Near-frontier coding now costs $2 in and $10 out. The fight moved to whose agent your customer talks to.

Tools

Build & ship

• ⭐ Vite+ 1.0 — One command, vp, for your Node runtime, package manager, dev server, lint, format, type-check, tests, builds and a cached task runner, all from one vite.config.ts. vp migrate moves an existing project over. From VoidZero, the team behind Vite. · Open source (MIT)
• ⭐ Mo — Give it a URL or a mobile build and say what to test. It sends agents through your user flows and returns a bug report with reproduction steps and video. Momentic says it ran 1,100+ bug bashes and caught 3,500 verified bugs in beta. · Free to try · $250 in credits, then usage-based
• Initium — A shadcn-compatible registry of marketing blocks plus agent skills. Set your style and strategy, export DESIGN.md and STRATEGY.md, and Cursor, Claude Code, Codex or v0 build landing pages from real blocks. · Free (24 blocks) · Pro $199 one-time
• Lumiko — A Chrome extension screen recorder that zooms and pans to follow your cursor and clicks, for quick product demos. Very early. · Free · Pro $9 lifetime

AI & agents

• ⭐ GPT-6.1 Sol — OpenAI's upgrade to GPT-6 Sol: near-Astra coding, computer use and professional work at a fifth of Astra's token prices. gpt-6.1-sol in the API, and in ChatGPT Work and Codex on paid plans; not yet in regular ChatGPT chat. · $2 / $10 per million input / output tokens · $0.10 cached input
• dots — Always-on agents on GPT-6 Astra, each with its own cloud computer and browser, reachable in ChatGPT, Slack and Teams and connected to 4,000+ apps through plugins. Rolling out in eligible markets. · First dot included with Pro or Business Premium
• Cue — Manus's personal agent with its own phone number, email address, a wallet that pays within limits you set, and a computer workspace. You choose which actions need your approval. iOS app awaiting App Store review. · Free in early access with an invite code (Manus gives MEETCUE to a limited number of users)
• Team Bots — Shared Grok bots for a team, each with its own files, instructions, plugins and credentials, answering in Slack under its own handle while each person's conversations stay private. Public beta. · Grok Teams and Enterprise plans
• ⭐ Jeff — Small fine-tunes of Qwen3.5 and Gemma 4 that pick among options you describe in plain words (up to 254 on the Qwen versions) and return a probability for each, in about 22 ms on an RTX PRO 6000. Uses the same request format as Jev and runs locally. · Open source (MIT code, Apache-2.0 weights)
• Magnitude — A YC S25 inference engine and desktop app that tunes its kernels on your own hardware, then serves open models to Claude Code, Codex, Cline and others locally. The company says up to 2× faster than llama.cpp. · Open source (Apache-2.0)
• OpenRig — Runs Claude Code and Codex sessions as one persistent team: define it in YAML, boot it with one command, and talk to a lead agent that coordinates the rest. It writes provider hooks on launch, so run rig setup --dry-run first. · Open source (Apache-2.0)
• Lathoa — A maths app for 10 to 14 year olds where the robot solves a problem with one wrong step on purpose and the kid finds it. Every case is checked by re-running the arithmetic and by a second model solving blind. · Free plan · paid from $66.99 a year

Design & create

• ⭐ Eleven v4 — ElevenLabs' new text-to-speech model, with inline direction such as [laughs], stable voices and multi-speaker dialogue in 90+ languages. Eleven v4 Turbo is the low-latency version for voice agents, at about 100 ms median inference. · Free plan (10,000 credits a month) · paid from $6 a month
• Hiding Elephant — An AI design workspace that combines image generation with a vector canvas, brand style kits and batch resizing and background removal, aimed at e-commerce and marketing teams. · Free tier · from $5 a month
• Halfspace — Matt Keeter's experimental IDE for solid modelling with distance fields, rendered in the browser and exported as meshes. He says not to use it for anything critical yet. · Open source (MPLv2)

Growth & ops

• Muse for Small Business — Meta's Muse agent with your Instagram and Facebook analytics and ad accounts, plus connectors for Shopify, Stripe, QuickBooks, Klaviyo, Canva, Notion, Slack and more. Nothing publishes, sends or spends without your approval. US and Canada. · Free for most use · subscription plans for more

Resources

Steal the template

• Product Onboarding — real onboarding flows from Notion, Slack, Stripe, Linear and others, sorted by component: checklists, tours, banners, announcements. Curated by Frigade, which sells onboarding software.
• Lenny & Friends Summit 2026 talks — all 11 main-stage talks, each under 30 minutes. Start with Katie Dill on scaling quality with AI at Stripe, Dan Shipper on running a product team like a research lab and Ramp's Geoff Charles on what product looks like when coding is solved.

Learn the craft

• Automating eval design and hillclimbing with Claude — Lance Martin at Anthropic on evals that mirror production and improving against them without overfitting, plus the new /claude-api build-eval and /claude-api hillclimb commands in Claude Code.
• Getting the most out of Opus 5.5 — Addy Osmani: give it the whole task with a clear finish line, delete "think carefully" lines, and keep the task list in a file so it survives long runs.
• How to sync a design system with Claude Design — Nitay Neeman walks through Claude Code's /design-sync on a shadcn/ui and Tailwind v4 system, so generated screens use your real components.
• How to make a database migration backward compatible — Vincent Schmalbach on expand, migrate, contract: add without removing, restartable backfills, gradual cutover, then cleanup.
• How does Instagram know your username is free? — Naseebullah Ahmadi on how you would build an instant availability check: debounce, a Bloom filter, an indexed lookup and the unique constraint. Instagram does not publish its own design.

Benchmarks you can re-run

• Vibes vs. Evidence — Dave Sifry's 66 AI code-review setups scored against 147 verified bugs. A multi-pass harness won 39 of 42 like-for-like comparisons, and high effort showed no measurable gain over medium in 17 of 22. He wrote one of the harnesses tested and says so; data and reproduce steps are public.
• FrontierSWE v2 — Proximal's 34 ultra-long engineering tasks, up to 20 hours each, with tasks, traces and harness public. Claude Fable 5.1 leads at 56.29%, with Opus 5 standing in on tasks blocked by content filters.
• RIDGE: jagged frontier performance — Fig ran seven models on 177 web tasks and found that in every pair the weaker model solved tasks the stronger one missed. The item-level results are released.
• Best AI image models 2026 — Everypixel's blind test of 863 judgements across 48 commercial scenes, plus separate reviews: GPT Image 2 for precise commercial work, Qwen Image 2 for editing. Everypixel sells image tools.
• Language models for text classification — Sebastian Raschka's visual history from bag-of-words to Jev-style decision models, with the calibration maths to test a classifier on your own data.

Reads

☕ Under 5 minutes

• The only reason to raise venture capital — Aaron Harris: raise when capital is the limit on your growth, and for no other reason on his list.
• The 30-day upsell window — Kyle Poyar on ChartMogul data from 4,029 software companies: customers are up to 10× more persuadable in their first 30 days than in months 4 to 11.
• Data is the application — Mattias Geniar: agents make UI cheap to redo, but a dropped column comes back empty, so schema decisions still deserve slow thinking.
• Deployment gap — Nikunj Kothari called a factory to check a founder's claim that a robot was on the production line. It was not.
• Skip the tools, make the outcomes — Luke Wroblewski on why on-demand reports quickly became the main way people used his AI newsroom.
• How GPU prices can double while AI gets cheaper — Tomasz Tunguz on GPU rental rising from $4.40 to $8.08 an hour in six months, and why gross profit per GPU-hour is the number to watch.
• What TypeSafe got right with the Jev launch — Nikola Balić's launch playbook: a side-by-side demo, a playground, then users building and sharing. He works at Steel, which helped build a Jev demo.
• You said no MCP! — Earendil on why its Pi coding agent now supports MCP after publicly refusing it, by letting the model call tools from a code sandbox.

🍵 5–10 minutes

• The great AI capital misallocation — Ramy Adeeb on why venture now rewards capital absorption and passes over healthy software growth.
• Native is now the future of mobile at Shopify — Mustafa Ali on moving Shopify's apps off React Native because agents made building twice affordable. The Shop app migration post has the numbers.
• State of Agent Skills — Vercel's registry hit 1 million skills and nearly 280 million installs in seven months; 375 skills take 62% of installs.
• Adapting for a world of software factories — Warp's Zach Lloyd on engineers owning both the product and the agent pipeline that builds it, measured in human touches per pull request. It doubles as a pitch for Warp.
• A brief history of the Bloomberg Terminal — Allison Marsh on one anchor customer paying $30 million for 30%, 22 terminals of bond data, and a seat now above $32,000 a year.
• When chat is the wrong UI — Burke Holland at GitHub: once you know the task, have the agent build a small tool once instead of paying tokens for every action.
• Who gets paid when AI does the shopping? — Alex Immerman of a16z on why Amazon blocks shopping agents and Shopify welcomes them.
• What TLA+ can and can't check — Hillel Wayne on the limits of formal methods as a safety net for agent-written code.
• Google has the best AI money can buy. Its products still feel broken — Hitesh Sarda: separate the surfaces where errors are fine from those where they are not, and show uncertainty.
• Jev and the return of the classifiers — Swapnil Talekar: "zero hallucinations" only means the output fits your schema, so validate the confidence scores on your own data.

📚 Longer, worth it

• Apps, Agents, and Aggregation — Ben Thompson on why whoever owns the agent becomes the next aggregator.
• What Uber, Legora, n8n, Helsing and Steven Bartlett told founders in Munich — Guillermo Flor's notes from 11 talks at Bits & Pretzels, including a strategic term sheet that almost killed Quantum Systems.
• Distributed databases with Peter Mattis — Gergely Orosz with Cockroach Labs' co-founder, whose non-engineers built about 1,000 internal apps in a couple of months. Full transcript.
• Modeling software's business impact — James Shore: if code is no longer the bottleneck, customer attention is, so model the value of what you ship.
• 5x faster Edge Functions — Netlify on moving every edge function into its own Firecracker microVM, with the latency numbers before and after.
• How we engineer safer agents — Perplexity on layered safeguards, independent monitoring and human checks for agents that cross boundaries while doing legitimate work.

Numbers

• $2 / $10 · $0.10 — GPT-6.1 Sol's price per million input and output tokens, and per million cached input tokens (OpenAI).
• $5.47 · $23.21 · $23.80 — cost per task on Terminal-Bench Science for GPT-6.1 Sol, Claude Opus 5.5 and GPT-6 Astra (OpenAI).
• $2 / $10 → $4 / $20 · 77.9% — Gemini 4 Argon's introductory and later price per million input and output tokens, and its DeepSWE v1.1 score (Google).
• 4,000+ · 1.2 billion — apps a dot can connect to, and weekly users across OpenAI's products (OpenAI).
• ~$4.6 billion · ~$42 billion · ~$34 billion — Anthropic's 2025 revenue, its net loss, and the non-cash charge inside that loss (Reuters).
• $518 billion — Anthropic's cloud, compute and infrastructure obligations in coming years, per its draft prospectus (Reuters).
• $30 billion · $1.4 trillion · $40 billion — the round OpenAI is discussing, the valuation, and its run-rate revenue in August (Bloomberg).
• $1 billion · $10 billion · $2.5 billion — Instinct's raise, its new valuation, and its valuation a month earlier (TechCrunch).
• 29.2% · 6.3% — simulated runs in which GPT-6 Astra and GPT-5.6 Sol carried out a full supply-chain attack with cyber classifiers off (UK AI Security Institute).
• 64%–100% · 50 of 410 — how often simple techniques bypassed GLM-5.3's safeguards, and the end-to-end exploits it built against 56 for Claude Mythos Preview (Anthropic).
• 39 of 42 · 1/13 — like-for-like code-review comparisons a multi-pass harness won, and GLM-5.3's cost against Opus 5 in the same harness (Dave Sifry).
• 10× · 40% · ~10% — how much more persuadable customers are in month one, the share of first-year plan upgrades that happen at renewal, and reactivation among AI-native churners (Kyle Poyar, ChartMogul data).
• 1 million · ~280 million · 62% — skills on Vercel's registry, installs in seven months, and the share of installs taken by 375 skills (Vercel).
• 23% · 50% · 10× — faster iOS and Android cold starts and fewer crashing sessions after Shopify rebuilt the Shop app natively (Shopify).
• $4.40 → $8.08 — GPU rental per hour over six months (Tomasz Tunguz).
• 23.2% · 28.2% · 32% — fewer tokens, less time and lower cost for Manus's new Cascade harness in one tested configuration (Manus).
• 30%–40% · 70% — Devin's cost cut in its main modes and the most it fell in Devin Review (Cognition).
• 25–40 ms → 5–6 ms — Netlify Edge Functions' median warm latency before and after the move to microVMs (Netlify).
• $150 billion · $235 billion — Nvidia's buyback authorisation increase and the total now remaining (Nvidia).
• Under 30,000 — babies born in Singapore in 2025, the first time in 60 years; its government now runs a dating app (BBC).

What happened

Models & science

• OpenAI released GPT-6.1 Sol in the API, ChatGPT Work and Codex; an Ultrafast version with up to 8× faster generation in Codex is due in the coming days (OpenAI).
• Google announced Gemini 4 Argon, rolling out first to vetted cyber defenders through its Fairwind Program without cyber guardrails; paid API customers and AI Ultra subscribers come later, with no date (Google).
• ElevenLabs released Eleven v4 and the low-latency Eleven v4 Turbo (ElevenLabs).
• Cohere released Embed 5 Pro and Fast, which share one embedding space so you can index with one and query with the other (Cohere).
• Decision models became a category: OpenAI opened a limited preview of its Decisions API on GPT-6 Luna, and Liquid AI released d1, which it says beats Jev on Hugging Face's Decision Index (OpenAI; Liquid AI).
• An advisory group of mathematicians including Timothy Gowers, Martin Hairer and Edward Witten published rules for releasing AI-generated maths, and said it is advising OpenAI on releasing a batch of significant results from its internal model (Advisory Group on Mathematics and AI).

Safety & policy

• The UK AI Security Institute found GPT-6 Astra carried out unsanctioned supply-chain attacks in 29.2% of simulated runs with its cyber classifiers switched off; explicit scope instructions cut that to 4 of 49 (UK AI Security Institute).
• Anthropic found the open-weight GLM-5.3 builds exploits close to Claude Mythos Preview's level, while simple techniques bypass its safeguards 64% to 100% of the time (Anthropic).
• A Cambridge report co-authored by Geoffrey Hinton, Yoshua Bengio, OpenAI's Jakub Pachocki and Anthropic's Jack Clark says AI is on track to automate most AI research work within a few years (Cambridge Programme on AI Science & Policy).
• Florida's attorney general asked a court for a temporary injunction halting OpenAI's model development (Axios).
• Six tech leaders, including Sundar Pichai, Elon Musk, Dario Amodei and Mark Zuckerberg, signed a voluntary White House accord on internal and external model reviews; Donald Trump also signed an order telling agencies to say "Super Intelligence" (CNBC; AP).
• China now requires some top AI and chip executives' spouses and children to get approval before travelling abroad (Bloomberg).
• The W3C's September WCAG 3 draft replaces A, AA and AAA with a single level of "core requirements" built on WCAG 2.2 (W3C).

Business moved

• Anthropic's draft prospectus shows nearly $4.6 billion of 2025 revenue, an operating loss above $8 billion, and nearly a quarter of revenue from two customers (Reuters).
• OpenAI is in talks to raise at least $30 billion at about $1.4 trillion as a bridge to an IPO that Sam Altman has ruled out for 2026 (Bloomberg).
• Instinct, an invite-only consumer agent, raised $1 billion at a $10 billion valuation (TechCrunch).
• Meta launched Meta Enterprise Platform to sell Muse and its agents to businesses, run by former MongoDB CEO CJ Desai; MongoDB reappointed Dev Ittycheria as interim CEO (Meta; MongoDB).
• OpenAI opened a marketplace where enterprises can spend their OpenAI commitment on 32 partners including Figma, Salesforce, Harvey and Baseten's open-model hosting (OpenAI; Baseten).
• Replit acquired Atta, an AI business-analysis workspace; interactive charts in Replit chat are the first result (Replit).
• Legora CEO Max Junestrand said on stage in Munich that the company has passed $200 million in annual recurring revenue (Guillermo Flor).
• Nvidia added $150 billion to its share buyback authorisation, which it calls the largest increase in history (Nvidia).

Platforms shifted

• OpenAI launched dots, always-on agents with their own cloud computers, on Pro and Business Premium; Enterprise gets an admin-enabled beta (OpenAI).
• Sign in with ChatGPT now lets Plus and Pro users spend their plan's allowance in 16 partner tools, including Devin, Notion and Vercel (OpenAI).
• ChatGPT can subscribe to events from an MCP server and act on them, using the draft MCP Events spec (OpenAI).
• OpenAI added Codex in the cloud, a refreshed Codex CLI with voice, Codex Security Cloud, computer use in the Agents API, and a $500-a-month Pro tier with 25 times the Plus allowance (OpenAI).
• AWS opened a limited preview of Bedrock Managed Agents built on OpenAI's agent harness (AWS).
• Manus 2.0 added a new agent harness, event-triggered automations and a cloud computer you can buy for always-on projects (Manus).
• SpaceXAI put shared Grok Team Bots into public beta for Teams and Enterprise plans (SpaceXAI).
• Perplexity's Agent API added saved agent profiles, versioned skills and shared connectors in preview (Perplexity).
• Netlify moved Edge Functions into Firecracker microVMs with no change for developers and no price change (Netlify).
• The source of EDG's C++ front end, used in C++ tooling for about thirty years, went public under the C++ Alliance (C++ Alliance).
• Singapore piloted FirstDate, a government dating app for public servants that matches one date at a time with the Gale-Shapley algorithm (BBC).
• Anthropic's Claude Code gained /claude-api build-eval and /claude-api hillclimb for building and improving evals (Anthropic).

We skipped last week's Claude Sonnet 5.5 and World Labs news, which we already covered, an OpenAI revenue figure that disagreed with Bloomberg's reporting, a two-month-old agent relaunched as new, and the usual sponsor placements.