- August 19, 2026 → OpenAI — date OpenAI and Replit announced Free Mode powered by GPT-5.6 Luna
The meter that used to stop beginners
On August 19, 2026 OpenAI, Replit chief executive Amjad Masad said the economics of building software had finally flipped. The experience now lets anyone “build all sorts of software in Free Mode,” he said — and it runs on GPT-5.6 Luna.
See the OpenAI announcement.
Free Mode is Replit’s bet that the last real barrier to software creation was not skill but cost. For years, building with an AI coding agent meant watching a token meter climb with every prompt, every debugging loop, every failed deploy. Replit was an early user of GPT-3, and Masad has spent the partnership arguing that model price performance — not model cleverness — decides who actually gets to ship.
The math finally bent his way.
What GPT-5.6 Luna actually buys
The model behind Free Mode is the smallest of OpenAI’s latest family OpenAI. GPT-5.6 Luna costs $0.20 per million input tokens and $1.20 per million output tokens on the OpenAI API OpenAI API, roughly a tenth of the price of the larger GPT-5.6 Sol tier. Cheap inference at scale is what lets Replit promise “millions of users” a free experience without bleeding money on every conversation.
That price gap is the whole story. Masad was blunt about it: OpenAI’s recent “price cuts” are what let Replit open Free Mode to “millions of users.” The cut is recent, not ancient — and it is the reason Free Mode exists now.
But free has a ceiling. When a task needs heavier reasoning, Replit routes the work to GPT-5.6 Sol, then returns to Luna while keeping the project’s context intact. The free tier is a doorway, not the whole building.
Power users will still hit the door.
Who this actually helps
Replit’s own customer page lists companies already running the agent in production Replit customers — evidence OpenAI points to as proof the model scales OpenAI: Rokt built 135 internal applications in 24 hours, and Zinus says it saved more than $140,000 while cutting development time by half. Those are paid, heavy users — the opposite of the curious beginner Free Mode is built to catch.
Masad frames the change as a fairness problem OpenAI. He calls software “such an empowering tool” and says widening the pool of people who can build by 100x would spread that leverage far more widely. If anyone online can build a product or a startup and simply get going, he argues, the payoff is “a renaissance-level entrepreneurial boom” the world has never seen.
The promise is sweeping. The risk is that free access manufactures a funnel: capture the beginner, then charge once the idea gets serious. Replit is not hiding that — it is the business model.
The bigger bet on model economics
OpenAI has been pushing the GPT-5.6 family as proof that capability and affordability can rise together OpenAI API. If true, Free Mode is less a Replit feature than a preview of what cheaper frontier models do to who gets to build.
For the model lineup behind this move, see OpenAI’s GPT-5.6 Sol, Terra, and Luna breakdown zbrandco, and how the smaller GPT-5.6 models can now run agents zbrandco.
Masad’s renaissance will be measured in the people who open a tab, type an idea, and never see a bill for trying.
