| Sunday · May 3, 2026 |
Issue № 003 |
14 min read |
|
|
The FindSkill Weekly Brief
The Skill
Just for FindSkill Pro members. The AI news that actually matters for your work.
|
|
Pro Members Only
|
A Private Brief
|
Not Published Anywhere
|
|
|
|
Mia
AI Learning Editor · FindSkill.ai
|
Welcome to Issue 003.
I'm Mia, and every Monday I land in your inbox to make sense of the week's AI news — without the jargon, the hype, or the "10 things you MUST know" energy. This brief doesn't get published anywhere else. It goes to Pro members. That's it.
Three things happened this week that look like three different stories. They're one story. Anthropic priced itself at $900 billion. Microsoft and OpenAI quietly broke up. Apple confirmed it's paying Google a billion a year for a custom Gemini. The era of one-AI-for-everything ended this week. Last issue was the lens for actions (agents — when AI does things). This issue gives you the lens for choice (which AI for which thing).
Quick callback: I asked last week for the funniest things your inbox-triage agent did. Half the replies said "it auto-archived a wedding invitation." One said it drafted a reply to her therapist scheduling lunch. Section 03 below is a different experiment — same five prompts, three different models, five days. The pattern isn't what I expected.
— Mia
|
01 |
This Week in AI
Three headlines. One pattern.
|
Business
Anthropic priced itself at $900B — and the gap to OpenAI isn't the news
Bloomberg reported this week that Anthropic is taking funding offers at a $900 billion valuation, with a May board meeting to settle terms. For comparison, OpenAI's last round closed at $852B. Anthropic was at $380B in February — so this is more than a doubling in two months. The headline number is shocking. The number under the number is more interesting: Anthropic claims $30B annualized revenue against OpenAI's $25B. The wedge isn't ChatGPT clones. It's Claude Code. Anthropic's revenue mix is roughly 80% enterprise API plus Claude Code; OpenAI's is consumer-weighted. The market is voting on which kind of AI company makes more money — and right now, "the one developers build with" is winning.
|
What it means for you
If you're on an annual ChatGPT or Claude plan, there is real upward price pressure coming this year. We wrote the full subscriber-pricing decision framework last week ( the $900B vs $852B post →). Short version: don't commit to annual unless you'd accept paying 1.5× current rates. Test what your top three real prompts feel like in two other models before your renewal — the next twelve months are the right time to be portable, not loyal. |
Business
Microsoft and OpenAI quietly broke up. Then Microsoft launched a 4-model agent platform.
April 27, Microsoft and OpenAI restructured their partnership. Three things changed: Microsoft's IP license is now non-exclusive (still through 2032), OpenAI got multi-cloud freedom (live on AWS Bedrock the next day), and the contractual AGI trigger that locked Microsoft in is gone. Forty-eight hours later, Microsoft Agent 365 went GA at $15/user/month — a multi-model orchestration layer where Anthropic Claude, OpenAI, Mistral, and an "Auto" mode (Microsoft picks) all sit in one dashboard. Word, Excel, PowerPoint can each call a different model now. Not because Microsoft hates OpenAI. Because Microsoft watched the market and concluded one model is no longer the right answer at the platform layer.
|
What it means for you
Your employer's Microsoft 365 tenant will get Agent 365 controls within weeks if it doesn't have them already. If you're in IT, security, or compliance — go look at the multi-model toggle. If you're a knowledge worker — start expecting that your "Copilot" might quietly be Claude on a Word doc and GPT on an Excel sheet, depending on the task. We have the operational read on the breakup ( Microsoft-OpenAI exclusivity end →). Read it before your next compliance meeting. |
Business
Apple paid Google a billion a year for a custom Gemini. That's the whole story for Siri.
April 22, Google's Cloud chief Thomas Kurian publicly confirmed at Google Cloud Next that Gemini will power the next Siri. The deal was first signed in January. Bloomberg reports Apple is paying Google approximately $1B/year for a custom 1.2-trillion-parameter Gemini variant. Phase 1 is already live — iOS 26.4 uses Gemini for context awareness and on-screen recognition. Phase 2 ships with iOS 27 in September: "Conversational Siri." Apple, the company that wrote "designed by Apple" on every product since 1997, just told the market: even Apple buys models from someone else. The single-vendor era is over not because of policy or regulation. It's over because the math stopped making sense.
|
What it means for you
This week's three stories all say the same thing — one AI is no longer enough. Anthropic won by being the better second choice. Microsoft built the layer that picks. Apple skipped the question and rented the answer. The implication for your week: stop trying to find your single "best" AI. Start picking the right one for the right task. Section 02 names the concept. Section 03 shows the five-day experiment. Section 04 is the prompt to set it up.
---
|
|
02 |
Term of the Week
The one concept to understand this week
|
| |
|
Term №003
Model router
< model router *(no expansion — the term IS the description)* >
A layer that sits between your app — or your prompt — and a bunch of AI models, deciding which model is the right one for this task right now.
|
Think of it like this →
A maître d' for AI. You walk in and say what you need — "fast and cheap," "deep reasoning," "code that runs," "image with text on it." The maître d' walks you to the right table. Same restaurant. Different chef per dish. You don't hire the chef. You hire the person who knows which chef to give your ticket to.
|
|
⚠ Common misconception
"We need to pick one AI." No. You pick a router. Different models are good at different things, and you're already paying for two or three of them. A router lets you use the best one per task without thinking about it. The biggest cost mistake of 2025 was treating AI like one product. The biggest opportunity for 2026 is treating it like a stack.
|
Where you'll hear it: Microsoft Foundry's "model router" feature in Azure. Microsoft Agent 365 (Story 2 above — that whole thing is a router). OpenRouter (consumer-friendly, single API to a hundred models). Portkey (enterprise gateway). LiteLLM (open-source). Perplexity itself is a model router for research — it picks Claude, GPT, or its own Sonar variant behind the UI. IDC predicts 70% of top AI-driven enterprises will use multi-model routing by 2028; 37% are already running 5+ models in production. Cost savings from intelligent routing: 40–85%. The category went from theory to mainstream while everyone was watching agent demos.
Routing Claude Code to a different model in practice →
|
|
03 |
Deep Insight
I gave Claude Opus 4.7, GPT-5.5, and Gemini 2.5 Pro the same five prompts for a week. Here's what each one wins.
|
Most "Claude vs GPT vs Gemini" posts run one prompt through three models and call it journalism. I ran five real prompts through three models for five days and tracked which one I actually trusted with the output. The result is less binary than the headlines suggest — and more useful.
I built this experiment to settle a question I keep getting from Pro readers: which one should I cancel? The answer I came in expecting was "Claude for everything except code." The answer I came out with was "stop trying to cancel any of them — start using them as a stack." I'll show you the five tasks, the model that won each, and the small piece of paper now stuck to the side of my monitor.
Five tasks I do every single week: (1) summarize a 30-page PDF into one paragraph; (2) write a one-to-one sales email; (3) draft the same email for a list of 100 people; (4) translate plain English into a SQL query against a schema I describe; (5) brainstorm 10 hooks for a piece of content. Three models: Claude Opus 4.7 (Claude.ai), GPT-5.5 (ChatGPT Plus), Gemini 2.5 Pro (AI Studio). Same wording across all three. No per-model tuning — most readers don't tune, they paste, so I didn't either.
Five days, five tasks, three models | Task | Claude Opus 4.7 | GPT-5.5 | Gemini 2.5 Pro |
|---|
| 30-page PDF → 1-paragraph summary | Best structure, easiest to skim. Quietly drops the boring middle. | Most thorough — keeps all the data, worse for skimming, better if I'm citing. | Surprisingly good but hallucinated one stat. Caught it on second read. | | 1-to-1 sales email | Reads like I wrote it. Voice transfer is the thing it's better at than anything. | Reads like a sales person. Polished but generic. | Reads like a competent intern. Workable but not me. | | Mass send to 100 people | Survives sending to a list. Caught my "I noticed your" cliché. | With voice memory pinned: best balance. Predictable across the 100. | Adds three emojis I didn't ask for. Useful starting point if I delete them. | | English → SQL | 4/5 correct. Failed on a window function. | 5/5 syntax, 4/5 intent. Named the wrong column once. | 5/5 syntax AND 5/5 intent. Caught a data-type mismatch I hadn't flagged. | | 10 content hooks | 7 useful, 3 overcooked. Voice solid. | 5 useful, 5 generic ("Unlock the secret to..."). | 8 useful, 2 hyperbolic. Best at concrete, surprising angles. |
After five days I stopped picking favorites and started picking roles. Four patterns held up across every test:
| i. | "Voice-heavy → Claude." When the output has to sound like a specific person — sales emails, replies, edits to my writing — Claude is the cleanest. Less smoothing. More retention of the input voice. This is the thing it does better than any other model for my work, and I'd cancel ChatGPT before I'd cancel Claude on this one job alone. | | ii. | "Reasoning-and-data → Gemini." SQL, structured data, math, "find the contradiction in this document." Gemini was sharper than I expected and clearly ahead on the parts where one wrong answer matters. The hook generation also surprised me — Gemini wrote the most concrete, specific hooks I'd actually use. The bias against Gemini in your circle is a year out of date. | | iii. | "Volume + general-purpose → GPT-5.5." When I'm running the same prompt through 100 inputs (mass emails, batch summaries, scaling a one-off into an automation), GPT-5.5's consistency wins. Not always the best per output. The most reliable. Reliability scales. This is also why every "agent that does work for you" platform defaults to GPT under the hood — it's the one whose output you can plan around. | | iv. | "The combined verdict." I built a small router. Not a software router. A Notion table with five tasks, three models, and the assignment for each. It lives stuck to my monitor. The 30 seconds of decision-up-front is worth not having to redo the work — and I now use all three subscriptions I was about to cut down to one. |
What the router looks like Input · how I used to work Open Claude. Paste prompt. Read output. Sometimes okay, sometimes wrong. Try GPT instead. Sometimes worse, sometimes better. Settle for the second answer because I've already burned 10 minutes. Repeat three times a day, every day. Paying for two subscriptions but only really using one. | | | Output · how I work now Glance at the router. Pick the model. Paste once. Read once. Done. Five tasks per day, the right model on the first try, ten minutes a day saved. Paying for three subscriptions and using all three. Net cost up. Net work down. The math finally works. |
Try the same five prompts in the three models you already pay for. Don't use the "best" one — use each one. Reply with which model surprised you and which one you stopped trusting. The pattern in your work is probably different from mine, and the router for your work is the actual deliverable.
|
04 |
The Workflow
One way to use AI at your job this week
|
| |
Stop asking "which AI is best." Ask "which AI for this task."
Spend 15 minutes today writing a model router for your own work. The format that beats every alternative I've tried is a five-row table. Column 1: the five things you ask AI to do most often this week. Be specific — "summarize meeting transcripts," "draft client emails in my voice," "translate plain English to SQL," "edit my writing for tightness," "research one topic across 10 sources." Column 2: what makes the output good — "matches my voice," "stays under 100 words," "syntactically correct," "cites the source," "scans 10+ pages without hallucinating." Column 3: the model you're assigning. Use the patterns from Section 03 — voice-heavy → Claude, reasoning/data → Gemini, bulk + consistency → GPT-5.5, deep research with citations → Perplexity. Pin the table where you write. For seven days, do not improvise. Use the assigned model. Yes even if you're "sure" the other one would be faster.
Why it works: The cognitive overhead of choosing isn't zero. Decision fatigue compounds. By moving the choice out of the moment of work and into a one-time setup, you remove the lookup tax forever. The thing the multi-model era is actually selling isn't more capability — it's a way to stop guessing. Pinned to your monitor, it's a router. Without that table, you're paying for three AI subscriptions and using whichever one you opened last. With it, you pay for three and use three.
|
Do this week
Make the table. Don't make it good. Make it exist. Refine it Friday. After two weeks you'll know what's working and what's not, and you'll add a column for the prompt itself — at which point you've reinvented something a $40-per-seat enterprise tool charges for. You built it in a Notion page in 15 minutes. Welcome to model routing.
|
|
|
05 |
The Side Play
One way to make money with AI this week
|
| |
| ◆ Income Idea · Play №003 |
Be the person who tests every AI for one specific thing. That's a job now.
Pick one niche you already know. Not "AI for entrepreneurs." Too generic. Pick: "Claude vs GPT for legal contract review." "Gemini vs Claude for real-estate listing photos." "ChatGPT vs Perplexity for medical literature search." "The three model routers for small-business owners." A specific use case crossed with the specific people who do that work. Format: weekly, same template. "I gave [X model] and [Y model] and [Z model] this prompt. Here's what each one did. Here's which one I'd actually use." 8–12 minutes if video. 600–900 words if newsletter. Five tweets if X. Distribution: pick one platform you'd realistically maintain. YouTube and Substack pay better long-term. X builds the relationships. Realistic revenue: $200–$2,000/month within six months — sponsorships, affiliate links to the AI tools, then your own products once you have an audience that trusts you.
Why it works: Demand is exploding. Every model release means thousands of "is this any good for X?" questions go unanswered for weeks. Supply is small — most reviewers do "best AI of 2026" content that's useless to a niche professional. The professional doesn't want a leaderboard. They want one person who tests the same five tasks across every model release in their niche, every week. Pick that lane. Most of the audience you need is already searching for you. They just can't find you because you haven't published yet.
|
Do this week
Write the format spec for one episode. Not the channel name. Not the about page. Just the structure of one episode: three models, one prompt from your niche, three outputs, your verdict. If you can write that template in 30 minutes, you can do this. Hit publish on episode one before the weekend.
|
|
|
06 |
The Stack
Three tools I'm testing this week
|
Not affiliate picks. Not sponsored. Just three multi-model interfaces at three price points so you can feel the router thing without writing code.
01 |
Poe
Multi-model chat aggregator · $20/month
The fastest way to feel what a "model router" actually means without writing a line of code. One subscription, dozens of bots powered by Claude, GPT, Gemini, DeepSeek, Mistral, even open-source models. Type the same question into three of them. You'll learn more in 10 minutes than from any blog post about routing. The UI is not the most polished. But for teaching yourself the concept, nothing else gets you there faster.
|
02 |
OpenRouter
API gateway for 100+ models · Pay per token
For anyone who builds anything. One API key, every model. The reason it sees 90,500 monthly searches and the related Claude-Code-Router project has 31,000+ GitHub stars: developers are voting with their integrations. If you've ever wished Claude Code could swap to a cheaper model for the boring parts, this is the layer. Paying for ChatGPT and Claude and Gemini? Cancel two of them and route through OpenRouter — for most personal use cases, you'll spend less.
|
03 |
t3 Chat
Multi-model chat for builders · ~$8/month
The newer entrant from Theo Browne. Lean, fast, multi-model — opens in milliseconds, syncs across devices, costs less than Poe, and skews more dev-friendly. Built specifically for "I want to use the right model and don't want a slow UI in the way." If you live in chat all day and the UI lag of ChatGPT/Claude bothers you, the speed alone is the feature.
|
|
07 |
Inside FindSkill
What's new for members this week
|
| New |
Claude Code with DeepSeek V4
For devs who want Claude Code's interface but DeepSeek V4's pricing. The exact "swap the model under the chat" workflow Section 02 is about. Start the course →
|
|
| New |
Enterprise AI Rollout Playbook
For Pro members leading rollouts at work. Vendor selection, governance, the unsexy parts. Start the course →
|
|
| New |
ChatGPT Workspace Agents for Non-Engineers
For HR, ops, marketing. Build a working agent without writing code. Start the course →
|
|
|
|
Which task in your week wants a different model than the one you're using for it?
Reply to this email. I read every one. Bonus: tell me which model surprised you most in the five-prompt experiment, and I'll factor it into next week's brief.
↵ Hit reply
|
|
The Skill · by FindSkill.ai
|
|