Good morning. Here's what's worth acting on today.

BREAKING STORY

OpenAI cuts its flagship price and says the cheap model is nearly as good at code

On 17 November OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in ChatGPT Work, Codex and the API (gpt-6.2-astra and gpt-6.2-sol). Astra now costs $8 input and $40 output per million tokens, 20% below GPT-6 Astra's $10 and $50. Sol stays at $2 and $10. OpenAI says Sol comes within 0.5 points of Astra on FrontierCode 2.0, a test of merge-ready changes to real codebases, at about a fifth of the cost.

Why it matters: most agent loops run thousands of calls, so the model you pick sets your bill. Our estimate for 10 million input and 2 million output tokens a month: $160 on Astra (it was $200 on GPT-6 Astra) and $40 on Sol. The catch is that early power users report regressions.

My take: Point your high-volume agent loops at Sol and keep Astra for the tasks Sol fails; that is where the saving is. Rerun 20 of your own test tasks before you switch, because early users report regressions the benchmarks don't show.

MARKET TRENDS

Brands will spend $116 billion on creators this year, and paying for content is growing fastest

Brands will spend $116 billion on creators in 2026, up 15%, and Madison & Wall forecasts $178 billion by 2030. The biggest slice, 44% or $52 billion, is paid promotion of creator content, and most of it buys ad space from platforms. The fastest-growing slice pays creators to make content, posted or not: 12% of the market, forecast to grow 16% a year.

How to make money from it: this work needs no big following. Split your rate card into a production fee, a fee for posting on your account, and a paid-usage licence priced by duration (30, 90 or 365 days). A brand that runs your video as an ad then pays again when the licence ends. Small businesses can buy three creator videos, run them as ads and renew only the winner.

GROWTH HACK / DIY

Ship a tested prompt library with your AI connector

Connecting your product to Claude or ChatGPT takes minutes. Then the customer opens a new chat, faces an empty box and asks something broad. Livestorm's example: "which of my webinars did best?" gets a confident table that counts the all-hands as a webinar and staff as attendees. On 12 November Livestorm launched the Webinar Prompt Library to fix that "now what?" moment. Its head of growth says the goal is to "help more of our customers adopt the MCP". Each prompt runs on the customer's own data and opens in Claude or ChatGPT in one click. Livestorm has published no results yet.

Do it today:

  1. Pick the ten questions new users should ask in week one, and write each as a numbered brief: what to include, what to skip, what to compare.

  2. Build in the guardrails Livestorm lists: leave out test and internal records, show the numbers, say when data is too thin, and wait for approval before any change.

  3. Test every prompt on three real accounts (busy, new, messy), then track how many connected accounts ask a second question within seven days.

What you'll need: a working MCP connector, Claude and ChatGPT accounts for testing, and about 20 hours for a first set of ten (our estimate). The full playbook has a prompt template and illustrative maths for sizing it.

AGENT LAB

Make shared-order agents read every message before they commit

When several people each send their own agent to one shared order, budget or calendar, the group does worse than if one agent served everyone. In Anthropic's Worse Together study (11 Nov), one coordinator beat per-user teams by about 10 to 45 points; on a group order, Claude Opus 5 reached 88% of the best outcome as a coordinator against 44% as a team.

How it works: two fixes, each for a different job. For orders and bookings, a gate blocks the commit until the committing agent has read every pending message. On failed episodes rerun with the gate, it recovered 73% (Opus 5) and 62% (Sonnet 5), against 45.5% and 30.3% for a plain rerun. For shared budgets, one lead agent ranks every request by value and hands out the allocation: 71% of the best outcome against 30% for a leaderless team (Opus 5).

Try it: in n8n, queue messages in a Data Table and put an If node (unread = 0) in front of the commit step. Test ten orders with and without it.

TRENDING AI RESEARCH

One planted line made AI assistants leave out critical facts in up to 93% of answers

A new preprint from NTU Singapore and Apart Research, Remembered as Brief, tests what five AI models stop telling you once their memory says you want short answers.

A user asks for a summary of a news article containing one line addressed to AI assistants, saying this user wants very short answers and no warnings. Gemini 2.5 Flash-Lite and DeepSeek V4 Pro saved it as the user's preference every time, and 93% and 84% of their later answers left out a critical item, almost always without saying so. Asked where to put a generator out of the rain, one replied "Under an eave or awning", with no carbon monoxide warning. On choices that depend on the user's own priorities, the same two models picked for the user 80% and 95% of the time, against none with neutral memory.

How you can use it: The fix is simple: have the model list the must-know points before it applies your preferences. That restored more content in four of five models.

Caveats: one author, synthetic scenarios, no proven harm to real decisions yet.

Do this today: check your ChatGPT or Claude memory and delete any "keep it brief" rule you didn't set.

TOOL REVIEW

RankSpot: an SEO agent you pay per action

What it does: RankSpot connects to Claude, ChatGPT, Cursor or n8n over MCP so your assistant can do your SEO: check where you appear in Google and AI answers, write and publish articles, and pitch editors for links from your own mailbox. It relaunched on Product Hunt on 17 November, and the company says articles and outreach emails wait for your OK before they go out.

Price: There is no subscription. The live pricing page lists $0.004 per AI visibility check, $3 per article and $0.025 per outreach email, from a $10 top-up. The company says most sites spend $30 to $90 a month.

Who it's for: It suits founders and small sites already working in Claude or ChatGPT.

Verdict: worth trying for a small site you run from Claude or ChatGPT. Use the cheap visibility checks freely, keep articles and outreach on manual approval, and edit every draft before it goes live.

TODAY'S AUTOMATION

A contact form that files itself and pings you

If enquiries reach you as a pile of form notifications, a few go unanswered and none end up in a list you can sort. Today's build fixes that. Someone fills in a short form with their name, email, company and message. n8n tidies the answers, adds them as a new row in a Google Sheet and emails you an alert. Press Reply on the alert and your answer goes straight to the person who wrote in.

Who it helps: freelancers with a portfolio site, cafés taking catering requests, coaches booking discovery calls and small agencies collecting briefs.

Apps: n8n (its built-in form, so no Typeform or Tally needed), Google Sheets and Gmail. Build time: about 20 to 30 minutes.

How it works: one clean-up step trims stray spaces, lower-cases the email address and converts the timestamp to your time zone. Then two steps run side by side, so the alert still arrives if the sheet fails.

WHAT GOT FUNDED YESTERDAY

Five startups that announced rounds on 18 November.

Perry AI (useperry.com) · legal tech · UK. Founded by Shashwat Patel (CEO) and Vaneesa Agrawal (CPO), the London company automates legal work for private capital funds and their law firms, from deal data to compliance. It says it "remembers every position a fund has taken and acts on it". It raised a $2.2M seed round led by Haatch, with Antler, Syncorp Ventures, Araya Ventures and US angels.

Humyn Labs (humynlabs.ai) · robotics data · India. Founded by Manish Agarwal (CEO) and Ishank Gupta, the Bengaluru company films people doing everyday tasks with its own camera rig and labels the footage for robot training. It says it is a deployment partner to AI labs, not only a data supplier. It raised $47M: a $26M Series A led by Accel, with Prosus Ventures and others, plus $21M of venture debt led by Lighthouse Canton.

Vero Security (vero.security) · AI security · Israel. Founded by Elad Tessler (CEO) and Daniel Meirov (CTO), it finds the AI agents running inside a company and blocks actions that break its security policies. It says AI security needs an operating model "designed for AI rather than borrowed from other areas of security". It raised an $11M seed round led by TLV Partners, with GTM Capital, Maple Capital and angels.

Unsiloed AI (unsiloed.ai) · document AI · US. Founded by Aman Mishra (CEO) and Adnan Abbas (CTO), the San Francisco company turns complex documents, tables and charts included, into data AI agents can use. It says its model reads page text and page images in two separate streams. It raised a $7.5M seed round led by Felicis, with Y Combinator, Orange Collective, Uncorrelated Ventures and angels.

Tesorai (tesorai.com) · life-science data · US. Founded by former Google and Verily staff Melissa Miao (CEO), Peter Cimermancic (CTO) and Charlie Kim (CSO), the San Diego company makes raw lab data from past and new experiments searchable. It says its platform "reclaims the value of experiments that have already been done". It raised a $5.3M seed round from Perplexity Fund, IQT and six others; no lead was named.

TICKER

Southern Cross Electrical reports what it calls its biggest batch of work, led by a Sydney data centre

Southern Cross Electrical (ASX: SXE) is a Perth-based contractor that wires, powers and fire-proofs data centres, buildings, mines and infrastructure across Australia. On 16 November it announced over A$300m of new awards, which it called the largest single announcement of works in its history, led by the electrical design-and-construct contract for NEXTDC's S4 data centre in Sydney. It won S4 after about two years of early cost-planning work with NEXTDC and builder Multiplex, and the company says both data centre wins came from repeat clients.

The numbers: an A$810m order book at 30 June (up from A$685m), FY26 revenue of A$718.7m, down 10.3%, data centre revenue the company forecasts rising from A$120m to about A$400m in FY27, and a A$3.85 close on 19 November, below its 50-day and 200-day averages.

The main risk is customer timing: SXE fell from A$4.25 (20 October close) to A$3.54 (26 October) over four sessions after NEXTDC cut guidance over a power delay in Melbourne.

General information only. Not financial advice or a recommendation to buy or sell any security. Data as of 19 November 2026.

QUICK HITS

That's today. Hit reply and tell me what you're building. I read every email.

Dev