OpenAI Slashes Prices 80% as Its Billion-User Goal Arrives Late

Column Overview

Two numbers landed within a week of each other, and if you only read the headlines you'd think OpenAI was having its best month ever. On July 30, 2026, the company slashed prices on its GPT-5.6 Luna model by 80% — both for the text you feed in and the text it hands back. A week later, on August 6, OpenAI announced that ChatGPT had crossed 1 billion people using the product. Put side by side, it reads like total dominance: cheaper, bigger, faster. Look closer at the timing and the fine print, though, and a different story emerges — one about a company cutting prices because it has to, and celebrating a milestone that arrived later than it was supposed to.

The discount, broken down

The headline cut applies to GPT-5.6 Luna, OpenAI's mid-tier model: input tokens dropped from $1.00 to $0.20 per million, and output tokens fell from $6.00 to $1.20 per million — an 80% reduction on both sides of the ledger. GPT-5.6 Terra, one rung down, got a smaller 20% trim. GPT-5.6 Sol, the newest and most capable model in the lineup, saw no price change at all. That asymmetry is the first clue that this wasn't a blanket "thank you for your business" gesture — it was targeted at the tiers where OpenAI is under the most competitive pressure, while the frontier model stayed untouched.

OpenAI's official explanation leans on efficiency, not charity. The company says GPT-5.6 Sol was put to work optimizing its own production infrastructure — reportedly rewriting and tuning the GPU inference kernels that serve live traffic, trimming end-to-end serving costs by roughly 20%. That detail deserves to be pulled out and looked at on its own. This isn't a model writing an app or debugging someone's code, the kind of AI-assists-productivity story that's become routine. This is a model being pointed at the machinery underneath itself — the software layer that decides how efficiently GPU clusters turn electricity into tokens — and improving it. Whether or not that alone explains an 80% cut on Luna, it's a genuinely new kind of story: AI companies starting to use their own models to shrink the cost of running AI.

Why now: the pressure is coming from outside

Efficiency gains rarely explain an 80% price cut on their own, and the more convincing driver sits in the competitive data. On OpenRouter, the routing marketplace many developers use to compare and switch between models, Chinese-built models briefly captured 46% of U.S. enterprise token volume — at one point overtaking American models entirely for that segment of usage. That's a striking number for a market segment, U.S. enterprise buyers, that Western labs have generally treated as their home turf.

This lines up with a trend that's been building for a while: the aggressive per-token pricing coming out of Chinese labs, DeepSeek's roughly ¥0.001 per thousand tokens being the most cited example, has been reshaping what buyers consider a normal price. That pressure used to look like a regional story, mostly relevant to price-sensitive markets in Asia. The OpenRouter numbers suggest it isn't regional anymore. It has crossed the Pacific and is now showing up in the enterprise budgets that OpenAI depends on most. An 80% cut on a mid-tier model, timed the way it was, reads less like generosity and more like a company defending share against competitors it didn't expect to be this close, this fast.

The billion-user headline has an asterisk

The August 6 announcement described ChatGPT reaching "1 billion people" — a "people" figure, not a "weekly active users" figure, and the distinction matters more than it looks. The last time OpenAI gave a precise weekly-active-user number was back in February 2026, when it cited just over 900 million. Between then and the billion announcement, the company shifted to softer, broader language, which makes the two figures harder to compare directly and easier to spin.

More telling is the schedule slippage: by OpenAI's own earlier internal targets, this milestone was supposed to arrive roughly seven months before it actually did. A billion users is still a genuinely enormous number, and nothing about hitting it late erases that. But "big and on time" and "big and seven months behind" tell two different stories about how a company is tracking against its own expectations, and OpenAI's growth curve has clearly been the second one.

What this actually tells us

None of this means the AI growth story is fake. Usage keeps climbing every time prices fall, which is the textbook signature of Jevons paradox playing out in real time — cheaper compute doesn't shrink the market, it uncorks demand that was being held back by cost. That dynamic looks intact, and it's arguably the healthiest sign in this whole episode: people are actually using more of this stuff when it gets cheaper, not just switching providers to save money.

But "the underlying trend is real" and "the company is hitting its own targets on schedule" are two separate claims, and this month is a reminder not to conflate them. It's the same tension that's shown up elsewhere in AI's business story lately — the technology working roughly as promised while the capital and growth timelines run ahead of what the numbers actually support. A billion users seven months late is still a real accomplishment. It's also evidence that the pace investors and executives have been promising is harder to hit than the pace the market is actually delivering.

It also sharpens a divide worth watching. OpenAI is choosing to compress margins and burn cash to defend market share against faster, cheaper competitors — a strategy that only pays off if scale eventually converts into pricing power. Anthropic, by contrast, has been reported to be pushing gross margins toward the 70%-plus range, effectively betting that discipline beats volume. Both companies are responding to the same price war; they're just choosing opposite sides of the trade-off between growing fast and staying profitable. Industry observers believe that as this price war intensifies — squeezing even a company as dominant as OpenAI into an 80% cut — the gap between those two strategies is likely to widen before it narrows, and which one turns out to be right may not be clear for another year or two.