OpenAI cut Luna's price 80%. Read what it left untouched.

On July 30, 2026, OpenAI cut the price of GPT-5.6 Luna by 80% and Terra by 20%, and left Sol alone. Luna's input price fell from $1 to $0.20 per million tokens, its output from $6 to $1.20. Terra dropped from $2.50 and $15 to $2 and $12. Sol, the top coding model, held at $5 and $30. The models were about three weeks old.

Read the shape of the cut, not the headline. OpenAI took its cheapest tier to the floor, trimmed the middle, and did nothing to the top. That asymmetry says more than the 80% does. It marks where OpenAI still sets its own price and where the market now sets it for them.

If you make money on top of the API, that map matters more than the discount. The tiers falling fastest are the ones you can no longer charge a premium to resell. The tier that held is the one still worth paying for. Below is what changed, why the cut is lopsided, and what to do if your margin rides on token cost.

What actually changed

Three GPT-5.6 models, one announcement, three different treatments.

Luna, the cheap tier, got the full cut: $0.20 per million input tokens and $1.20 per million output, down from $1 and $6. Terra, the mid tier, got 20%: $2 and $12, down from $2.50 and $15. Sol, the coding model, stayed at $5 and $30, and picked up a new fast-response mode instead of a discount.

OpenAI credited efficiency gains, optimizations across its training and inference stack, with some reporting putting the serving-cost improvement near 20%. Take the reason at face value and it still leaves the spread unexplained. Efficiency gains apply to the whole stack evenly. Cutting one tier 80% and another 0% is a choice about where to compete, made on top of whatever the engineering saved.

Read the asymmetry

Price cuts flow to where the competition is.

At the cheap end, the competition is everyone. Terra now undercuts Anthropic's Claude Sonnet 4.6, which runs $3 and $15. Luna at $0.20 input sits in the same bracket as open-weight models teams can host themselves and the cheaper tiers from Google. That end of the market is a commodity, and commodities compete on price. So OpenAI cut to the number the market was going to force anyway.

At the top, Sol did not move because nothing is pushing on it. Buyers who need the best coding model do not switch to save a few dollars per million tokens. They need the output to be right. Demand there is thick, substitution is hard, and OpenAI keeps its pricing power. The flat bar in the chart is the tell. The tier OpenAI protected is the one it knows you cannot easily leave.

What it means if you build on the API

Cheaper tokens sound like pure upside. For a while, they are. Your input cost on Luna just dropped 80%. So did everyone else's, on the same day, at the same rate. When every competitor gets the same cost cut on the same morning, the floor moves and your position above it does not.

Two things follow for anyone selling something built on the API.

First, stop pricing your product against token cost. If your margin is the gap between what OpenAI charges you and what you charge a customer for the same tokens, that gap only shrinks. Token prices move one direction. The analysts reading this cut expect the declines to keep coming: cheaper inference pushes pilots into production and makes agentic workflows that were previously uneconomic start to pencil out. That is real demand you can sell into, but only if you charge for the outcome. The margin has to live in the workflow you wrap around the model, the data you bring, and the result you stand behind.

Second, watch which tier you depend on. If your product runs on Luna, you are building on a commodity, and your supplier just proved it will race that tier to the floor. If it runs on Sol, you are paying for pricing power that gets passed straight to you. The per-call economics of an agent already swing from $0.50 to $2.00 a resolution depending on how many model calls it makes and which model it calls. This cut widens that spread. Route the cheap, substitutable work to Luna and reserve Sol for the calls that need it, and your unit cost falls without touching quality.

There is a platform layer reading the cut the same way. Marketplaces that take 3% to 25% of every agent transaction do not care what a token costs. They tax the volume on top of it. Falling token prices grow that volume, so the rake wins either way.

The honest caveat

Two ways this read could be wrong.

The efficiency story might be genuine. If OpenAI really did cut serving costs and is passing them through, the move is cost, not defense, and the tier spread could just reflect where the savings landed. The pricing-power read and the cost-pass-through read both fit the same numbers. What tips it toward pricing power is the timing: three weeks after launch, right as cheaper rivals gained ground, OpenAI cut the tier most exposed to them and left the rest.

And Sol's flat price is not a guarantee. Coding demand looks inelastic today. A strong enough competitor at the top could change that fast, and then the model OpenAI protected this week gets cut next. Pricing power is a snapshot, not a moat.

What to watch now

Two signals. Whether Google or Anthropic answers Luna's $0.20 within the month, and whether Sol ever moves. The first tells you the price war at the bottom is still live. The second tells you the top has stopped being safe. The number to track is not the size of the discount. It is which tier holds its price.

Sources

• InfoWorld, OpenAI drops GPT-5.6 Luna and Terra API prices by up to 80% (July 30, 2026)
• American Bazaar, AI just got cheaper: OpenAI slashes prices by up to 80% (July 31, 2026)
• Cryptopolitan, OpenAI cuts GPT-5.6 Luna price 80% as cheap rivals gain ground
• Yahoo Finance, OpenAI cuts GPT-5.6 Luna and Terra prices by up to 80%
• VentureBeat, AI price wars: OpenAI cuts GPT-5.6 Luna prices by 80% as model competition shifts toward cost

This is not financial advice.