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

BREAKING

Hugging Face now sends its language-model calls through OpenRouter

Since 16 November, language-model calls through Hugging Face Inference Providers go to OpenRouter, which picks the provider for each call based on live price, speed and reliability. If that provider fails, the call is retried on another. Hugging Face says it adds no markup, so you pay the provider's rate on your Hugging Face bill. OpenRouter says 160 models from 38 labs are available this way today.

A suffix on the model ID steers routing: :cheapest picks the lowest price, :fastest the highest throughput, and a provider name such as :novita uses only that provider. One breaking change, also reported by TestingCatalog: your own provider API keys no longer work for language models. They still work for image and video providers.

My take: adding :cheapest to a model ID now finds the cheapest provider on every call. If you use your own provider keys through Hugging Face, check bills and error logs this week.

MARKET TREND

Enterprise AI spend hits $176.8B, and the open room is in narrow tools

Menlo Ventures' fourth State of Enterprise AI report (3 Nov) estimates companies are spending at a rate of $176.8 billion a year on using AI, 5.4 times a year earlier. It is a September run rate from a survey of nearly 400 large US buyers, and Menlo invests in Anthropic and OpenRouter, so read it as an estimate. Coding takes $34 billion, while marketing ($1.2B), sales ($1.4B) and creator tools ($1.5B) are small categories, and marketing and sales have no settled winner. Menlo says money goes where an agent can act, has the context it needs, and produces a result someone can check.

How to make money from it: pick one marketing, sales or creator job. Build it only if the tool can act in a system your customer already uses, the context fits in a short rules file, and the result can be checked on the day and with a number within four weeks. Write the pass/fail checklist first, run it free for three customers for four weeks, then price per job with a monthly floor.

GROWTH HACK (PLG)

A year-in-review your users send to their clients

The idea: the Wikipedia apps will show readers a Year in Review in early December, joining Spotify Wrapped, which drew more than 300 million engaged users and 630 million shares in 2025. If your users serve clients of their own, give your recap a second job: alongside each user's year in numbers, make a one-page "your 2026 with us" card per client, to send as a year-end thank-you. A small "Made with [Product]" line shows the client where it came from. Users get a ready-made client report; you get return visits and warm introductions. Skip it where client data is sensitive, such as health or legal work.

Do it today:

  1. Pick three to five numbers per side: business figures for the user, relationship figures (projects, hours, turnaround) for the client. Never show revenue, rates or margins to clients.

  2. Make privacy the default: cards only for clients with three or more jobs and no disputes or overdue balances. The user previews, edits and sends every card, and share links expire after 60 days.

  3. Check 20 recaps against raw records, then email every active user theirs in the first week of December, with a referral parameter on the "Made with" link.

What you'll need: your usage data, an HTML template rendered to image or PDF, and your email system. Estimate one to two weeks for a solo builder. Judge it on send rate, activated signups and January retention.

AGENT LAB

Your agent needs evals

More customers now send an AI agent to use your product for them. Vercel's blog laid out a method (16 Nov) for testing that: give an agent a real request, let it work through the whole task in a copy of your product with fixed test data, and have a small script check the result against each requirement. Run it several times, because "one good manual test can't tell you whether a fix worked."

Why it matters: in Vercel's illustrative example, an agent that succeeds 60% of the time almost always gets the job right once in five tries (pass@5, about 99%) but gets all five right only about 8% of the time (pass^5). For orders, bookings and payments, the second number is the one your customers live with.

Try it: write four tasks, including one impossible request where the right move is to tell the user. Run each five times with the agents your customers use, score every requirement separately, then read the failed runs. Vercel's first run found 9 of 10 "passes" on its impossible task were agents giving up without looking. Estimate: a weekend.

AI RESEARCH

Coding agents finish 95% of single requests and 51% of 20-step chains

What it says: in BOBA-Bench (1 Nov), two Amazon researchers ran ten coding agents through 280 chains of 20 consecutive changes from real Python projects. The agents completed 95% of requests one at a time but only 51% of chains. An agent's own tests pass, it reports success, and a missed mistake makes the next request fail about twice as often. Few mistakes were fixed later, and models judging finished work were right 53% of the time.

Why it matters: most people chain requests, and single-request scores hide this. The best model, Claude Opus 5.5, completed 66.7% of chains. A plain "that's wrong" after a failed step, with one repair attempt, lifted completion to 59%. Caveats: a preprint, Python only, with code not yet public on 18 November.

How you can use it: write your check before each request, keep it where the agent cannot edit it, and commit after every step that passes. When a check fails, say so and name the failing test before any new request. A second AI's approval is no substitute.

TOOL REVIEW

readywhen: an email assistant that chases the replies you are owed

What it does: you email readywhen a job such as "Chase Sam about the invoice". It drafts the follow-up, sends it from your own inbox when you reply "send", then watches the thread and offers another nudge if nobody answers. Nothing goes out without your OK. The email version launched on Product Hunt on 16 November and finished #3 product of the day.

Price: as of 18 Nov 2026 it is free in beta, with no pricing page; the co-founder says premium pricing will come "shortly".

Who it's for: freelancers, small agencies, trades and consultants waiting on invoices, quotes and sign-offs. It reads a lot of your email, per its privacy notice, so connect only your main business inbox.

Verdict: Try it now while it is free if unpaid invoices and unanswered follow-ups slip through your week; wait for published pricing before you build a routine around it.

TODAY'S AUTOMATION

A morning email that summarises your reading list

This n8n workflow reads a blog's RSS feed at 7am each day, keeps the five newest posts, asks an AI model to write one plain sentence about each, and emails you the result. RSS is the public list of recent posts most blogs publish, so no login is needed.

It suits anyone following a few industry blogs or competitors: one email, five sentences, five links, and you open only the posts that matter. The tutorial estimates this cuts 10 to 15 minutes of checking a day to about 2, roughly an hour a week.

Apps: n8n (Cloud or the free self-hosted version), an OpenAI API key and Gmail. Build time: 20 to 30 minutes the first time.

How it works: a Schedule Trigger starts the run, an RSS Read node fetches the feed, a Code step merges the five posts so the AI runs once, and Gmail sends the summary.

WHAT GOT FUNDED YESTERDAY

Five startups that announced rounds on 17 November.

  • Remanence (remanencelabs.ai) · AI infrastructure · France. Founded by former H Company staff Pierre-Louis Cedoz (CEO), Hamza Benchekroun and Axel Moyal, it turns AI agents' logs into training for open-weight models a company owns. It says AI companies "spend more on intelligence every quarter and own none of it". It raised a $30M seed co-led by EQT Ventures and FirstMark, with Zero Prime, Emblem, Drysdale, Global Founders Capital and Yellow.

  • Fermat (fermat.app) · fashion design software · Spain. Founded in Barcelona by Pol Baladas (CEO) and Oscar Cuatrecasas, it turns designers' sketches into realistic renders and, soon, factory-ready specifications. Its CEO says generative AI is "the first technology that can match" fashion's need for speed. It raised an $8M seed led by Kibo Ventures and Lunar Ventures, with PROfounders Capital, Masia Ventures, Blue Wire Capital, Common Magic and Betaworks.

  • Asynchronics (asynchronics.com) · space software · Poland. Founded in Warsaw by Adam Chikha (CEO) and Serge Barral (CTO), it builds digital copies of spacecraft parts so engineers can test flight software before the hardware arrives. It says its open-source engine, NeXosim, is "the only simulation framework" that spreads work across processor cores automatically. It raised a $3.1M seed led by b2venture, with Movens Capital, Heartfelt and Hard2beat.

  • Lemonpeel (lemonpeel.ai) · AI video · India. Founded by Hardik Goyal (CEO), Riddhiman Pal (CPO) and former TVF director Abhinav Nikam, it gives creative teams AI agents that make, edit, check and deliver videos. It says its engine routes work across more than 300 AI models while keeping characters, logos and products consistent. It raised a $1M pre-seed round led by Shastra VC, with Rukam Sitara and angel investors.

  • Treat (treat.health) · health tech · Finland. Founded by former Swappie CEO Sami Marttinen, doctor Antti Mauno and CTO Onni Wahlberg, it sorts a patient's symptoms with AI before a doctor who has read their history takes over the chat, offered through employers. It says it "makes zero revenue from lab tests, clinic visits, or referrals". It raised a €3M pre-seed round co-led by byFounders and Illusian.

BELOW THE TICKER

Similarweb: AI deals lift revenue growth to 16%

What it does: Similarweb estimates traffic across websites and apps and sells that data to marketers, investors and, increasingly, AI companies through an API or MCP. Self-serve plans start at $125 a month (as of 18 Nov 2026); the AI-facing access is sold by contract.

Why it's in the news: it reported its third quarter on 17 November. As of 17 Nov 2026, revenue was $83.0 million, up 16% after 9% growth in the second quarter. GAAP earnings per share were $0.04, against a $0.05 loss a year earlier, and net revenue retention rose to 104% from 98%, so existing customers now pay more than a year ago. It says AI-related deals drove about two-thirds of recurring revenue added.

Main risk: growth leans on fewer, larger contracts and on AI companies, a young customer group, and "AI-related" is the company's own label with no published definition. In February, longer sales cycles led to a revenue miss and a 34.6% one-day share fall.

General information only. Not financial advice or a recommendation to buy or sell any security. Data as of 17 Nov 2026.

QUICK HITS

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

Dev