Issue #3 · June 27, 2026 · ~5 min read

The week the industry stopped paying for wasted tokens

Welcome back to Zero to MVP Weekly. Same deal: one signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. Let's go.

1. The signal of the week

Everyone led with a new model this week. The story underneath it is that the industry stopped paying for wasted tokens.

On June 26, OpenAI previewed GPT-5.6 — three models, Sol, Terra, and Luna, sold as a price ladder as much as a power ladder: Luna runs around $1 / $6 per million tokens against Sol's $5 / $30, with caching rules built to make repeated context cheap. You can't actually buy Sol yet — the preview is gated to a handful of partners — so the usable news isn't the model. It's the shape of the offer: pick a tier, cache aggressively, pay for outcomes instead of bragging rights.

And that's exactly what the buyers are doing. The same week, CNBC reported big AI spenders unwinding what Meta internally nicknamed "tokenmaxxing" — treating tokens burned as a proxy for productivity. Leaders are pushing back on usage-for-its-own-sake, capping per-seat spend, and at least one startup reportedly moved its traffic off a frontier model onto a cheaper open one to crash its cost curve. The question flipped from "how much AI are we using?" to "what did it actually get done?"

For a builder, that's permission to do the smart thing out loud: route the boring bulk — refactors, migrations, test passes — to a cheap or open model, and save the frontier for the parts that earn it. And it's not theoretical anymore. Shopify described an AI stack built to not care which model survives; a Show HN this week dropped model routing straight into Claude, Codex, and Cursor. Same "route, don't switch" idea from issue #2 — now with a business case and the tools to do it.

One caveat to keep us honest: most of the dramatic numbers are big-company anecdotes reported through one outlet, and the cheapest new model is one you can't touch yet. The mindset shift is real. Treat the specifics as direction, not gospel.

2. My take

Tokenmaxxing was a phase. Routing is the skill that will outlast it.

The old instinct was to throw more context and more tokens at the model. Bigger prompt, bigger model, better answer. Sometimes that works. But it is not a product strategy.

A real AI product needs to know which task deserves the expensive model and which task can be handled by something smaller, faster, cheaper, or local. Summarization, extraction, code generation, planning, validation, and final writing are not the same workload. Treating them as one workload is lazy architecture.

That said, there is a trap here. For a pre-revenue MVP, model routing can easily become a beautiful distraction. Nobody cares that your token bill is optimized if the product itself is not useful yet.

So I would treat routing as a habit, not as a feature. Build with the idea that models are replaceable. Keep the boundaries clean. Measure where the expensive calls actually matter.

But ship first. Optimize after the product earns the right to be optimized.

3. Tool worth trying

Google AI Studio Starter Tier — prompt an app, then actually deploy it, no credit card.

I tested Google AI Studio Starter Tier this week, and it earns its place for fast prototyping. The big win: you can deploy a full web app to a live URL straight from the studio — with no credit card and no Google Cloud billing setup at all. For testing a hypothesis or vibe-coding, it's a great fit.

It all happens under the hood. The AI generates the frontend and the backend, and one tap on Publish spins up Cloud Run for hosting, a database (Firestore or PostgreSQL), and even Google sign-in. The platform takes the routine infrastructure and IAM setup off your plate and hands back a working link.

There are limits, of course: the free tier holds up to two active apps per personal account, and traffic quotas are capped. Even so, for an MVP, an internal tool, or showing a client a proof of concept, it's more than enough. Worth a try — especially when you need to build and demo a working prototype fast, without losing time to DevOps.

4. MVP idea of the week

A token-spend co-pilot for small dev teams.

Problem: the moment a team turns on coding agents, the bill goes invisible until it's scary — nobody knows which repo, model, or person is burning tokens, or whether it bought anything.
Solution: a lightweight dashboard over your agent/API usage that attributes spend to repo, person, and task, flags waste, and suggests "route this to a cheaper model" wins.
For whom: small teams and agencies on Copilot / Claude Code / Codex with a real monthly bill.
Minimal v1: ingest one provider's usage export, show spend-by-repo and a weekly "biggest waste" email. No live control yet.
Demand check: offer ten teams a free audit of last month's AI spend. If they want it monthly, you have a product.
Why now: GPT-5.6's tiered pricing plus the public "tokenmaxxing" backlash just turned cost-per-outcome into a number people actually watch.

5. Useful links

OpenAI previews GPT-5.6 (Sol / Terra / Luna) — skip the IQ talk; read the pricing tiers and caching rules. The product is the cost card.
The shift from "tokenmaxxing" to efficiency — the macro behind today's signal: why "more tokens" stopped meaning "more work."
How Shopify built a stack that doesn't care which model survives — model portability as architecture, not ideology.
Smart model routing in Claude, Codex and Cursor — a Show HN you can try this weekend to put "route, don't switch" into practice.
GLM-5.2 is a step change for open agents — the open-weights case keeps getting stronger; the through-line from issue #2.