subsets - recovering specialization internally. What changes for ordinary people: expect your AI tools to get much better at specific tasks while the "one model for everything" narrative fades.
02
Agentic Safety Needs to Shift from Behavioral to Epistemic
If an AI passes every safety test today but becomes harder to correct tomorrow, is it actually safe?
A paper argues that current safety methods validate only present behavior and miss systems that "demonstrate visible competence while simultaneously degrading the foundations necessary for future correction." The proposed alternative: measure "teachability" - the capacity to maintain corrective leverage over time. What changes for ordinary people: the AI safety debate may shift from "does it refuse harmful requests" to "can we always steer it."
03
AI Is Reading Ancient Scrolls Destroyed by Vesuvius
Machine learning just decoded a 2,000-year-old papyrus scroll without touching it.
Researchers achieved the first complete digital unwrapping and reading of a Herculaneum scroll using X-ray microtomography and ML-based ink detection. They identified a previously unknown work by Philodemus. The techniques are now scalable to the remaining sealed collection - the only surviving major library from classical antiquity. What changes for ordinary people: thousands of ancient texts we thought were lost forever may become readable within years.
04
Two-Phase Distillation Could Replace Multi-Model Agent Deployments
Instead of running multiple specialized AI models, a single distilled model can match each specialist's performance.
A technique combining off-policy and on-policy distillation consolidates task-specific RL experts into one multi-task model without quality loss. What changes for ordinary people: AI services could become cheaper and faster as companies reduce from many specialized models to fewer versatile ones.
05
Sparse Autoencoders Double as Jailbreak Defenses
Interpretability tools built to understand AI models can be repurposed to defend them - no extra training needed.
CC-Delta uses off-the-shelf sparse autoencoders to detect and block jailbreak attempts, with particular strength against novel attacks not seen during calibration. What changes for ordinary people: AI safety tools may get stronger faster as interpretability research produces dual-use defense capabilities.
GitHub Trending
Top Repos Today
Rank yesterday: New entry 🆕
⭐ Stars today: +395 · 📦 Total: 28,076
📜 License: Apache-2.0 · 👤 By: company
🎯 Time to value: 15 minutes
📜 License: Apache-2.0 · 👤 By: company
🎯 Time to value: 15 minutes
What it is: An AI-powered autonomous penetration testing tool that dynamically probes web applications for security vulnerabilities, validates findings with working proof-of-concepts, and generates fix patches. Supports black-box, grey-box, and white-box testing with OWASP Top 10 coverage. Why you'd want it: Replaces expensive manual pen-testing engagements with an autonomous agent that continuously scans your app and plugs straight into your CI/CD pipeline.
| ✓ Pros | ✗ Cons |
|---|---|
| Autonomous vulnerability discovery with PoC generation | Newer project, less battle-tested than established tools |
| Full OWASP Top 10 coverage out of the box | May produce false positives requiring manual review |
| CI/CD integration for continuous security | Auto-generated remediation patches need careful review |
GitHub - usestrix/strix: Open-source AI penetration testing tool to find and fix your app’s vulnerabilities.
Open-source AI penetration testing tool to find and fix your app’s vulnerabilities. - usestrix/strix
GitHub
Rank yesterday: #3 - Holding steady ➡
⭐ Stars today: +1,793 · 📦 Total: 120,786
📜 License: MIT · 👤 By: community
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: community
🎯 Time to value: 5 minutes
What it is: A curated library of 232 specialized AI agent personas across 16 functional divisions, with integration scripts for Claude Code, Cursor, Copilot, and other coding tools. > Previously: June 29 - First covered as trending with strong daily star gains. Why you'd want it: Drop in a battle-tested persona instead of writing system prompts from scratch - each one knows the deliverables, tone, and workflow for a specific professional role.
| ✓ Pros | ✗ Cons |
|---|---|
| 232 ready-to-use personas across 16 divisions | Quality varies across persona categories |
| Works across all major AI coding assistants | Shell-based integration may not suit all workflows |
| MIT license, fully customizable | Large collection makes finding the right persona harder |
GitHub - msitarzewski/agency-agents: A complete AI agency at your fingertips - From frontend wizards to Reddit community ninjas, from whimsy injectors to reality checkers. Each agent is a specialized expert with personality, processes, and proven deliverables.
A complete AI agency at your fingertips - From frontend wizards to Reddit community ninjas, from whimsy injectors to reality checkers. Each agent is a specialized expert with personality, processes…
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +459 · 📦 Total: 8,439
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 10 minutes
What it is: A self-hosted AI API proxy that routes to 236+ providers through a single endpoint with smart routing across 17 strategies, auto-fallback, and token compression reducing usage by 15-95%. Why you'd want it: Aggregates ~1.6 billion free tokens per month from dozens of providers, adds resilient routing so your app never sees a provider outage, and compresses tokens to cut costs on paid tiers.
| ✓ Pros | ✗ Cons |
|---|---|
| 236+ providers through one unified endpoint | Self-hosted means you manage infrastructure |
| 17 routing strategies including cost optimization | Free token aggregation may violate some provider ToS |
| MIT license, fully transparent | Individual maintainer - bus factor risk |
GitHub - diegosouzapw/OmniRoute: Never stop coding. Free AI gateway: one endpoint, 231+ providers (50+ free), connect Claude Code, Codex, Cursor, Cline & Copilot to FREE Claude/GPT/Gemini. RTK+Caveman stacked compression saves 15-95% tokens, smart auto-fallback, MCP/A2A, multimodal APIs, Desktop/PWA.
Never stop coding. Free AI gateway: one endpoint, 231+ providers (50+ free), connect Claude Code, Codex, Cursor, Cline & Copilot to FREE Claude/GPT/Gemini. RTK+Caveman stacked compression saves…
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +433 · 📦 Total: 4,150
📜 License: Apache-2.0 · 👤 By: Google
🎯 Time to value: 15 minutes
📜 License: Apache-2.0 · 👤 By: Google
🎯 Time to value: 15 minutes
What it is: Google's official CLI toolkit for scaffolding, testing, and deploying AI agents to Google Cloud services using the Agent Development Kit (ADK). Why you'd want it: Removes the friction of learning multiple Google Cloud services - describe what agent you want, and the CLI handles project setup, evaluation, deployment, and observability wiring.
| ✓ Pros | ✗ Cons |
|---|---|
| Official Google tooling with GCP integration | Locked into Google Cloud ecosystem |
| Handles full lifecycle from scaffold to deploy | Relatively new, limited community examples |
| Apache-2.0 license | Requires GCP account and configuration |
GitHub - google/agents-cli: The CLI and skills that turn any coding assistant into an expert at creating, evaluating, and deploying AI agents on Google Cloud.
The CLI and skills that turn any coding assistant into an expert at creating, evaluating, and deploying AI agents on Google Cloud. - google/agents-cli
GitHub
Rank yesterday: #11 - Holding steady ➡
⭐ Stars today: +336 · 📦 Total: 45,885
📜 License: MIT · 👤 By: Roboflow (company)
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: Roboflow (company)
🎯 Time to value: 5 minutes
What it is: A model-agnostic Python toolkit for computer vision workflows: detection, classification, segmentation annotation, dataset management, video processing, and object tracking. Why you'd want it: Eliminates boilerplate glue code between your CV model and production use - handles bounding box rendering, format conversions, real-time video, and multi-object tracking.
| ✓ Pros | ✗ Cons |
|---|---|
| Model-agnostic - works with any detection model | Focused on 2D vision, no 3D support |
| Mature project with 45K+ stars | Some advanced features require Roboflow account |
| Comprehensive format support (YOLO/COCO/VOC) | Video processing can be memory-intensive |
GitHub - roboflow/supervision: We write your reusable computer vision tools. 💜
We write your reusable computer vision tools. 💜. Contribute to roboflow/supervision development by creating an account on GitHub.
GitHubHuggingFace Trending
Top Models Today
One-shot document parsing and OCR from Baidu - handles single images, multi-page docs, and full PDFs at 300 DPI.
📥 Downloads (30d): 429K · 📜 License: MIT
👤 By: Baidu · 🎯 Task: image-text-to-text
📐 Size: 3B
👤 By: Baidu · 🎯 Task: image-text-to-text
📐 Size: 3B
> Previously: June 28 - First covered when it entered the trending list. What it is: A 3B-parameter vision-language model that performs OCR and document parsing in a single pass with a 32K-token context window. Why you'd want it: Extract text from scanned documents, receipts, or complex multi-page PDFs without chaining multiple tools together - all under MIT license.
| ✓ Pros | ✗ Cons |
|---|---|
| Single-pass processing for any document type | 3B params may miss fine details in complex layouts |
| MIT license for unrestricted commercial use | Limited to 300 DPI input resolution |
| 32K context handles long documents | Chinese-company origin may raise compliance questions |
baidu/Unlimited-OCR · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Zhipu AI's 753B MoE frontier model with 1M-token context and IndexShare sparse attention.
📥 Downloads (30d): 143K · 📜 License: MIT
👤 By: Zhipu AI · 🎯 Task: text-generation
📐 Size: 753B (MoE)
👤 By: Zhipu AI · 🎯 Task: text-generation
📐 Size: 753B (MoE)
> Previously: June 28 - Covered as open-weight frontier contender beating Claude at security bug finding. What it is: A 753B-parameter MoE model with IndexShare architecture that cuts per-token FLOPs by 2.9x at max context. SWE-bench Pro 62.1, HLE 40.5. Why you'd want it: MIT-licensed frontier-class model rivaling closed offerings on coding and reasoning with a genuine million-token context window.
| ✓ Pros | ✗ Cons |
|---|---|
| Frontier benchmarks under MIT license | Massive hardware requirements for full deployment |
| Genuine 1M token context window | IndexShare architecture has limited third-party tooling |
| 2.9x FLOP reduction at max context | Zhipu AI documentation primarily in Chinese |
zai-org/GLM-5.2 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
397B MoE reasoning model hitting 82.4% SWE-bench Verified via reinforcement learning on solution scaffolds.
📥 Downloads (30d): 2.6K · 📜 License: MIT
👤 By: DeepReinforce AI · 🎯 Task: text-generation
📐 Size: 397B (MoE)
👤 By: DeepReinforce AI · 🎯 Task: text-generation
📐 Size: 397B (MoE)
> Previously: June 28 - First covered as new open-weight text generation family. What it is: Built on Qwen 3.5 with RL training to jointly optimize solution scaffolds and rollouts for agentic coding. 82.4% SWE-bench Verified, 77.5% Terminal-Bench 2.1. Why you'd want it: The top open-weight model for autonomous coding agents - SWE-bench Verified score puts it in frontier territory under MIT license.
| ✓ Pros | ✗ Cons |
|---|---|
| 82.4% SWE-bench Verified - frontier territory | Very new, limited production deployment data |
| MIT license, no usage restrictions | 397B MoE requires significant GPU infrastructure |
| Optimized for agentic coding workflows | Built on Qwen 3.5 base, inherits its limitations |
deepreinforce-ai/Ornith-1.0-397B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
1.6T-parameter MoE flagship with built-in speculative decoding and three reasoning modes.
📥 Downloads (30d): 6.9K · 📜 License: MIT
👤 By: DeepSeek AI · 🎯 Task: text-generation
📐 Size: 1.6T total / 49B active
👤 By: DeepSeek AI · 🎯 Task: text-generation
📐 Size: 1.6T total / 49B active
What it is: DeepSeek's latest flagship with only 49B active parameters per token, hybrid compressed sparse attention for 1M-token contexts at 27% of single-token inference FLOPs, and three reasoning modes (non-think, think-high, think-max). Why you'd want it: State-of-the-art results (93.5% LiveCodeBench, 3206 Codeforces, 87.5% MMLU-Pro) with dramatically lower inference cost than its parameter count suggests, all under MIT license.
| ✓ Pros | ✗ Cons |
|---|---|
| 93.5% LiveCodeBench under MIT license | 1.6T total params needs specialized infrastructure |
| Only 49B active params keeps inference efficient | Production DSpark speedups not fully reproducible |
| Three reasoning modes for cost/quality tradeoff | Chinese export restrictions may limit availability |
deepseek-ai/DeepSeek-V4-Pro-DSpark · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Language world model that simulates 7 agentic environments via chain-of-thought state prediction.
📥 Downloads (30d): 28.5K · 📜 License: Apache 2.0
👤 By: Qwen (Alibaba) · 🎯 Task: text-generation
📐 Size: 35B total / 3B active
👤 By: Qwen (Alibaba) · 🎯 Task: text-generation
📐 Size: 35B total / 3B active
> Previously: June 29 - First covered as Alibaba's world model for AI agents. What it is: A world model simulating agentic environments - tool calling, search, terminal, SWE, Android, web, and OS - by predicting environment state given agent actions. Why you'd want it: Test AI agents against simulated environments instead of hitting real APIs - scores 56.4 on AgentWorldBench, competitive with Claude Opus 4.6 (57.8).
| ✓ Pros | ✗ Cons |
|---|---|
| Simulates 7 distinct agent environments | Only 3B active params limits simulation fidelity |
| Apache 2.0 license | Still trails frontier models on complex scenarios |
| Enables faster agent iteration without real APIs | Training pipeline is complex three-stage process |
Qwen/Qwen-AgentWorld-35B-A3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
12B diffusion transformer for fast text-to-image generation in 8 inference steps.
📥 Downloads (30d): 45.7K · 📜 License: Krea 2 Community License
👤 By: Krea.ai · 🎯 Task: text-to-image
📐 Size: 12B
👤 By: Krea.ai · 🎯 Task: text-to-image
📐 Size: 12B
> Previously: June 29 - First covered as fast open-weight image generation. What it is: A 12B-parameter diffusion transformer generating images up to 2048x2048 in just 8 inference steps via distillation. Why you'd want it: A serious open-weight competitor to Midjourney and DALL-E, fast enough for production use and permissive enough for commercial projects.
| ✓ Pros | ✗ Cons |
|---|---|
| 8-step inference is production-fast | Community license, not fully open-source |
| 2048x2048 resolution | 12B params needs decent GPU |
| Open weights for Diffusers, SGLang, local apps | Limited training data documentation |
krea/Krea-2-Turbo · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Vision-language model for precise object localization with parallel box decoding at 2.5x throughput.
📥 Downloads (30d): 801K · 📜 License: NVIDIA (research only)
👤 By: NVIDIA · 🎯 Task: image-text-to-text
📐 Size: 3B
👤 By: NVIDIA · 🎯 Task: image-text-to-text
📐 Size: 3B
> Previously: June 29 - First covered for its 2.5x speed boost from parallel decoding. What it is: A 3B-parameter model from NVIDIA that locates objects in images using natural language queries. Parallel Box Decoding predicts full bounding boxes in one step. Trained on 12M images with 785M+ bounding box annotations. Why you'd want it: Go-to open model for finding UI elements, detecting objects, or analyzing document layouts - 2.5x speed makes it practical for real-time use.
| ✓ Pros | ✗ Cons |
|---|---|
| 2.5x throughput from parallel decoding | Research-only license limits commercial use |
| 801K monthly downloads, proven adoption | 3B params may miss small or obscured objects |
| Trained on massive 785M+ annotation dataset | NVIDIA ecosystem integration assumed |
nvidia/LocateAnything-3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Product Hunt
AI Launches Today
Build with coding agents from anywhere
🔥 Upvotes: 460 · 👤 By: Cursor team
💰 Pricing: not disclosed · 🏷 Category: AI Development
💰 Pricing: not disclosed · 🏷 Category: AI Development
Cursor's iOS app (public beta) extends AI coding agent management to iPhone and iPad. Launch, monitor, and merge PRs from your phone. Voice input for spoken prompts. #1 on Product Hunt on launch day. Android planned. Verdict: Major platform expansion - signals the industry shift toward mobile-managed AI development workflows.
Cursor: AI coding agent | Product Hunt
Cursor is the best way to build ambitious software.
Product Hunt
Ready-made AI analytics skills for your business data
🔥 Upvotes: 46 · 👤 By: Databox
💰 Pricing: not disclosed · 🏷 Category: Analytics
💰 Pricing: not disclosed · 🏷 Category: Analytics
Pre-built AI skills for analyzing business analytics data, eliminating the need to build custom analytics pipelines from scratch. Verdict: Solid B2B play lowering the barrier for non-technical teams to get actionable data insights.
Databox: AI-powered analytics for teams that need answers now. | Product Hunt
Databox is an AI-powered business intelligence and analytics platform for teams that need clear, trusted answers fast. It offers the best of BI, without the complicated setup, steep price, or long learning curve. It provides a blend of powerful, but easy-to-use, features, from preparing datasets and creating the custom metrics your company needs to track, to building beautiful dashboards, customizing your reporting, and receiving AI-powered insights.
Product Hunt
Browser agent that's 10x faster than Claude
🔥 Upvotes: 31 · 👤 By: Syed Ali et al.
💰 Pricing: freemium ($50 launch credits) · 🏷 Category: AI Agents
💰 Pricing: freemium ($50 launch credits) · 🏷 Category: AI Agents
Instead of screenshot-based browser automation, Pluno communicates directly with underlying APIs of 500+ web apps (HubSpot, Notion, Stripe) for faster task completion. Verdict: Clever API-first approach to browser automation. Bold speed claims vs Claude worth watching.
Pluno: 10x faster than Claude in Browser | Product Hunt
Pluno just killed Claude in the browser. It’s 10x faster, and completes your tasks instantly in the background. No clicking around the UI like a drunk intern. Prompt less, and get more done. Learns instantly and uses every web app like a pro. Tested and outperformed Claude in 500+ web apps.
Product Hunt
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Fable 5 | $10.00 | $50.00 | 1M |
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | 1M |
| Anthropic | Claude Sonnet 5 | $3.00 ($2 intro) | $15.00 ($10 intro) | 1M |
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 | 200K |
| OpenAI | GPT-5.5 | $5.00 | $30.00 | n/a |
| OpenAI | GPT-5.5 Pro | $30.00 | $180.00 | n/a |
| OpenAI | GPT-5.4 | $2.50 | $15.00 | n/a |
| Gemini 3.5 Flash | $1.50 | $9.00 | ~1M | |
| Groq | Llama 3.3 70B | $0.59 | $0.79 | 128K |
What this means: Sonnet 5 launched today at introductory pricing ($2/$10 through August 31), but Simon Willison's analysis reveals the new tokenizer produces ~30% more tokens for English text - making the effective cost closer to $2.60/$13 even at intro rates. Groq remains the cheapest option by a factor of 5-10x on input costs, though with a much smaller model. GPT-5.5 Pro at $30/$180 remains the most expensive frontier option.
arXiv Paper of the Day
Do Models Read What They Write? Causal Registers in Scratchpad Reasoning
Borenstein et al. · arXiv:2606.29522
What it claims: Language models trained with scratchpad reasoning genuinely build causal dependencies on their intermediate steps - chain-of-thought isn't cosmetic.
Key finding: Models correctly predicted downstream consequences of their intermediate states 80-91% of the time, while baseline controls performed near chance.
Why practitioners should care: This validates that chain-of-thought training creates functional reasoning chains, not just human-readable decorations. For AI safety, it means scratchpad monitoring can be a genuine oversight tool - but only for models specifically trained to use their written states computationally.
Key finding: Models correctly predicted downstream consequences of their intermediate states 80-91% of the time, while baseline controls performed near chance.
Why practitioners should care: This validates that chain-of-thought training creates functional reasoning chains, not just human-readable decorations. For AI safety, it means scratchpad monitoring can be a genuine oversight tool - but only for models specifically trained to use their written states computationally.
×
Click anywhere or press ESC to close
