Good morning. Here's what's worth acting on today.
BREAKING STORY
Anthropic files to go public as ANT, on $16.3B in half-year revenue
Anthropic publicly filed its S-1 with the SEC on 26 October to list on the Nasdaq Global Select Market under the ticker ANT. The filing shows revenue of $16.3 billion for the six months to 30 June 2026, up from $1.16 billion a year earlier, and more than 1,900 customers paying $1 million or more over the 12 months to 30 September.
The filing says its revenue is "primarily consumption-based, and much of it is not subject to long-term contractual commitments". The headline net loss of $80.7 billion is mostly non-cash: about $72.4 billion is the change in fair value of convertible notes, derivatives and SAFEs. The operating loss was $693 million.
My take: Copy metered, per-task pricing over flat seats. Anthropic earns most of its money by charging for what customers actually use, so your buyers are getting used to paying that way. Your main model supplier now answers to public shareholders, so make sure you can swap models in a day: keep prompts and evals in your own repo, send every model call through one thin layer, and test a second provider on your top three tasks this month.
MARKET TRENDS
Developers saving 10+ hours a week with AI fell from 68% to 26%
Over 95% of developers now actively use AI, yet the share who say it saves them 10 or more hours a week fell from 68% in 2025 to 26% in 2026, according to Atlassian's State of Developer Experience survey. 72% report the same or more friction than before AI, and Atlassian says 49% spend 4+ hours a week reviewing AI output and correcting prompts. InfoQ notes the 1,560 responses are self-reported and come from a vendor that sells developer tools.
How to make money from it: The bottleneck has moved from producing work to checking it, so sell verification. Offer a fixed-price weekly QA retainer that reviews a small team's AI-written code or copy, a review checklist pack for one stack, or eval templates that catch the errors a client keeps fixing by hand. Price it against those four hours a week.
Harvard Business Review Analytic Services on why fast AI answers create coordination gaps in B2B service. Read the briefing paper.
GROWTH HACK / DIY
The proactive first message: let your product speak first, only when it matters
The idea: Your product checks the user's own data every few hours and messages them only when something needs a decision, with the next step already drafted and one tap to approve it. Claude Duet and Kiwi, "a free AI assistant that texts you first", both launched on this pattern in the past week. Neither has published retention results, so test it on your own users.
Do it today:
From support tickets and churn notes, list three to five moments a user would want a heads-up about: an invoice 7 days overdue, a client waiting on a reply, a trial ending with setup unfinished.
Write each as a trigger rule (condition, data to show, action to offer) and run checks every few hours. Cap it at one message per topic per day and three a week, with an optional daily brief for the rest.
Put the drafted action in the message (the reminder email, the reply) with Send, Edit and Not now buttons, and Mute and Change frequency links in every footer. Keep a holdout group on your old drip so you can compare.
What you'll need: Event data with timestamps, a scheduler (cron, n8n or Zapier), one channel users already read, signed one-tap action links and a preferences page. Allow about a week for the first two triggers; cost is mostly your existing sending costs. Track nudge-to-action rate against mute rate for each trigger.
AGENT LAB
The read-back loop: make your agent confirm every detail before it acts
Sierra's τ-Entity benchmark tests whether voice agents capture names, phone numbers, addresses and emails exactly right. When agents chose their own checking strategy, only 14–41% of tasks succeeded across all three audio conditions. A prescribed routine (ask, spell or read back, confirm and correct, submit) raised that to 37–54%, at a cost of 21–28 seconds per call. Use it, in voice or chat, wherever one wrong character costs you: emails, order IDs, phone numbers, postcodes.
How it works: raw value → normalise in code (strip spaces, map "at" and "dot") → validate the format → check it against the world (does the order exist, does the email domain have MX records) → read it back character by character → a yes submits, a correction restarts from the top → after two failed retries, hand off to a human.
Try it: Take the one field your agent gets wrong most often and put it behind a field spec (type, regex, lookup, read-back format, retry limit) held in a state store, so the model can't talk past a step. Sierra found only 24–37% of wrong captures that agents verified were corrected, so test the repair step as hard as the check.
TRENDING AI RESEARCH
Trial by Trace: agents that pass your test can still break the rules
What it says: Trial by Trace, a new paper from TraceLite AI and co-authors, scored 3,420 runs across four agent harnesses and nine models. 31.1% of runs completed the task while failing at least one harness criterion, and only 8.4% passed every check on both the task and the harness.
Why it matters: Checking only the final answer hides process failures such as repeated side effects, swallowed errors and sub-tasks that never report back. The rankings shift too: on the project's results table, Claude Code led on task success (44.0%) and OpenClaw on harness score (72.0%). The authors ran one execution per cell, so treat the rankings as indicative.
How you can use it: Pull ten recent "successful" runs from your own agent and check each one. Did every tool call get a result? Were errors and truncated output passed back? Did any send, pay or delete run twice? Did delegated tasks return? Does the final "done" match the real file or inbox? Any pass with a failed check goes into your regression set. The benchmark code is on GitHub under MIT.
TOOL REVIEW
Loqui: Mac dictation that writes in your tone, in 90+ languages
What it does: Loqui drops cleaned-up dictated text wherever your cursor is: hold fn, speak, and it adjusts tone to the app (professional in email, conversational in chat, technical in code editors). It also records meetings and writes minutes split into decisions, actions, risks and open questions.
Price: As of 27 October 2026, Free with a weekly quota, or Pro at US$8/month (US$80/year) with no dictation cap and 3 meeting hours a month. Pro + Meetings is US$16/month with 30 hours a month. It needs macOS 26 on Apple Silicon; there is a Windows version too.
Who it's for: Solo builders and creators who write a lot of short text every day (client emails, Slack, prompts) and want meeting minutes without a separate notetaker. Audio is processed in the cloud: Loqui's privacy policy says it goes to ElevenLabs for recognition and OpenAI models for formatting, so don't dictate anything you wouldn't paste into a third-party AI tool.
Verdict: Try it. Spend a week on the free quota with three daily writing jobs; if most outputs need only light edits, Pro pays for itself.
WHAT GOT FUNDED YESTERDAY
5 startups that announced rounds yesterday, 26 October.
Nevis (neviswealth.com) · wealth-management fintech · UK. Founded by Mark Swan, Philipp Burda and Ivan Chalov, Nevis builds an AI operating system whose agents handle advisers' admin: CRM, meeting notes, planning and custodial paperwork. It says its agents finish the work rather than suggest it, with 100,000+ tasks completed. It raised a $70M Series B led by EQT Ventures, with Charles Schwab, QuantumLight, Sequoia Capital, ICONIQ and Ribbit Capital.
Matiks (matiks.com) · social gaming · India. Sudhanshu Bhatia, Sushant Timmapur and Mohan Kumar run an app of competitive maths, memory, logic and puzzle games with friend challenges and daily streaks. It says it has 300,000+ daily active users and will use the money to launch an AI tool for building games without code. It raised a $10M pre-Series A led by General Catalyst, with Ronnie Screwvala, Info Edge Ventures, Tanglin Venture Partners and Untitled Ventures.
Ailytics (ailytics.ai) · vision AI for industrial safety · Singapore. Co-founders Lenard Tan and Prateek Manocha turn existing CCTV cameras on construction and heavy-industry sites into real-time safety monitoring. It says its Ailyssa engine builds a 3D picture of a site from a single 2D feed and works on decades-old cameras, with 400+ deployments in 15 countries. It raised a US$7.4M Series A led by Tin Men Capital, with SEEDS Capital, Safie, a family office and existing investors.
Colare (colare.com) · hiring tech · US. Co-CEOs Esther Thomas and Nain Abdi screen mechanical, electrical and embedded engineers with simulations of real design work such as CAD, PCB and thermal tasks. It says rubrics from senior hardware engineers score both a candidate's approach and output, saving the roughly 40 hours of senior time a hire typically takes. It raised a $2.5M pre-seed from Afore Capital, Garage Capital and Acadian Ventures; no lead named.
Medikea (medikea.co.tz) · healthtech · Tanzania. Co-founded by Dr Elvis Silayo (CEO) and Desirè Ruhinda, who has since stepped away, Medikea combines 24/7 virtual consultations with its own clinics, pharmacies and labs. It says it is vertically integrated, serving cash-paying and insured patients across one network. It raised a $1.2M pre-seed from Baobab Network, Catalyst Fund, develoPPP Ventures, Madica, Impacc, Hustle Fund, O'Shaughnessy Ventures, Madiro and the Boehringer Ingelheim Foundation; no lead named.
QUICK HITS
OpenAI released GPT-6.1 Terra on 26 October for computer use and writing, at $1 input, $0.05 cached and $5 output per million tokens, 50% below GPT-5.6 Terra.
Canva launched Craft, its first in-house image model, with 2K output and 30 aspect ratios, in Dream Lab and the Canva API.
New paid-tier Gemini API projects now store Generate Content API requests by default (store=true); existing projects keep their setting, and you can opt out with store=false or in Logs and Datasets settings.
That's today. Hit reply and tell me what you're building. I read every email.
Dev
