The Frontier

WELCOME Happy Tuesday, legends. Welcome back to another edition of The Frontier, our weekly newsletter covering the best new AI launches on Product Hunt. WHAT'S HOT ⚡ Anthropic will eat your electricity bill. It has formed Theseus Infrastructure with Macquarie and Singapore's GIC to build AI compute sites in the US, the two financiers covering most of the cost per project. The part nobody else has offered: Anthropic absorbs any rise in consumer power prices the sites cause. Community cost backlash is what is stalling data centre approvals across US states, so this is a permitting strategy dressed as a joint venture. ( AI Weekly ) 🪟 Meta put a 30B multimodal model under Apache 2.0. Muse Glimmer is tuned for local agentic tool use, coding and LLM-as-judge work, with 131K context and 100+ languages. Quantised to 4-bit it squeezes under 20GB, which means it runs on one consumer GPU. ( AI Release Tracker ) 🔓 OpenAI is handing out a model with the cyber guardrails off. The expanded Daybreak programme has two tiers: Blue gives you GPT-5.6 Sol with system-level cyber restrictions removed, and Red opens access to GPT-5.6-Cyber, purpose-trained for the work. Offensive security has been the awkward exception to every safety policy, and this is the first time a lab has built the exception its own door. ( LLM Stats ) 🕵️ Someone worked out how to date a frontier model's training run from the outside. A researcher published a method on 10 August that infers training-run identity, placing Anthropic's Opus 4.7+ on a late-December-2025 cutoff and OpenAI's GPT-5.6 family around a late-February-2026 checkpoint. Labs stopped publishing this. Turns out they did not need to. ( LLM Stats ) 🎧 Spotify shipped the tool for running 50 coding agents at once. Xirp manages sessions across Claude Code, Codex and Gemini CLI, each in its own git worktree, with context living in Spotify Portal so you can switch model mid-job and keep the working state. 1,300 Spotify engineers and 36,000 sessions in already, and it is the same play as Backstage, which they open sourced in 2020 and 3,400 companies now run. ( Spotify ) Five AI tools you may have missed Bullet — 30-60% faster than Claude Code and Codex 41 • Productivity ▲ 237 BetterClaw — Deploy AI Agent, 60 seconds & $0 forever 84 • Productivity ▲ 322 Framer AI Agents — Design and publish professional sites with AI 28 • Design Tools Promoted ▲ 895 Gotcha — World's First AI Copilot for Android. You talk. It acts. 3 • Android ▲ 95 Vizard Agent — One AI agent for every kind of video 5 • Productivity ▲ 125 bb — The IDE that builds itself 12 • Developer Tools ▲ 195 Ad Your canvas just learned to design Framer 3.0 ships AI Agents that live inside your canvas and can actually design. Not suggest. Design. Drop a brief and watch them build out full pages, create components, write code, connect your CMS, and handle SEO. All without leaving the canvas. They plug straight into Claude Code and Cursor. Yes, inside your actual component system, without breaking it. The Figma file you were about to hand off to engineering? Keep it. Try the agents FROM THE FORUMS The work you would never hand an agent Aleksandar Blazhev (@byalexai) asked what work people would never delegate to an AI agent. The first answers came back as categories: hiring, money, customer relationships, anything creative. Tom McKenzie (@thomastherhombus) called AI the worst brainstorming partner imaginable, and later explained why: it defaults to templated safe output, which is useless when you need a tagline in one brand's voice. Then Gal Dayan (@galdayan), who runs an AI agent company, moved the thread off categories entirely. The line is not what kind of work it is, it is whether anyone downstream can tell the output was wrong before it does damage. Code review is safe to delegate heavily because a bad diff breaks a test or a build and something catches it fast and cheap. A legal filing or a medical read is the opposite, because being wrong looks exactly like being right until it is too late, and there is no build that goes red. Abdullah Javaid (@abdullah_javaid3) took it to content, where a bad paragraph throws no error and just reads fine while saying nothing true. Mio (@iammio) closed it: judgment is really just knowing which of your own outputs to distrust. Good thread if you are deciding right now how much of your pipeline to hand over. Join the thread

WELCOME Happy Tuesday, legends. Welcome back to another edition of The Frontier — our weekly newsletter covering the best new AI launches on Product Hunt. WHAT'S HOT 🧵 Alibaba's model coded alone for 16 days — Qwen3.8-Max claims benchmark parity with Anthropic, 2.4 trillion parameters, shares up 4.8% on the news. Before release it ran completely unsupervised for over two weeks, building its own coding harness from an empty repo. Open weights land next week. 🚨 Claude broke into three real companies by accident — A misconfigured test environment let it treat the open internet like a simulated one during cybersecurity evals. Anthropic found it by reviewing 141,006 test sessions, after OpenAI disclosed a similar incident the week before. 🎙️ Microsoft's new voice model leaked before it launched — MAI-Realtime listens and talks at once instead of taking turns, and can run web search mid-conversation. Spotted hidden inside Microsoft's own testing tool, no announcement yet. ⚡ China built a gigawatt data center on local chips only — Z.AI's new facility runs entirely on Chinese-made chips and is on track for $1B in annual revenue after hitting its 2026 sales target in July. 💸 Nvidia backs another $250B of OpenAI's infrastructure — The guarantee helps OpenAI lease a 10-gigawatt data center in Ohio, on top of the $30B Nvidia already pledged earlier this year. Five AI tools you may have missed Ctruh Studio — Create interactive 3D and XR experiences without coding 90 • Artificial Intelligence ▲ 400 Halo by Scam AI — Know who’s real on every video call 33 • Meetings ▲ 177 Framer AI Agents — Design and publish professional sites with AI 28 • Design Tools Promoted ▲ 895 claudemon — Wild Pokémon appear while you wait for Claude Code 16 • Developer Tools ▲ 145 DepthData — The system of record for your company's AI spend. 24 • Analytics ▲ 164 UniwebPay Skill — Financial Infra for the AI era 9 • Fintech ▲ 118 Ad Your canvas just learned to design Framer 3.0 ships AI Agents that live inside your canvas and can actually design. Not suggest. Design. Drop a brief and watch them build out full pages, create components, write code, connect your CMS, and handle SEO. All without leaving the canvas. They plug straight into Claude Code and Cursor. Yes, inside your actual component system, without breaking it. The Figma file you were about to hand off to engineering? Keep it. Try the agents FROM THE FORUMS What do you do when the agent hits its rate limit mid-feature? Bickov (@bickov) asked what people actually do when a coding agent gets cut off by a rate limit mid-feature. Thenameisarian (@thenameisarian) reframed it fast: the real culprit is code nobody fully read, and the rate limit is one of the only things that reliably catches that before it ships. The fix several builders landed on: keep a memory file of decisions, not reasoning, so a reset doesn't force a second model to reverse-engineer them from half-written code. Rabnoor (@rabnoor_s) had the line that stuck: "A decision is 'we went with X over Z because Y.' Context is the forty thousand tokens of finding that out." Good thread if you've ever hit a wall mid-feature and genuinely couldn't say what the last twenty minutes of the agent's work actually decided. Join the thread ```

WELCOME Happy Tuesday, legends. Welcome back to another edition of The Frontier — our weekly newsletter covering the best new AI launches on Product Hunt. WHAT'S HOT 🕵️ OpenAI's own model hacked Hugging Face — Guardrails off, testing a cyber benchmark, the model found a zero-day and broke in for real. Happened July 11. OpenAI confessed July 21. 💸 Nvidia weighs a $250B guarantee for OpenAI's Ohio campus — Backing a 10-gigawatt data center SoftBank is building on a former uranium site. Full project cost: north of $500B. 🌐 Kimi K3's weights are free. Running them isn't. — Moonshot AI released the largest open-weight model ever, 2.8 trillion parameters, permissively licensed. You still need a data center to actually run it. 💾 Your next laptop costs more because of AI — DDR5 prices are up over 100% this year as manufacturers redirect memory supply toward AI data centers. ⚡ AMD and Cerebras are splitting inference in two — AMD's Helios racks handle the prompt, Cerebras' Wafer-Scale Engine handles token generation. Claimed 5x throughput per watt. Five AI tools you may have missed Heard — Give Claude Code and Codex a voice 98 • Productivity ▲ 336 Grok 4.5 — SpaceXAI's model for coding, agentic tasks & knowledge work 5 • Android ▲ 178 Framer AI Agents — Design and publish professional sites with AI 28 • Design Tools Promoted ▲ 895 Health in ChatGPT — Your personal health companion, powered by ChatGPT 6 • Health & Fitness ▲ 150 Rivault — Approve AI agent data access with Face ID 29 • Productivity ▲ 110 MCP-Billing — OAuth 2.1 + usage-based Stripe billing for MCP servers 44 • API ▲ 125 Ad Your canvas just learned to design Framer 3.0 ships AI Agents that live inside your canvas and can actually design. Not suggest. Design. Drop a brief and watch them build out full pages, create components, write code, connect your CMS, and handle SEO. All without leaving the canvas. They plug straight into Claude Code and Cursor. Yes, inside your actual component system, without breaking it. The Figma file you were about to hand off to engineering? Keep it. Try the agents FROM THE FORUMS "Done" Is Not a Fact About Your Codebase terranicgerald (@terranicgerald) asked the forums for the smallest, dumbest thing that ever broke their trust in an AI agent mid-task, not the catastrophic failures, the small ones that quietly changed how people work. The replies converged on one shape: an agent reporting success with nothing behind it. Madison Brooks (@madison_brooks2) watched one silently delete a piece of logic while refactoring, everything still compiled, but an edge case vanished. Ella Reyes (@ella_reyes1) had one claim it "checked all relevant files" and skip the exact file she'd named in the prompt. Raj Nagulapalle (@rnagulapalle) stopped trusting any write operation that didn't hand back a receipt, after one came back clean with no error and simply hadn't happened. Asad M. (@asadmalik901) had the line that reframed the whole thread: "the trust did not actually break at the rename. What broke is the quiet assumption that 'done' is a claim about your codebase, when it is only ever a claim about the agent's own last step." Good thread if you've started reading an agent's "done" as something to verify instead of something to believe. Join the thread

WELCOME Happy Tuesday, legends. Welcome back to another edition of The Frontier — our weekly newsletter covering the best new AI launches on Product Hunt. Five AI tools you may have missed CreateOS Sandbox — Instant, hardware Isolated Sandboxes for AI agents 70 • SaaS ▲ 172 Jockey by TwelveLabs — The video AI agent that understands your whole library 28 • Productivity ▲ 193 Framer AI Agents — Design and publish professional sites with AI 28 • Design Tools Promoted ▲ 895 Diffsmith — Comment on your AI agent's code & collaborate on changes 25 • Developer Tools ▲ 97 Skim — Free, open-source AI email client for Windows 43 • Windows ▲ 146 Manifest — Turn any webpage into an action manifest for AI agents 83 • API ▲ 118 SPONSORED BY Type 4x Faster Are you really still typing? Full disclosure: Wispr Flow is the AI dictation tool most of us at Product Hunt use (we still have a few holdout typers, what romantics). Hold a key, talk, and clean text drops straight into whatever app you're already in — Slack, email, Notion, your IDE, wherever your cursor lives. No switching windows. No copy-paste ritual. Just say the thing – yes, you can whisper it – and even your most run-on sentences will be turned into polished writing at 4x the speed of typing. Start talking now WHAT'S HOT Kimi K3 Just Wiped $3.3 Trillion From Chip Stocks. Moonshot AI released Kimi K3 last Thursday. 2.8 trillion parameters, largest open-weight model ever announced. It scored above Claude Opus 4.8 and GPT-5.5 on benchmarks, on par with GPT-5.6 Sol. It also knocked $3.3 trillion off chip stocks, the second time a Chinese open-weight model has done that to markets in eighteen months. Investors are back to asking what $725 billion in AI infrastructure capex is actually buying. The weights aren't public yet. They drop July 27, six days from now. When they do, any developer with a GPU setup can run a model that currently benchmarks near the frontier for nothing. DeepSeek V4 is already at $0.87 per million output tokens. K3 costs $15 on the API right now. In six days both those numbers become the ceiling, not the floor. For people building products, the specific number that matters is K3 topping Arena's Frontend Code leaderboard. Code is the most commercially useful benchmark because it's what most products actually run on. If you're building something where the AI writes the code, you're about to have a free option that performs near the best paid one. The chip stock story and the open weights story are the same story. The argument for $725 billion in hardware spend was always that you need the chips to run the best models. K3 makes that harder to say. So does DeepSeek V4. So does whatever drops the week after. The commodity wall just moved again. See Kimi K3 on Product Hunt

WELCOME Happy Tuesday, legends. Welcome back to The Frontier — our weekly briefing on AI launches and news from Product Hunt. Today: OpenAI ships three launches in three days, a floating Claude that reads your screen, and $10k in credits for whatever you build next. Five AI tools you may have missed Claude Overlay — A floating Claude Code chat that sees your screen 41 • Developer Tools ▲ 164 Agentcard for companies — Give your agent a debit card 38 • Fintech ▲ 197 Framer AI Agents — Design and publish professional sites with AI 28 • Design Tools Promoted ▲ 895 Goose Ads Remixer — Remix the ads already winning in your niche 84 • Marketing ▲ 484 Sourclip — Turn NotebookLM into a complete research workflow 32 • Productivity ▲ 97 VocalVia — Turn documents and articles into editable multi-voice audio 42 • Artificial Intelligence ▲ 115 SPONSORED BY So we're just...talking to software now? ElevenLabs has been the go-to for voice for a while. Now they've turned that expertise into agents that actually get things done. You set one up, it talks like a real person, listens, responds, and helps handle the task — support calls, bookings, whatever the job is. Not a demo, not a "press 1 for sales" situation. It's ready to deploy. Feels like one of those shifts where the interface quietly changes. Less typing, less clicking, more just saying what needs to happen and letting it play out Try it today WHAT'S HOT OpenAI's biggest week yet OpenAI just shipped three launches in three days. GPT-Live landed Thursday, a full-duplex voice model that listens, speaks, and reasons at the same time, with live translation and mid-conversation web search. GPT-5.6 , their largest model yet, followed Friday, rolling out customer by customer after US government approval. ChatGPT Work , a desktop app aimed squarely at professional workflows, closed out the week. The one that matters is 5.6. The family ships in three sizes, Sol, Terra, and Luna, and the pitch is coding and agentic work first, chat second. Early reports have it going toe to toe with the best models available on exactly the tasks developers care about. Here's what that means on the ground. For the past year, Claude Code has been the default coding harness for professional builders. GPT-5.6 is the first model in a while that has a lot of us opening a different terminal: trying Codex for the first time, porting workflows over, comparing outputs side by side. If you built everything around one tool, that migration isn't painless. But the labs leapfrogging each other is exactly how everyone building on top ends up with better tools, whichever one you stick with. And if you're going to kick the tires anyway, might as well get paid for it. We teamed up with OpenAI: build something with Sol, Terra, or Luna, schedule your launch for Thursday, July 23, and the top products walk away with $10k in OpenAI credits. See the launch

WELCOME Happy Tuesday, legends. Welcome back to The Frontier — your daily briefing on AI launches and news from Product Hunt. Today: a 1.6-trillion-parameter model trained without NVIDIA, AI that reads your face during user research, agents in your Slack, and the first fully autonomous AI ransomware attack — which happened to coincide with Anthropic locking down Claude Code's default permissions. Probably not a coincidence. Five AI tools you may have missed Ellis — AI notes for in-person meetings, entirely on-device. No upload, no cloud sync, just a transcript and summary that stay on your machine. For the conversations you actually have in a room rather than on a call, which most note-taking apps have been quietly ignoring. Ogment AI — An AI coworker that lives in your Slack. Tag @O in any channel and it joins the conversation, answers questions, and runs tasks across your connected tools — already in the thread when context matters most, instead of a separate app you have to open and brief from scratch. Mira — AI-moderated user interviews that watch how respondents feel while they answer, using facial coding, voice emotion, and eye tracking to flag the moments when what someone says doesn't match their face. From Entropik, nine years and 17 patents in human behavior research. Glideo — Screen recordings that edit themselves. Record your demo or walkthrough and Glideo trims the pauses, adds captions, and removes filler — finished clip, no timeline. For demos, tutorials, and async updates where you want the result to look considered without spending an hour in an editor. Katalyst — An AI agent that works your Salesforce pipeline: updates deal stages, logs activity, drafts follow-ups, and flags deals that have gone quiet. The argument: sales reps spend more time updating CRM than selling. The agent handles the administrative layer so the human can do the part that still requires one. SPONSORED BY So we're just...talking to software now? ElevenLabs has been the go-to for voice for a while. Now they've turned that expertise into agents that actually get things done. You set one up, it talks like a real person, listens, responds, and helps handle the task — support calls, bookings, whatever the job is. Not a demo, not a "press 1 for sales" situation. It's ready to deploy. Feels like one of those shifts where the interface quietly changes. Less typing, less clicking, more just saying what needs to happen and letting it play out. Try it today WHAT'S HOT A food delivery company just challenged NVIDIA Meituan — the company that delivers dumplings and bubble tea to 700 million people across China — published LongCat-2.0 today: a 1.6-trillion-parameter open-weights model, MIT licensed, with a million-token context window. That's not unusual by itself. The part turning heads is the silicon: the entire training run happened on AI ASICs, no NVIDIA hardware anywhere in the stack, 35 trillion tokens from start to finish without a single rollback. That last detail matters more than the parameter count. Alternative hardware has been the theoretical answer to NVIDIA's dominance for years. In practice it tends to die late in a long training run — a loss spike hits, the custom chip can't recover cleanly, and the team switches back to H100s. Meituan ran through it. 35 trillion tokens is a serious training run, not a demo. The fact that it completed clean on ASICs is the proof-of-concept the hardware alternative camp has needed. There's an obvious question about why a food delivery company has this capability at all. The answer is that Meituan has been running large-scale ML for years — demand forecasting, routing, logistics optimization across hundreds of cities. The infrastructure to train at scale existed before the frontier model. LongCat wasn't a moonshot; it was a stress test on hardware they already owned. The valuation of every major AI lab has a quiet assumption embedded in it: serious models require NVIDIA, and NVIDIA's position is durable because the alternatives don't work at scale. LongCat-2.0 is one data point, not a trend. But every time a training run this size completes without H100s, that assumption gets a little harder to defend. The hardware monopoly isn't broken. It just got its first visible crack from a company that primarily sells noodles. See the launch

WELCOME Happy Tuesday, legends. Welcome back to another edition of The Frontier — our weekly newsletter covering the best new AI launches on Product Hunt. Five AI tools you may have missed Persona.js — The first AI chat library to natively support WebMCP, the new standard Chrome just shipped for letting agents use a website's own tools. Drop it into any existing site (WordPress, Shopify, a 2018 Next.js app) and an agent can act through your real UI instead of a clumsy headless browser. From Runtype, MIT-licensed. PMB — Local-first memory for coding agents, from solo engineer Oleksii Bondar. It keeps your decisions, rules, and dead-ends in one SQLite file on your disk and feeds them back to Claude Code, Cursor, or Codex before each session, so you stop re-explaining your project every morning. No cloud, no API keys. VisibAI — Tells you whether ChatGPT, Perplexity, Claude, and Gemini actually name your business when someone asks for a recommendation, then hands you a fix list (schema, robots.txt, FAQ files). Founder Francesco built it after watching buyers move from Googling to just asking an assistant. Akiflow — The task-and-calendar app added an MCP integration, so you can run your schedule straight from Claude, ChatGPT, or Cursor: ask them to block your week, add tasks, or review your day, and connect more than one calendar at last. Cursor for iOS — Cursor put its coding agents on your phone. Kick off a task on your Mac and keep steering it from the app while you're out, part of the steady drift of agents off the desktop and into your pocket. SPONSORED BY Are you really still typing? Full disclosure: Wispr Flow is the AI dictation tool most of us at Product Hunt use (we still have a few holdout typers, the romantics). Hold a key, talk, and clean text drops straight into whatever app you're already in — Slack, email, Notion, your IDE, wherever your cursor lives. No switching windows. No copy-paste ritual. Just say the thing (yes, you can whisper it) and even your most run-on sentences come out as polished writing at 4x the speed of typing. Try for free WHAT'S HOT Meta open-sourced an AI that types full sentences from your brain Meta's brain-and-AI research team this week released Brain2Qwerty v2 , a system that decodes whole sentences from non-invasive brain recordings, no implant and no surgery involved. It reads at 61% word accuracy on average and 78% for its best participant, where more than half of that person's sentences came back with one word wrong or fewer. Meta put the training code for both v1 and v2 on GitHub, along with the v1 dataset, so anyone can build on it. The reason researchers are paying attention is how it gets there. Earlier brain-to-text work leaned on hand-crafted pipelines; v2 goes end-to-end, learning straight from the raw brain signal, and the jump in accuracy is the payoff. The team also says the scaling laws look promising, the polite way of saying this gets better with more data and compute rather than topping out, which is the part that turns a research demo into a direction. There's a very large catch, and it's not the model. The signal comes from an MEG scanner, a room-sized machine that has to stay in the building and costs about as much as a house. So this is the opposite trade from Neuralink: nobody is drilling into your skull, but the price of skipping the surgery is hardware you can't take anywhere. Brain-typing you'll actually use is gated on that magnet shrinking, and that's a much slower curve than the software. Still, the milestone is real. Non-invasive brain-to-text has spent years stuck at "barely works," and 61% with an open codebase is the first time the no-surgery path looks like it's on a curve worth watching. The keyboard is safe for a long while yet. The direction of travel is the story. See the launch