OpenAI's Mid Tier Now Costs Ten Times Its Cheapest

The line that shipped on Thursday

OpenAI's developer changelog carried one sentence on 30 July: starting that day, GPT-5.6 Luna costs 80 percent less and GPT-5.6 Terra costs 20 percent less. In money, Luna moved from one dollar per million input tokens and six dollars per million output tokens to twenty cents and one dollar twenty. Terra moved from two dollars fifty and fifteen dollars to two dollars and twelve dollars. Sol, the top model in the family, was not touched. The company said on its own account that the lower Luna and Terra prices are also reflected in how usage is counted in Codex and ChatGPT Work, so the same subscription absorbs more work.

The timing is the first thing worth noting. GPT-5.6 reached general availability roughly three weeks earlier. A tier that shipped at one dollar per million input tokens was repriced to twenty cents before most buyers had finished a full billing cycle on it, which tells you something durable about what a launch price is: it is an opening position, not a planning input. Anyone who signed an annual budget off the July launch sheet is now working from a number that survived twenty-one days.

One explanation, and it only fits one of the two cuts

OpenAI gave a reason. It said efficiency work during the development of GPT-5.6, including the model rewriting and optimising its own production serving code, cut the end-to-end cost of serving the model by about 20 percent and improved token generation efficiency by more than 15 percent. Read that against the two cuts and the arithmetic separates cleanly. Terra fell 20 percent. The stated cost saving was 20 percent. That is a pass-through, near enough to exact that it is fair to call Terra's new price a cost-driven price.

Luna fell 80 percent. That is four times the stated efficiency gain, so cost does not account for it and something else must. The obvious candidate is the competitive floor: cheap open-weight models, many of them Chinese, have been pulling the bottom of the market down all year, and a twenty-cent input tier is a defensive position against that floor rather than a reflection of what the model costs to run. The practical consequence for a buyer is that the two prices now have different physics. Terra's is anchored to OpenAI's own cost curve and should drift down slowly and predictably. Luna's is anchored to whatever the cheapest credible competitor is charging this quarter, which means it can move again, in either direction, without a corresponding change in cost.

Run one workload through both tiers

Take a support deflection workload that consumes 500 million input tokens and 100 million output tokens in a month, which is an ordinary volume for a mid-sized product with a busy help desk. Before 30 July, that workload cost 1,100 dollars on Luna and 2,750 dollars on Terra. After 30 July it costs 220 dollars on Luna and 2,200 dollars on Terra. Both bills fell, and every headline about the day is accurate on that point.

Now look at the gap rather than the levels. Before the cut, running that job on Terra instead of Luna cost 1,650 dollars a month extra. After the cut it costs 1,980 dollars a month extra, or about 1,822 euro and 1,564 pounds at prevailing rates. The premium for choosing the wrong tier went up by a fifth on the same day both prices went down, because the ratio between the tiers widened from 2.5 to 1 to exactly 10 to 1 on both input and output. This is the part that does not appear in the announcement, and it inverts the intuition: a price cut made routing discipline more valuable, not less. Teams that put everything on the mid tier for peace of mind were paying a modest insurance premium in June. They are paying a heavy one now.

The tag that reprices itself

The same changelog entry retired Priority Processing and introduced Fast mode, which delivers up to 2.5 times standard speed at twice the price, offered on Sol in the API across the responses and chat completions endpoints. The migration detail is the one to act on: requests already tagged priority will automatically use Fast mode. No code changes, no migration window, no opt-in.

That means a production system built months ago, by someone who may no longer work on it, can start billing at a different rate this week without anyone touching a repository. The upgrade is not unreasonable on its face, since 2.5 times the speed for twice the price is better value per unit of latency than the arrangement it replaces. But value per unit of latency is only the right measure if you actually needed the latency. Batch jobs, overnight enrichment, retry queues and anything else that inherited a priority tag from a copied configuration are now paying a speed premium for speed that no user is waiting on.

What an operator changes this week

Three things, in order of how much they are worth. Grep the codebase for the priority tag and strip it from everything that is not user-facing, because that is money leaving with no counterparty benefit. Then re-run your tier assignment: any workload parked on Terra out of caution now carries a ten-fold premium over Luna rather than a 2.5-fold one, and the honest test is whether you have ever measured a quality difference on your own traffic rather than assumed one. Finally, stop treating the published price as a planning constant, and hold a review whenever a tier moves rather than annually.

European buyers carry one extra exposure worth naming. These prices are set and billed in dollars, so a euro or sterling budget absorbs the currency move on top of the model price, and a cut announced as 80 percent is 80 percent in dollars only. That does not change the direction of any decision here, but it does mean the euro saving on a Luna migration is a range rather than a number, and a finance team that books the headline figure will be reconciling a variance later.

Read next: GPT-5.6 Terra Matches GPT-5.5 at Half the Price  |  The Door Left Open Was the Package Registry