Good morning. Here's what's worth acting on today.
BREAKING STORY
Gemini 4 Hydrogen: Google's cheaper, faster model, which it says edges out its flagship
Google released Gemini 4 Hydrogen on 12 November as its everyday model for agents (AI that carries out multi-step tasks on its own). It costs $1.50 per million input tokens and $7.50 per million output tokens as of 13 Nov 2026, the same rate across its whole 1-million-token context window (how much text it can read at once). It is in AI Studio, the Gemini API and the free Gemini app, where the automatic model picker chooses it when it fits. On Google's own benchmarks it scores 51 on the Artificial Analysis Intelligence Index, against 50 for Argon, Google's top model, and finishes tasks 2.5 times faster than Gemini 3.8 Flash.
My take: 3.8 Flash's introductory price of $0.75/$3.75 ends on 31 December and rises to Hydrogen's rate, so if you run on Flash you will pay Hydrogen's price either way. Test Hydrogen now at low or medium thinking (how much it reasons before answering). Artificial Analysis estimates, as reported by VentureBeat, that those cost $0.29 and $0.80 per benchmark task, against $1.87 at the default high. Run ten real tasks at medium, compare the output with what you get today, and use high only where you can see a difference.
MARKET TRENDS
78% of creators who use AI say it sped up their growth
In Adobe's Creators' Toolkit Report, an Adobe-commissioned survey of more than 2,000 US creators run with The Harris Poll, 78% of creators who use creative AI said it had accelerated the growth of their business or following. The more useful figure is the gap: 34% of full-time creators turned down a paid video job in the past three months because they lacked the skills to deliver it.
How to make money from it: Sell those creators the video work they turn away, at a fixed price. For example: clean up one recording (cut pauses, fix audio, add captions) and turn it into three short clips within 48 hours. Use AI editing tools to keep your time per job low, and price per video. To find buyers, look for creators in your niche posting long videos with no clips, and send each one a finished clip cut from their own footage.
GROWTH HACK / DIY
The need-shift upsell: sell when your user's situation changes
The idea: Most products ask for money on a calendar, such as a trial that ends on day seven. A need-shift upsell asks when the user's situation changes and they need something the free product doesn't do. Flo, the women's health app, says 40% of its subscriptions come more than six months after install, driven by new value for stages such as pregnancy or perimenopause. Its first perimenopause onboarding step cut conversion by almost 10%, because about 75% of users couldn't be placed in any phase; a rewrite won by 23%.
Do it today:
Map two or three moments that create a new problem, such as a first subcontractor or crossing 20 invoices a month. Log each as an event; where behaviour can't show it, ask one question and let anyone who doesn't fit carry on.
Treat each moment as a second first day: a few setup screens with a free part that is useful on its own, then one offer, one price and one button where the free version runs out.
Hold out a random 10 to 20% of users who hit the event, compare conversion and 60-day retention after four to six weeks, and confirm a winner with a removal test.
What you'll need: event tracking (as of 13 Nov 2026, PostHog's first 1M events a month are free), an in-app message slot, billing that handles mid-cycle upgrades (Stripe Billing pay-as-you-go costs 0.7% of volume), two days to build and four to six weeks for a result.
AGENT LAB
One AGENTS.md pointer got coding agents to read the docs 30 times out of 30
Coding agents such as Claude Code, Cursor or Codex rarely read the docs of the products they build on. Mintlify tested a fix on 10 November. A command that prints task guidance did nothing by itself: Claude Sonnet passed 16% of runs with it available and 20% without. A short pointer at the top of AGENTS.md (the Markdown file many coding agents read first) made agents run the command before their first edit in 30 of 30 tests.
How it works: user asks for a task → the agent reads the pointer, which names the task types (such as custom domains, publishing, access, and automations including CI) → runs the instructions command → reads the matching topic (common mistakes, steps, checks) → edits. Wording mattered: a generic pointer passed 66% of Mintlify's own tests, naming the task areas 83%, and rewriting around the mistakes agents made 100%.
Try it: turn your last 20 support messages into five to eight topic pages ("use when", "common mistakes", "steps", "done when") and add the pointer, plus a CLAUDE.md containing @AGENTS.md, since Claude Code reads CLAUDE.md. Run three real tasks five times without the pointer and five with it; keep it if passes rise by more than two in five.
TRENDING AI RESEARCH
Cheap AI API resellers often don't match the official API
What it says: Relays resell access to models such as GPT-6 Astra and Claude Opus 5, often below the official price. Six researchers at Queen's University, Kingston, paid for 20 relays (arXiv, 5 November). The first round flagged 7 of 10 low-price relays and 2 of 10 same-or-higher-price relays for behavioural differences from the official APIs. Four returned tool calls with tools switched off, and five reported 4,200 to 60,000 more input tokens (the units you are billed for) for identical requests.
Why it matters: A discount is only real if the token count is honest, and a tool call you disabled could still act if your app runs whatever it receives. The sample was not random, the gap between price groups missed statistical significance (p = 0.07), and seven of 35 relay-model pairs gained or lost a flag on retest.
How you can use it: If you buy through a reseller, run a 20-prompt check against the official API first. Send the same real-work prompts to both on the same day with the same settings, score the facts, and compare input-token counts with the change in your balance. Reject any tool call when tools are disabled, and rerun the check monthly.
TOOL REVIEW
Nod by n8n: a free agent that runs admin errands on your own computer
What it does: Nod is an open-source agent from n8n Labs, in research preview since 10 November. It runs on your computer with its own browser, works in Gmail, Calendar, Slack and Notion, and can log in to sites with no integration. It sends, pays, books or deletes only when your words asked for it; otherwise it asks.
Price (as of 13 Nov 2026): Nod is free under the Apache 2.0 licence. You pay your AI model provider, or use a ChatGPT Plus or Pro plan or a local model. Every task ends with a cost receipt, and Nod checks in once a task passes $5.
Who it's for: solo operators with weekly errands such as chasing invoices or downloading statements. Schedules run only while your computer is awake.
Verdict: Try it if you run repeat admin errands and want an agent on your own machine; skip until it leaves preview if you need it running 24/7.
WHAT GOT FUNDED YESTERDAY
Five rounds announced on 12 November.
Luminal (luminal.com) · AI infrastructure · US. Founded by Joe Fioti (CEO), Matthew Gunton and Jake Stevens, the San Francisco company builds a compiler that finds the fastest way to run an AI model on a given chip, plus a cloud to deploy them. It says that cloud reached about 90 trillion tokens a month within four months. It raised a $25M Series A led by Standard Capital, with Y Combinator, Felicis, SagaVC and Liquid 2.
Biomatter (biomatter.ai) · biotech · Lithuania. Founded by Laurynas Karpus (CEO), Vykintas Jauniškis, Irmantas Rokaitis (CTO), Donatas Repečka and Rolandas Meškys, the Vilnius company designs enzymes with AI for pharma and industry. It says it designs from the target chemistry rather than modifying natural enzymes. It raised an $18M Series A led by Astanor, with UVC Partners, Inventure and Metaplanet.
Opdo (opdo.ai) · optical engineering · Germany and US. Stephan Kuehr (CEO) and Yousef Arzhangnia (CTO) co-founded the company, whose AI agents automate optical engineering from lens design to manufactured parts. It says it works inside tools like Zemax and CODE V, and has cut development cycles from months to weeks for customers including Ford and Bosch. It raised a $2.3M pre-seed round led by Angular Ventures, with Inflection and Tiny.vc.
mistEO (misteo.co) · climate risk · India. Founded by Samuel John (CEO) and Ambarish Narayanan (CTO), the Kerala company tells banks, insurers and agri businesses how weather will affect a specific farm, asset or loan. It says its Indra weather model forecasts at 3 km resolution. It raised a Rs 18 crore (about $2M) seed round led by MGF-Kavachh, with Seafund.
Mondrio (mondrio.io) · pricing software · Netherlands and US. Gregor Biljardt (CEO) and Joep van den Nieuwenhoff (CCO) co-founded the company, which gives B2B software firms AI agents, backed by human pricing experts, to set prices and apply them in quotes and renewals. It says live deal results feed back into its pricing. It raised a $3.75M seed round led by Heartcore Capital, with Pretiosum VC, Hersir Ventures, Cambrena Capital, Momo VC and angels Godard Abel and Onno Oldeman.
QUICK HITS
X launched a native clipping tool; both the original creator and the person who clips the video can earn from clip views through Original Content Rewards, if they are eligible.
Gusto's small business survey found that 25% of small employers now have an AI policy, up from 15% last year.
Nectar launched a Creator Agent that searches more than 300 million Instagram and TikTok profiles and runs brand outreach, with MCP access planned for Claude, ChatGPT and Gemini.
That's today. Hit reply and tell me what you're building. I read every email.
Dev
