Airbnb matched frontier models with a 2B model at 1% of the cost. Above 30 requests a minute, the big API is the expensive choice.

Today's haul: 15 tools · 9 resources · 22 reads · 12 numbers · 26 things that happened. Every tool, resource and read below has a working link. No link, no listing.

The 60-second catch-up

For narrow, repeated jobs, a small model you tune yourself now beats renting the frontier.

Airbnb fine-tuned a model with 2 billion effective parameters for a production policy classifier. It lifted F1 from 0.563 to 0.902, ran at a 35 ms median against about 900 ms for frontier APIs, and cost roughly 1% as much to serve. The catch is traffic: below roughly 15 to 30 requests a minute, Claude Sonnet 5.5 and GPT-6.1 Sol were cheaper than a dedicated GPU (Airbnb). Hightouch found a middle path for identity resolution. A cheap model that passed only its uncertain cases to Opus, 22% of them, scored 98.8% at $0.88 per thousand, against Opus alone at 99.0% and $3.86. On a harder test the saving shrank to under 2×: $2.94 against $5.39 (Hightouch). Hosted review is falling in price too. Cognition says a typical Devin review now costs $0.25 to $0.75, down 82% per pull request since August (Cognition).

→ do this: count the requests per minute on your busiest single-purpose prompt. Above about 30, price a fine-tuned small model on one GPU. Below that, route only the hard cases to the big model and let a cheap one take the rest.

The best new hacking model mostly stays inside the job it was given. Patching is still the slow part.

OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and OpenAI now rates it "Critical" for cybersecurity, up from High. In a honeypot test with one approved target among forbidden ones, it went out of scope in 0.07% of runs, against 5.51% for its predecessor. Access is limited to approved Daybreak Red customers (OpenAI). Defenders are using the same models: Mozilla says early frontier access has helped Firefox ship more than seven times as many security fixes as in the same period of 2025, and it now ships weekly builds to close the gap between a public fix and an installed one (Mozilla). This week's bugs show how old the holes can be. OpenSSH on musl systems without PAM has accepted any 257 to 1,024-character password for locked accounts since 2016, now fixed in OpenSSH 10.6p2 (sshhole).

→ do this: if you run Alpine or another musl-based image with SSH password login on, upgrade OpenSSH today. Then give every agent you run a scope test: one allowed target, one forbidden one, and a log of which it touched.

Agents are becoming customers with their own permissions, and the products that serve them are being rebuilt.

Mercury now lets an agent prepare a payment or transfer but never send one: a person taps to approve, and an agent's permissions can never exceed its owner's role (Mercury). GitHub rebuilt its Git storage so that one repository could take 5,000 concurrent agents and 7,200 pushes a second in internal tests (GitHub). Kenneth Auchenberg argues that agent experience is not a developer-tools niche. Every product is becoming something agents use (Innovation Endeavors). Anita Kirkovska finds growth teams replacing SaaS tools with agents that have "the right data, instructions and access", one bottleneck at a time (Anita Kirkovska).

→ do this: list the actions an agent could take in your product. Split them into "prepare" and "commit", and require a human tap for every commit that moves money or deletes data.

Tools

Build & ship

⭐ Cursor live apps — Describe the tool you want and Cursor builds it, sets up a database, runs deterministic and agentic security checks, deploys it and gives you a link to share. · All paid Cursor plans
⭐ devvm — One isolated VM per project, with the project's toolchain ready, for you or a coding agent. Limits the damage from supply-chain attacks, leaked secrets and rogue agents. macOS and Linux. · Open source (MIT)
libghostty-vt 1.0 — The terminal-emulation core of Ghostty as a portable C library, now with a stable binary interface. JetBrains' terminal already runs on it. · Open source (MIT)
File Viewer — Previews 275 file extensions in the browser, including Office, PDF, CAD, 3D and archives, with no conversion server. Works with Vue, React, Svelte and Web Components. · Open source (Apache-2.0; the DWG CAD module is AGPL-3.0)
GG Coder — A desktop coding agent that runs on your existing AI subscriptions, API keys or local models, with local search of open-source code. Its speed claims come from the author's own benchmark. · Open source (AGPL-3.0)

AI & agents

⭐ Gauge — Runs real coding agents (Claude Code, Codex, Pi, OpenCode) against your Skill, CLI, SDK, MCP server or docs, shows where they get stuck, and reruns the same task after each fix. The local loop runs in Docker. · Open source (MIT) · agent usage billed by your provider
Devin Review 2.0 — Cognition's PR reviewer, now with incremental reviews, an "Ultra" effort level, custom reviewers in beta and a per-PR spend limit. · About $0.25–$0.75 per typical review
Gemini Omni 1.2 Flash — Google's video model, now with native 15-second clips and conversational editing. Generally available in the Gemini API. · $0.07 per second of 720p video (30% less than 1.1)
Personal Finance in Claude — Connects bank and card accounts through Plaid, read-only, so you can ask about up to two years of transactions. No ads or sponsored recommendations. · Beta on Claude Pro, Max and Ultra (US)
Kapwing MCP — Lets Claude, ChatGPT and other agents upload footage, add or translate subtitles, dub and export videos in your Kapwing workspace. · Included in paid Kapwing plans · AI features use credits
NVIDIA Cosmos 3 — Four open models for interactive world simulation of driving scenes and robot actions, with 16B trainable parameters each, according to NVIDIA. · Open weights (OpenMDW 1.1)

Design & create

⭐ Slab — 60 headless Web Components that handle keyboard, focus, ARIA, forms and events on top of native elements, and ship no visual CSS. Works with React, Vue, Svelte, Rails or plain HTML. · Open source (Apache-2.0)
beUI — Copy-paste animated components for React, Next.js and React Native, built with Motion and Tailwind and installed through shadcn. · Open source (MIT) · paid Pro tier
CSS Type Studio — Iterate on typefaces, sizes, spacing and OpenType features in the browser, then copy the CSS. · Free account saves 3 designs · Pro for 1,900+ typefaces
sidebearing-trim — Aligns text columns by where the letters' ink starts rather than by the glyph box, using OpenType metrics. Vanilla JS and React. · Open source (MIT, per npm)

Resources

Steal the template

How to book B2B demos from Meta ads: Alex Vacca's playbook from Frontal, which sends CRM sales outcomes back to Meta to train its targeting and books qualified demos at $600 each.
Trade the discount for the logo: when enterprise prospects asked Chamelio for a discount, it asked for logo rights in return. Never give a discount without getting something back.
The swap-the-logo test: put a competitor's name on your ads, landing page and cold emails. If they could post it unchanged, don't post it.

Learn the craft

The Ultimate MoE Training Handbook: Hugging Face's long guide to training mixture-of-experts models, with code and its own cluster runs. 18 of the 20 highest-scoring open-weight models are MoEs.
Agents aren't humans (but they kinda are): Cloudflare on why sites can't tell who is behind an agent, and the mechanisms needed before sites give agents more access.
Introduction to compute markets: the primary market: Ornn on how GPU owners sell capacity through reservations, on-demand rentals and spot contracts, and why the reservation is what gets the loan approved.

Benchmarks you can re-run

Frontier quality at 1% of the cost: Airbnb's method for testing a fine-tuned small model against frontier APIs on a real classifier, including the traffic level where self-hosting starts to win.
Design Systems Should Be Fast by Default: Den Odell times a 249-country dropdown and 100 checkboxes in ten design systems, and shows the Playwright tests that fail when a component gets too slow.
The benchmark where AIs have to lie to win: Dwarkesh Patel pits models against each other in the deception game Blood on the Clocktower across 386 games, a test of bluffing and coordination no other benchmark covers.

Reads

☕ Under 5 minutes

The "super consumer": Atlas members spend about $1 million a year on cards, so its founder started treating them like enterprise accounts. A few of his calls ended in six-figure trips.
Some things I learned working at Microsoft: Slava Akhmechet on scale, distribution, and how leadership tied every engineer's pay to security after the Midnight Blizzard hacks.
Discord is ditching Electron: Discord is moving its apps to shared Rust code. A preview Windows client used 290 MB of RAM. There is no release date yet.
I wish tech companies knew more about me: years of tracking, and the ads still show the gamepad he just bought.

🍵 5–10 minutes

The very long road to $20k MRR: Andris Reinman earned €750 in 18 months from optional licences for EmailEngine, then $2,000 in the first month he required a subscription.
The Random Cage Effect: baseball players who practised against pitches in random order improved 57%, against 25% for those who practised in a fixed order. Practise the way the real thing will arrive.
How Figma manages sandboxes at scale: warm pools of pre-started sandboxes cut provisioning latency by 90% and lifted availability from 99% to 99.95%.
Win the demo, lose the market: Proterra delivered more than 1,000 electric buses and grew revenue 24% in 2022, then went bankrupt. Better technology was not an advantage on its own.
A WebGPU game shipped in three days on Expo: a puzzle game drawn entirely in code went from first commit to App Store-ready in three days, with an AI agent doing the QA.
How NVIDIA uses AI agents for cluster operations: specialist agents for jobs, hardware, Kubernetes and observability, all behind one chat and run on-premises with local models.
A Jev that can ask Opus for help: Hightouch's hybrid judge, with the accuracy and cost per thousand for each setup.
Don't build the reusable component yet: wait until a pattern repeats three times for the same reason before you extract it.
How a split webfont slowed down text layout: when one word needs two font files, the browser lays it out in more than one pass. Keep basic Latin in one file.
I toured Unitree's factory: Josef Chen finds Unitree's robots still mostly built by hand, because labour is only about 8% of their cost.

📚 Longer, worth it

Growth is becoming an engineering problem: Anita Kirkovska on what AI-heavy growth teams are building, buying and replacing in their go-to-market stack.
AX is not the new DX, it's the new UX: Kenneth Auchenberg on why coding agents were only the first users of agent-ready products.
Is private equity cooked?: Jen Kha on why software buyouts now need twice the earnings growth they did in 2015, just as AI makes renewals harder to count on.
React.lazy, Suspense and the hidden 300 ms delay: the code loaded in about 100 ms. React's reveal throttle held it back, and a smaller bundle made no difference.
Inside GitHub's new Git storage architecture: how Rubicon splits Git objects from reference updates so storage and compute scale separately.
The Daytona playbook: Guillermo Flor on how Daytona, which gives every agent its own computer, recovered from a failed Series A and a few months of runway. The full piece is paid.
AI can generate the interface. What makes it worth caring about?: the same confetti burst fits someone launching their business and annoys someone finishing a routine task. Context decides.
Fortune telling and speculative design: Maggie Appleton on how GitHub Next collects weak signals and builds prototypes of possible futures.

Numbers

0.902: F1 score of Airbnb's fine-tuned 2B-parameter classifier, up from 0.563, at roughly 1% of the cost of frontier APIs (Airbnb).
15–30: requests per minute below which hosted frontier models were cheaper for Airbnb than a dedicated GPU (Airbnb).
48.5%: GPT-6 Cyber's score on ExploitGym, up from 32.1% for GPT-5.6 Cyber (OpenAI).
0.07%: share of honeypot runs where GPT-6 Cyber attacked a forbidden system, down from 5.51% (OpenAI).
7×: security fixes Firefox has shipped compared with the same period of 2025 (Mozilla).
82%: drop in Devin Review's cost per pull request since August (Cognition).
4.4%: fall in the S&P 500 on Thursday, its worst day since April 2025, led by AI data-center stocks (AP).
~2×: rate at which Anthropic is improving its models on well-specified tasks, compared with March 2024 to October 2025 (METR).
7,200: pushes per second to one repository in internal tests of GitHub's new storage, with 5,000 concurrent agents (GitHub).
$600: what Frontal pays per qualified B2B demo booked from Meta ads (Alex Vacca).
10–12%: yearly EBITDA growth a typical buyout now needs to return 2.5× in five years, against about 5% in 2015 (Jen Kha, a16z).
48%: drop in US Cybertruck sales last year, to 20,237, as Tesla plans to wind down production around mid-2027 (NBC News, Reuters).

What happened

Models & science

OpenAI released GPT-6 Cyber, a security model for finding and validating exploits, to approved Daybreak Red customers through the API and Codex (OpenAI).
Google made Gemini Omni 1.2 Flash generally available, with 15-second clips and a 30% lower price (Google).
NVIDIA released four open Cosmos 3 models for interactive driving and robot simulation (NVIDIA).
Meta added real-time voice conversations and reacting avatars to its Muse app in the US (Meta).
Discovered Materials, a YC startup, put up $1 million in prizes for five materials-science problems that frontier agents made no headway on (TechCrunch).
Mitchell Hashimoto shipped libghostty-vt 1.0, with a stable C interface for terminal emulation (Mitchell Hashimoto).

Safety & policy

METR, given internal access, estimated Anthropic is improving its models about twice as fast as before. Anthropic redacted most of the evidence METR found most informative (METR).
An OpenSSH bug on musl systems without PAM let locked accounts accept any long password. It was fixed in OpenSSH 10.6p2 (sshhole).
Researchers published a guest-to-host escape from a Firecracker microVM, gaining root on the host (PWN Research).
The FAA said it will replace the floppy-disk systems at more than 400 air-traffic-control facilities by the end of 2028 (Reuters, Engadget).
The UK will let registered design disputes use a small-claims track capped at £10,000 and give the IPO more power to reject copied registrations (Design Week).
The RealPage rent-pricing class action closed with nearly $360 million to renters (Multifamily Dive).
Meta president Dina Powell McCormick said election backlash will make data centers harder to build (TechCrunch).

Business moved

AI infrastructure stocks led the S&P 500 down 4.4% and the Nasdaq down 5.5%. Nvidia fell 9.3% (AP).
Greg Brockman said OpenAI could be ready for an IPO in 2027, after enterprise became most of its revenue (TechCrunch).
Google bought Crusoe's data-center site in southeast Wyoming (Crusoe).
EXL acquired Aimpoint Digital, adding about 250 AI engineers and consultants. Terms were not disclosed (EXL).
Airbnb said it will launch its own AI travel agent in the first half of 2027 (TechCrunch).
Aina opened preorders for a $399 smart ring with a screen, joystick and mic for controlling AI agents. It ships in January 2027 (TechCrunch).
Wood Mackenzie expects sodium-ion cells to reach cost parity with LFP by late 2026, with more than 70 GWh already under supply agreements (pv magazine).

Platforms shifted

Mercury added agent controls: agents can prepare payments but never send them, and removing an employee disconnects their agents (Mercury).
Vercel put Schedules into public beta, so apps and agents can create one-time and recurring jobs from code (Vercel).
Apple launched retention messages for the subscription cancellation screen. API access is by request (Apple).
Cursor now builds and deploys small live apps you can share with a link, on all paid plans (Cursor).
Discord is moving its desktop and mobile apps off Electron and React Native to shared Rust code (The Verge).
ElevenLabs users can now publish AI-generated tracks straight to SoundCloud (TechCrunch).

We left out the week's viral app videos, a Hacker News run of history and travel essays, and the usual sponsor placements.

Keep reading

The rest of this issue: Airbnb matched frontier models with a 2B model at 1% of the cost. Above 30 requests a minute, the big API is the expensive choice.

The tools, resources, reads and numbers are behind this point.

It costs an email address. The same address goes on Makersfuel, the MakersClaw briefing that runs Tuesday to Saturday, and every issue carries an unsubscribe link.

The address is used for the newsletter and nothing else. Privacy.