Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
936
advertiser trackers, 1,029 websites
ChatGPT has been quietly tracking you across the wider web
Top Story
200
x faster, 400x cheaper
Sorting beats writing for bulk work
One Thing to Tell Your Friends
One Thing to Tell Your Friends
ChatGPT has quietly built an ad-tracking network that can follow you across more than 1,000 other websites - and it still tags you even when you are logged out.
Summary
TL;DR
Top Stories
Trends
AI companies are quietly becoming data and advertising businesses, Sorting beats writing for bulk work, and Open AI models are being treated as public heritage.
Creative AI
Open video generation keeps climbing the charts and Open.
Dev Tools
A safer way to hand AI agents your secret keys and Inspect your database queries without leaving the page.
Research
What actually makes an AI coding assistant good and New open models keep raising the bar.
Business
Surprising
Should your AI ever tell you no?, Even signed, and 669,000 models is a measure of how much open AI there now is.
Worth Watching
Secure "key vaults" for AI agents, "Sovereign AI" as a movement, not a slogan, and Classification endpoints replacing chatbots behind the scenes.
GitHub
Leading repos: cloudflare/security-audit (+2,375), trycua/cua (+1,012), and addyosmani/agent (+729).
HuggingFace
Leading models: Qwen/Qwen3.8 (7.33), deepseek-ai/DeepSeek-V4.1 (497,000), and Lightricks/LTX (1.61).
Product Hunt
Top launches: Ami AI (525), AINA (475), and MosMos (409).
API Pricing
What this means: The cheapest capable models keep getting cheaper.
arXiv
An Empirical Study of Harness Design for Coding Agents — Managing the model's working context becomes more valuable as its memory budget tightens, and combining simple rule-based filtering with AI summarization delivers the best efficiency.
FYI
Hot off the Presses
01
ChatGPT has been quietly tracking you across the wider web
What this means for you: The chatbot you tell your most personal questions to may be linking that account to what you buy, read, and search on hundreds of other sites - whether or not you are signed in.
Previously: September 16 - OpenAI began building advertising into ChatGPT.
Today: An independent researcher took apart the network traffic around ChatGPT and found OpenAI is running a full advertising and conversion-tracking system, internally named "Bazaar," at a hidden web address (bzr.openai.com). It works like the ad-tracking machinery already run by Meta (Facebook's owner) and Google: a small piece of tracking code, called a pixel, sits on advertiser websites and reports your activity back. The new and unusual part is that this one is tied to an artificial intelligence chat product, where people share far more intimate details than they ever would on social media.
OpenAI's support team acknowledged the researcher's questions but did not answer them. The test only covered Chrome on Android, and the link back to your account was strongly inferred rather than directly proven - but the tracking cookie itself is plainly visible.
“On one device, a single tracking ID was seen following the user across 12 different shopping sites.”
- Your account gets a hidden ID - ChatGPT plants a cookie (a small tracking file) called
__obithat lasts a full year and is deliberately set to travel with you to other sites. - 936 advertiser trackers, 1,029 websites - the researcher counted OpenAI's pixel across a huge span of the web, including named retailers like Chewy, Wayfair, HelloFresh, and Coursera.
- Logged out does not mean private - signed-out users still get an "anonymous" device ID that persists at least 27 days.
- Some data travels in the clear - email, phone and names are scrambled first, but your country, region, city and postal code are sent as plain readable text.
02
A new site "rescues" free AI models by turning them into torrents
What this means for you: The free, open AI models that hobbyists and small companies rely on can vanish overnight when a host deletes them - this project makes them permanent.
Pirate Face is a preservation service that takes open-source AI models from Hugging Face (the main public library for free AI models) and converts them into peer-to-peer torrents - the same file-sharing technology used to share large files without a central server. While a model is still hosted normally, downloads come straight from Hugging Face. The moment a model is removed, a swarm of volunteer computers keeps it alive, and models marked "Rescued" stay downloadable indefinitely.
The team frames it as "permanent infrastructure for sovereign AI," protecting open models from takedowns, license changes, or corporate policy shifts. It treats open model weights the way libraries treat out-of-print books: public heritage worth keeping alive.
- Over 669,000 models are eligible - the service mirrors freely-licensed models (Apache-2.0 and MIT) synced live from Hugging Face.
- Downloads are guaranteed identical - every file carries an official checksum (a digital fingerprint), so rescued weights exactly match the originals.
- No code changes needed - a drop-in setting lets existing projects pull from Pirate Face instead, so nothing breaks.
03
A firsthand account of a company where AI writes nearly everything
What this means for you: This is an early warning about what "AI-first" workplaces can actually feel like - long hours, little thinking, and software nobody fully understands.
Developer and writer Simon Willison highlighted a firsthand account from a software engineer describing a large company where a coding assistant (Claude Code) now produces almost all technical work: the specifications, the code, the tests, the documentation, even the task tickets. Willison, who built widely-used open-source data tools, flagged the post as a cautionary tale, not a success story.
This is a single account, not a study. But it is a concrete, September-2026 snapshot of the difference between AI that supplements human judgment and AI that replaces it.
- 12 to 13 hour days, little real work - staff mostly execute the AI's suggestions rather than solving problems themselves.
- Everyone leans on it, junior to senior - engineers at every level default to asking the assistant instead of using their own judgment.
- The core risk is nobody understands the code - when AI output ships without real review, quality, knowledge, and long-term maintainability quietly erode.
04
The real AI cost story is not your token bill
What this means for you: If you run a business, judging AI purely by its monthly bill can make you miss where it actually pays off - and where a competitor could undercut you.
Analyst Nate argues that fixating on the price of AI usage (measured in "tokens," the small chunks of text an AI processes) misses the bigger prize: AI lowers the cost of serving each customer. He calls the trap a "visibility problem" - the AI bill is a clear, single line item, while the human and operational costs it replaces are scattered and hidden, so a naive comparison always makes AI look expensive.
He backs it with a customer complaint relayed by a Salesforce executive: despite promises that AI would "change everything," many products still feel unchanged - the gap between what AI can do and what businesses have actually rebuilt.
- Old pricing rules are up for grabs - service tiers and account minimums were built around expensive human attention, and all of it becomes renegotiable as automation gets cheap.
- Savings are not automatic profit - a rival can pour the same efficiency into lower prices or better service, so who keeps the gains is a fight.
- Real change means new outcomes - genuine transformation is offering customers things that were impossible before, not bolting an AI feature onto an old product.
Trends & Themes
Trends & Themes
AI companies are quietly becoming data and advertising businesses
Why this matters to you: The "free" or cheap AI you use is increasingly paid for with your data, the same bargain that shaped social media.
The pattern: the value in consumer AI is drifting from the cleverness of the model toward the data it can collect, sort, and act on. That is exactly the shift that turned search and social into advertising empires.
- The tracking is now visible - the ChatGPT ad network (a Top Story today) mirrors the cross-site tracking Meta and Google pioneered.
- Classification-for-hire is booming - a new tool called Jev (covered September 19) sorts and tags bulk personal data like contacts and inboxes for pennies.
- Analysts are reframing AI economics around the customer - Nate's cost-to-serve argument (a Top Story today) treats customer data and outcomes, not model size, as the real battleground.
Sorting beats writing for bulk work
Why this matters to you: A lot of "AI" work does not need paragraphs of generated text - it needs fast, cheap decisions, and that is getting radically cheaper.
When the job is bucketing, tagging, or routing, a "multiple-choice" AI call is dramatically cheaper than a "write me a paragraph" call. Expect more products to quietly swap expensive text generation for cheap classification behind the scenes.
- 200x faster, 400x cheaper - the Jev classification tool answers with a label and a confidence score instead of an essay, and reportedly tagged 348 research papers for 18 cents.
- The research agrees - today's arXiv paper of the day finds that smart handling of context, not brute force, drives efficiency in AI coding agents.
- Cost guides are going mainstream - Nate's newsletter shipped a 15-technique "token saver" guide aimed at cutting waste.
Open AI models are being treated as public heritage
Why this matters to you: The free models that power countless apps and hobby projects are fragile, and a movement is forming to make them permanent.
Centralized hosting means a takedown, license change, or policy shift can make widely-used weights disappear. The response is content-addressed, censorship-resistant distribution - the same idea that keeps other digital archives alive.
- A torrent lifeboat launched - Pirate Face (a Top Story today) mirrors 669,000+ open models so takedowns cannot erase them.
- Open models keep topping the charts - freely downloadable models from Qwen, DeepSeek, and others dominate this week's trending list (see Sections 7 and 13).
- The framing is political - Pirate Face calls its work "sovereign AI," a hedge against any single company controlling access.
The hard question is shifting from "can AI do it?" to "should it, and who checks?"
Why this matters to you: As AI gets more capable, the debates that affect you are less about raw ability and more about trust, ethics, and oversight.
The frontier of the conversation has moved. The interesting fights now are about guardrails, disclosure, and who is accountable when an AI acts.
- Should your AI ever refuse you? - writer Zvi Mowshowitz argues AI in professional roles should follow ethical codes (like a lawyer or doctor), not blind loyalty.
- Over-reliance has a human cost - the "AI writes everything" account (a Top Story today) shows what happens when nobody reviews the output.
- Verification is becoming a product - a security-audit tool that independently double-checks AI findings is among the fastest-rising projects on GitHub today (see Section 12).
Creative AI & Media
Creative AI & Media
Open video generation keeps climbing the charts
- LTX-2.5 turns a still image into short video and is one of the most-downloaded media models this week (1.61 million downloads).
- What it lets you do - animate a single photo or piece of art into a short clip without a studio or subscription.
- Try it: Lightricks/LTX-2.5 on Hugging Face
Open-source music generation you can run yourself
- YuE2-3B is a text-to-audio model for generating music, trending with strong community interest.
- What it lets you do - create original backing tracks or song sketches from a text description, on your own computer, no per-song fee.
- Try it: m-a-p/YuE2-3B on Hugging Face
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
What actually makes an AI coding assistant good
Why this matters: The secret to better AI coding help is less about a bigger AI brain and more about how you feed it information - which means cheaper models can punch above their weight.
- The finding - a large study varied how coding agents plan, act, and manage information across four different AI models.
- Key result - managing what the AI "remembers" matters more as its memory budget shrinks, and mixing simple filtering with AI summarizing works best.
- The twist - for weaker models, planning improves accuracy; for stronger models, planning mainly cuts cost.
- See Section 16 for the full paper.
New open models keep raising the bar
- DeepSeek-V4.1-Flash and Qwen3.8-27B are among the most-downloaded models this week, both freely available (see Section 13).
- Why it matters - these are the kinds of open models that services like Pirate Face (Top Story) are racing to preserve, and they power a large share of independent AI projects.
Business & Industry
Business & Industry
Surprising
Surprising & Under-the-Radar
Worth Watching
Signals to Track
01
Secure "key vaults" for AI agents
Why this is worth watching right now: as people run AI agents on remote machines, the messy problem of handing them secrets safely is finally getting real tools.
Today's llm-keys-ui is a tiny example of a pattern that will matter a lot: giving autonomous AI agents access to sensitive credentials without exposing those secrets in a chat log. Expect this to become standard plumbing. For ordinary people, it is the difference between AI helpers that are convenient and ones that quietly leak your passwords.
02
"Sovereign AI" as a movement, not a slogan
Why this is worth watching right now: Pirate Face turns an abstract worry - losing access to open AI - into working infrastructure.
The idea that open models are public heritage worth preserving in censorship-resistant networks is moving from talk to code. If it catches on, it changes who ultimately controls whether a useful AI stays available. For everyday users, it means the free tools you depend on are less likely to disappear at a company's whim.
03
Classification endpoints replacing chatbots behind the scenes
Why this is worth watching right now: the cheapest AI trick in business is quietly swapping "generate text" for "pick a label."
Tools like Jev hint at a broad shift: for sorting, tagging, and routing, companies are switching from expensive text generation to cheap "multiple-choice" AI. You will rarely see it, but it will make many apps faster and cheaper - and it means more of your data is being auto-sorted than you might realize.
GitHub Trending
Top Repos Today
Rank yesterday: New entry 🆕
⭐ Stars today: +2,375 · 📦 Total: 18,000
📜 License: MIT · 👤 By: company (Cloudflare)
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: company (Cloudflare)
🎯 Time to value: 15 minutes
What it is: A tool that lets an AI coding assistant run a thorough, multi-step security review of your code. It works in phases - exploring the code, hunting for issues, then independently double-checking each finding before reporting it. Why you'd want it: It automates the kind of careful security audit that normally requires an expert, and its self-verification step cuts down on false alarms.
| ✓ Pros | ✗ Cons |
|---|---|
| Independent verification reduces false positives | Requires an AI coding agent to run |
| Backed by a major infrastructure company | New project, still maturing |
| Free and open (MIT license) | Security output still needs human sign-off |
GitHub - cloudflare/security-audit-skill: A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings - cloudflare/security-audit-skill
GitHub
Rank yesterday: Rising ↑
⭐ Stars today: +1,012 · 📦 Total: 25,100
📜 License: MIT · 👤 By: startup
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: startup
🎯 Time to value: 30 minutes
What it is: An open platform for building AI that can control a computer - clicking, typing, and navigating apps - with isolated cloud desktops and local Mac virtual machines to run them safely. Why you'd want it: If you want to build or test an AI that operates real software, this gives you the sandbox and tools to do it without risking your own machine.
| ✓ Pros | ✗ Cons |
|---|---|
| Safe, isolated environments for risky automation | Computer-use AI is still error-prone |
| Cross-platform (Mac VMs plus cloud) | Setup takes some technical comfort |
| Includes benchmarks to measure quality | Broad scope can feel complex |
GitHub - trycua/cua: Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation.
Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation. - trycua/cua
GitHub
Rank yesterday: Rising ↑
⭐ Stars today: +729 · 📦 Total: 97,700
📜 License: MIT · 👤 By: individual (Google engineer)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: individual (Google engineer)
🎯 Time to value: 10 minutes
What it is: A free library of 25 "skills" - written playbooks that teach AI coding assistants the workflows and quality checks that experienced engineers use, covering planning, building, testing, and shipping. Why you'd want it: It makes an AI coding helper behave more like a disciplined senior engineer instead of a fast but sloppy junior.
| ✓ Pros | ✗ Cons |
|---|---|
| Huge, well-organized set of proven workflows | You still supply the AI agent to run them |
| Plain-text, easy to read and adapt | Value depends on your agent's quality |
| Very popular and actively used | Not a standalone app |
GitHub - addyosmani/agent-skills: Production-grade engineering skills for AI coding agents.
Production-grade engineering skills for AI coding agents. - addyosmani/agent-skills
GitHub
Rank yesterday: Rising ↑
⭐ Stars today: +461 · 📦 Total: 5,400
📜 License: Apache-2.0 · 👤 By: startup
🎯 Time to value: 60 minutes
📜 License: Apache-2.0 · 👤 By: startup
🎯 Time to value: 60 minutes
What it is: A system for reliably training very large AI models across many graphics chips, designed to keep going even when individual machines fail. Why you'd want it: For teams training big models, hardware failures are constant; this handles them automatically so long training runs do not collapse.
| ✓ Pros | ✗ Cons |
|---|---|
| Fault tolerance for long, expensive runs | Only relevant if you train large models |
| Scales to very large model sizes | Requires serious graphics processing unit (GPU) hardware |
| Open license (Apache-2.0) | Steep learning curve |
GitHub - higgsfield-ai/higgsfield: Fault-tolerant, highly scalable GPU orchestration, and a machine learning framework designed for training models with billions to trillions of parameters
Fault-tolerant, highly scalable GPU orchestration, and a machine learning framework designed for training models with billions to trillions of parameters - higgsfield-ai/higgsfield
GitHub
Rank yesterday: Holding steady ➡
⭐ Stars today: +382 · 📦 Total: 16,000
📜 License: AGPL-3.0 · 👤 By: startup
🎯 Time to value: 45 minutes
📜 License: AGPL-3.0 · 👤 By: startup
🎯 Time to value: 45 minutes
What it is: A self-hosted system for creating cloud-based coding workspaces, now built to run AI coding agents on your own servers without exposing your AI keys inside the workspace. Why you'd want it: It lets a company give developers (and their AI agents) consistent, secure environments that it fully controls.
| ✓ Pros | ✗ Cons |
|---|---|
| Keeps AI agents on your own infrastructure | Self-hosting takes real setup effort |
| Standardized, reproducible workspaces | AGPL license can deter some businesses |
| Mature and widely deployed | Overkill for solo developers |
GitHub - coder/coder: Secure environments for developers and their agents
Secure environments for developers and their agents - coder/coder
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +89 · 📦 Total: 5,200
📜 License: MIT · 👤 By: startup
🎯 Time to value: 20 minutes
📜 License: MIT · 👤 By: startup
🎯 Time to value: 20 minutes
What it is: A framework for building applications that are designed around AI agents from the ground up, rather than adding AI onto a traditional app. Why you'd want it: If you are starting a new AI-first product, it provides the scaffolding so you do not have to invent the plumbing yourself.
| ✓ Pros | ✗ Cons |
|---|---|
| Purpose-built for agent applications | Young project, small community |
| Clean, modern developer experience | Fewer examples than mature frameworks |
| Free and open (MIT) | Best for new builds, not retrofits |
GitHub - BuilderIO/agent-native: A framework for building agentic apps
A framework for building agentic apps. Contribute to BuilderIO/agent-native development by creating an account on GitHub.
GitHub
HuggingFace Trending
Top Models Today
A powerful open model from Alibaba that handles both text and images, topping this week's downloads.
📥 Downloads (30d): 7.33 million · 📜 License: open (Apache-style)
👤 By: Alibaba (Qwen team) · 🎯 Task: text + image understanding
📐 Size: 27B
👤 By: Alibaba (Qwen team) · 🎯 Task: text + image understanding
📐 Size: 27B
What it is: A large, freely-downloadable AI model that can read both text and pictures and respond in text. It is a general-purpose workhorse for people who want a capable model they can run themselves. Why you'd want it: Frontier-style capability without a subscription, and you keep your data on your own hardware.
| ✓ Pros | ✗ Cons |
|---|---|
| Handles text and images together | 27B size needs a strong computer |
| Free to download and run | Setup requires technical skill |
| Huge, active user base | Not as simple as a hosted chatbot |
Qwen/Qwen3.8-27B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A fast, large open model tuned for quick responses across text and images.
📥 Downloads (30d): 497,000 · 📜 License: open
👤 By: DeepSeek · 🎯 Task: text + image understanding
📐 Size: very large (mixture-of-experts)
👤 By: DeepSeek · 🎯 Task: text + image understanding
📐 Size: very large (mixture-of-experts)
What it is: A speed-focused version of DeepSeek's latest open model, built to answer quickly while staying capable. "Flash" signals it is optimized for fast, cheaper responses. Why you'd want it: A high-end open model for teams that need speed and want to avoid ongoing per-use fees.
| ✓ Pros | ✗ Cons |
|---|---|
| Fast responses at large scale | Very large; demanding to self-host |
| Free and open | Best on serious hardware |
| Strong momentum and support | Newer, less battle-tested |
deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A popular open model that turns still images into short videos.
📥 Downloads (30d): 1.61 million · 📜 License: open
👤 By: Lightricks · 🎯 Task: image-to-video
📐 Size: mid-size
👤 By: Lightricks · 🎯 Task: image-to-video
📐 Size: mid-size
What it is: An AI model that animates a single photo or illustration into a short video clip. It is one of the most-downloaded creative models this week. Why you'd want it: Make simple animations or motion clips from artwork without a video studio or subscription.
| ✓ Pros | ✗ Cons |
|---|---|
| Free image-to-video generation | Clips are short |
| Very popular and well-supported | Needs a capable GPU |
| Runs on your own machine | Quality varies by input |
Lightricks/LTX-2.5 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
An open model for generating music from a text description.
📥 Downloads (30d): 17,400 · 📜 License: open
👤 By: M-A-P community · 🎯 Task: text-to-audio (music)
📐 Size: 3B
👤 By: M-A-P community · 🎯 Task: text-to-audio (music)
📐 Size: 3B
What it is: A relatively small, freely-available model that creates music from a written prompt. Its modest size means it can run on more modest hardware than most media models. Why you'd want it: Sketch out original music or backing tracks for videos without licensing fees.
| ✓ Pros | ✗ Cons |
|---|---|
| Smaller size, easier to run | Not studio-quality yet |
| Free and open | Music generation is hit-or-miss |
| Good for prototyping audio | Smaller community than big models |
m-a-p/YuE2-3B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A highly compressed 27B model packaged to run efficiently on everyday hardware.
📥 Downloads (30d): 1.91 million · 📜 License: open
👤 By: prism-ml · 🎯 Task: text generation
📐 Size: 27B (compressed)
👤 By: prism-ml · 🎯 Task: text generation
📐 Size: 27B (compressed)
What it is: A version of a large model that has been shrunk (a process called quantization) and packaged in the "GGUF" format so it runs on ordinary computers rather than data-center chips. Why you'd want it: It brings a big model within reach of a normal laptop or desktop.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs a 27B model on modest hardware | Compression can slightly reduce quality |
| Very high download count | Still needs decent memory |
| Free and open | Format requires compatible software |
prism-ml/Ternary-Bonsai-2-27B-gguf · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Product Hunt
AI Launches Today
Lovable for getting customers.
🔥 Upvotes: 525 · 👤 By: Ami AI team
💰 Pricing: freemium · 🏷 Category: sales / AI assistant
💰 Pricing: freemium · 🏷 Category: sales / AI assistant
An AI sales assistant that focuses on turning leads into customers, aiming to read the emotional cues in a conversation to improve conversion. It targets small teams that cannot afford a full sales operation. Verdict: Strong launch-day interest; the real test is whether "emotionally intelligent" selling holds up beyond the demo. Product Hunt
Find your job search blind spots with an AI coach.
🔥 Upvotes: 475 · 👤 By: AINA team
💰 Pricing: freemium · 🏷 Category: careers / coaching
💰 Pricing: freemium · 🏷 Category: careers / coaching
A career coaching tool that reviews your resume and interview answers and points out weaknesses you might miss. It is aimed at job seekers who want honest, structured feedback. Verdict: Useful niche; value depends on how specific and non-generic its feedback really is. Product Hunt
Voice writing that works before, during, and after meetings.
🔥 Upvotes: 409 · 👤 By: MosMos team
💰 Pricing: freemium · 🏷 Category: productivity / transcription
💰 Pricing: freemium · 🏷 Category: productivity / transcription
A voice-and-meeting tool that transcribes in real time and turns talk into usable notes and summaries. It competes in the crowded meeting-notes space by spanning the whole meeting lifecycle. Verdict: Handy if the summaries are accurate; the category is competitive, so polish matters. Product Hunt
Learn English as you type Chinese.
🔥 Upvotes: 258 · 👤 By: Keysake team
💰 Pricing: freemium · 🏷 Category: education / language
💰 Pricing: freemium · 🏷 Category: education / language
A language-learning tool that gives real-time English feedback while you type in Chinese, blending everyday work with learning. A clever "learn while you do" angle. Verdict: Genuinely novel approach; appeal will hinge on how smooth the typing experience feels. Product Hunt
Every agent accessible via your Mac's notch.
🔥 Upvotes: 185 · 👤 By: Pushary team
💰 Pricing: freemium · 🏷 Category: productivity / macOS
💰 Pricing: freemium · 🏷 Category: productivity / macOS
A Mac tool that puts your AI agents in the notch at the top of the screen for instant access. It bets on ambient, always-available AI rather than a separate app. Verdict: Neat convenience play; whether it earns permanent screen space is the open question. Product Hunt
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Opus 5 | $5.00 | $25.00 | up to 1M |
| Anthropic | Sonnet 5 | $2.00 | $10.00 | up to 1M |
| Anthropic | Haiku 4.5 | $1.00 | $5.00 | up to 1M |
| OpenAI | GPT-5.6 Sol | $5.00 | $30.00 | large |
| OpenAI | GPT-5.6 Terra | $2.00 | $12.00 | large |
| OpenAI | GPT-5.6 Luna | $0.20 | $1.20 | large |
| Gemini 3.8 Flash | $0.75 | $3.75 | large | |
| Gemini 3.1 Pro | $2.00 | $12.00 | large |
Prices below are per one million tokens (a token is a small chunk of text; a million tokens is very roughly 750,000 words). "Input" is what you pay to send text to the model; "output" is what you pay for its reply.
What this means: The cheapest capable models keep getting cheaper. Google's Gemini 3.8 Flash and OpenAI's budget "Luna" tier are priced for high-volume, everyday tasks, while the premium tiers (Opus 5, GPT-5.6 Sol) cost 5 to 25 times more for the hardest work. Establishing baseline; changes will be flagged in future editions. Batch processing typically halves these prices, and caching can cut input costs further.
What this means: The cheapest capable models keep getting cheaper. Google's Gemini 3.8 Flash and OpenAI's budget "Luna" tier are priced for high-volume, everyday tasks, while the premium tiers (Opus 5, GPT-5.6 Sol) cost 5 to 25 times more for the hardest work. Establishing baseline; changes will be flagged in future editions. Batch processing typically halves these prices, and caching can cut input costs further.
arXiv Paper of the Day
An Empirical Study of Harness Design for Coding Agents
Run-Ze Fan, Zihao Zhang, Simin Ma, Yebowen Hu, Shouju Wang, Kaiqiang Song, Fei Liu, Hamed Zamani, Xiaoyang Wang · arXiv:2609.20804
What it claims: The "harness" around an AI coding agent - how it plans, what actions it can take, and how it manages information - matters as much as the underlying model. The team systematically varied these design choices across four different models to see what actually helps.
Key finding: Managing the model's working context becomes more valuable as its memory budget tightens, and combining simple rule-based filtering with AI summarization delivers the best efficiency.
Why practitioners should care: It suggests you can get better, cheaper results from an AI coding assistant by improving its scaffolding rather than paying for a bigger model - and that planning helps weaker models get answers right but mainly helps stronger models save money.
Key finding: Managing the model's working context becomes more valuable as its memory budget tightens, and combining simple rule-based filtering with AI summarization delivers the best efficiency.
Why practitioners should care: It suggests you can get better, cheaper results from an AI coding assistant by improving its scaffolding rather than paying for a bigger model - and that planning helps weaker models get answers right but mainly helps stronger models save money.
×
Click anywhere or press ESC to close
Member discussion