Decisions API Opens Public Beta as Builders Debate Evals, Moats, and AI-Native Workflows

Key Takeaways

1. Decisions API is now in public beta for all developers, with image input and a claimed speed advantage over GPT-6 Luna via the Responses API.
2. Alex Lieberman argues evals are the top blocker to becoming AI-native, and outlines grader types plus online/offline eval workflows.
3. A 21-year-old student's guitar app reached $25K/month, with Stripe and RevenueCat data showing web and mobile revenue split.
4. Acquire and Flippa listings show small SaaS and e-commerce businesses for sale, including an AI short-video editor at $185K MRR.
5. Codie Sanchez and Nick Huber both emphasize failure tolerance and skill-building over credentials for early-career operators.

1. AI Models and Developer Tooling

1. Decisions API is now open in public beta to all developers, supports image input, and is claimed to make decisions up to 10x faster than GPT-6 Luna through the Responses API, according to Tony Dinh. Arvid Kahl also shared the public beta announcement, framing the API as letting apps choose the right model, tool, or action in near real time, while joking that a "grilling session" turned out to be more about Claude Code than charcoal. 1 2 3
2. Alex Lieberman argues that evals are the number one problem blocking enterprises from becoming AI-native, saying manual prompt tuning and test inputs are not enough. He defines an eval as a grader applied to a trace, where a trace is the full record of an agent run including inputs, each step, and outputs; graders can be deterministic code checks, reference-answer comparisons, rubric scoring without a reference answer, or trajectory checks such as whether an order was looked up before a refund. He distinguishes online evals for production monitoring from offline evals run in a lab or CI, says both are necessary, and describes a promotion loop that turns bad production traces into offline cases. His eval-driven development flow is: define cases and pass criteria, run and observe failures, modify the agent, repeat until passing, then run the full dataset to prevent regressions, with each case run multiple times and thresholds set as all-of-k for high-risk behavior or n-of-k for acceptable variance. 1
3. Alex Lieberman also commented on SpaceX reportedly selecting the best backend model per task, including leading APIs such as Claude Opus 5.5, MidJourney, and Suno, calling it a good way to pressure teams to build the world's best models. Separately, he said he is tired of the "morning briefing" use case and that his current scheduled task monitors birthdays on his personal calendar, reminding him weeks ahead and suggesting gifts. 1 2

2. AI-Native Products and Go-to-Market

1. Alex Lieberman sees a large opportunity in Muse-specific content creation, arguing most AI creators focus on Claude, Codex, and GrokBot while Meta has 3 billion monthly active users, and that non-AI-native audiences will need help learning new technology. He will host a "How to Muse Mog your Friends" session with AI trainer and builder @JJEnglert, who has experimented hundreds of hours since Meta's new personal agent Muse launched; the referenced post says Meta just released Muse, Facebook account users can get a personal AI agent for free for the first time, the app has topped the App Store, and demos span chat, memory files, calendar, computer control, and phone calls. 1 2
2. Starter Story profiled 21-year-old student Kyan, whose first app reached $25K/month. He used ChatGPT to solve a guitar tone problem, found the same unsolved pain point in Reddit and Instagram comments, and built V1 over Thanksgiving week with Cursor without handwriting code, using Supabase, Vercel, Stripe, Mailgun, and OpenAI plus Tavily API, first as a web app and then as a Swift native app with RevenueCat and Superwall. Stripe and RevenueCat data show $11.5K in web revenue and $14K in mobile revenue over the last four weeks, 397 mobile active subscribers in the first month, 1,400 new subscribers in three months, and over 100K total users. He posts three pieces of content daily on TikTok, Instagram, YouTube, and Facebook; after a piece works, he makes 50 micro-variations, hires UGC creators to replicate the format 60 times per month, and adds ad spend, with the quoted approach: "Ride it until it dies. Then repeat." The summarized methodology is to pick a hobby you do weekly, find a small unsolved pain point, ship a web V1 that night, start charging the next day, and post related content daily. 1
3. Noah Kagan built a landing page for a beehiiv alternative in response to a request from Marc, after the referenced post said beehiiv raises prices annually and the customer felt ripped off and asked for the best alternative. Kagan also launched a free Wispr Flow alternative: unlimited for the first week, then permanently 2,000 words per week, or a one-time $20 for unlimited. He compared it with Wispr Flow at $15/month or $180/year, noted that Wispr Flow uploads voice to its servers, and said dictation should not cost that much. He is testing whether free converts better than $20 and will publish data after 30 days; the tool supports Apple Silicon Mac. 1 2

3. Startups, Acquisitions, and Operator Lessons

1. Acquire listed several startups for sale: an Australian men's underwear e-commerce brand with 48% repeat purchase rate, lean Shopify operations, $151.9K TTM revenue, asking $45K; an AI-native PLM SaaS for apparel brands with $1M–$10M+ revenue, enterprise demand and vertical expansion potential, $142K TTM revenue, asking $300K; an AI short-video editor SaaS with $7.2M cumulative revenue, $185K MRR, 3.2M users, $3.1M TTM revenue, asking $2.3M; and a 6+ year old AI e-commerce and personalization platform SaaS with $546.8K TTM revenue, asking $2.6M. Acquire also shared the view that building a startup for $100K and selling for $5M is a 50x return, that software needs no store, factory, large inventory, or hundreds of employees, and that a few people with laptops and $100K can sometimes create a multi-million-dollar asset, though it takes years and a lot of work. It introduced Guided by Acquire, which connects sellers with M&A advisors, assesses buyer fit, and advances deals. 1 2 3 4 5 6
2. Flippa relayed Elliot Omanson of OWLFI on The Exit EP 316 with Steve McGarry: most founders wait for a comfortable moment to sell, and that delay is often exactly when the best buyer window closes. Flippa also listed: a 10-year-old Australian motorcycle accessories group running three Shopify stores, needing only 25–30 owner hours per week, $882K annual revenue, $18.7K average monthly profit, $149 average order value, asking $605,576, in First Access; a one-year-old UK dental implant distributor with exclusive contracts to clinics and labs through December 2027, $75.9K annual revenue, $67.6K annual profit, 2–5 hours per week, asking $233,763, in First Access; and a New Zealand teeth-whitening brand launched in 2025 mainly serving US customers with a contractor team that can stay, $3.04M annual revenue, $52K average monthly profit, $41 average order value, asking $270K. 1 2 3 4
3. Codie Sanchez listed recent failures: paying $275K to a technical vendor and still failing to launch; hiring people who built a team costing $1M per year, going in the wrong direction and halving the close rate; and spending $125K on ads for a failed launch. She said everyone you admire fails more than 10x as often as you do per year. She also argued that being the hardest worker in your own company is a danger sign, meaning poor prioritization, weak boundaries, or still doing low-value work someone else should handle, and said she has been there. Her view on persuasion: you do not convince customers by what you say, but by what you get them to say to themselves. 1 2 3
4. Nick Huber told young people entering the workforce that no one wants to hire them because they lack real skills, that companies must invest time and money to train them, and that college has done little to make them an asset. He said your job is to make the company more profit than your salary and to proactively learn profit-creating skills. Easier skills he listed: learn to write and communicate, be brief and clear, respond quickly with positive high energy, prioritize the most valuable work, stay organized and push projects to completion on time, and ask for more work when done. Harder skills: learn a higher-level skill such as marketing, sales, engineering, finance, accounting, or web design; learn to manage or lead others; practice making decisions with unknown outcomes and improving; and learn to explain why what you do adds value. He emphasized learning these yourself or queuing up for mediocrity. 1
5. Ben Tossell shared the view that the only remaining moat is caring about a project for more than two days, said many things launched today, and predicted that dot, bot, muse, and instinct decisions will die once pi bot is born. He also said he has become comfortable with local tools, files, and skills but questions the claim that cloud is better, asking how exactly to migrate. 1 2 3 4