GPT-5.6 Sol Is Half Price Right Now, on Exactly Two Gateways, Until September 18. Do Not Rearchitect Around a 30-Day Promo.
GPT-5.6 Sol went to half price on OpenRouter and Vercel AI Gateway between August 17 and 18. Input fell from $5 to $2.50 per million tokens, output from $30 to $15. The discount runs through September 18, 2026, and applies automatically, so if you route through either gateway your bill already dropped without you touching anything.
Five days ago I wrote about how I would route between frontier models given that Grok 4.6 matched Sol at roughly half the price. That gap just closed from the other direction. It closed temporarily, on two specific platforms, and I am not changing a single line of routing logic because of it.
The exact shape of the deal
The numbers are worth stating precisely because "50% off" is doing a lot of hiding.
It is 50% off both input and output, which is unusual. Most promotional pricing discounts input only, since input is where the volume sits and output is where the margin sits. Cutting output from $30 to $15 is the expensive half of this.
It applies with no code change. That is the friendly-sounding detail that makes it a trap. There is no flag to set, no new model ID, nothing in your codebase that records that you are on promotional pricing. The change is invisible at the call site, which means on September 18 the reversal will also be invisible at the call site.
And it is confined to two platforms. Not the OpenAI API directly. Not Azure. OpenRouter and Vercel AI Gateway.
Why "which two platforms" is the whole story
OpenRouter and Vercel AI Gateway are not just resellers. They publish usage data. OpenRouter in particular has become the thing people screenshot when they want to argue about which model is winning, because it is one of the few sources of cross-vendor traffic share that anyone can look at.
Analysts have raised the possibility that shifting perceived market share, rather than acquiring customers, is the actual objective here. I want to be careful with that: it is a suggested motive from commentators, not something OpenAI has stated and not something anyone has demonstrated. Treat it as a hypothesis.
But you can evaluate the hypothesis on its own terms. If your goal were customer acquisition, you would discount where the customers are, which is the direct API. If your goal were to move a public number, you would discount precisely where that number gets computed. The promotion is confined to exactly the two platforms in the second category. That is at minimum a striking coincidence, and it is enough for me to treat the price as marketing spend rather than as information about what inference actually costs.
The distinction matters because those two things decay differently. A real price cut reflects improved economics and tends to persist or go further. A marketing spend reverses on schedule.
What five months of routing bills taught me
I have been routing across three model families since roughly March. The thing I got wrong early was treating price as a first-class routing input.
It is not. Price is the most volatile variable in the entire stack and the least predictive of whether a given call succeeds. In my own logs, the difference between a well-matched model and a poorly-matched one on the same task is frequently a 3x to 5x swing in total tokens consumed, because the wrong model needs more retries, produces longer preambles, and fails validation more often. A 50% list-price discount does not survive contact with a model that needs two attempts.
So my routing decisions are made on task fit and observed failure rate, and price enters as a tiebreak. That ordering means a temporary discount changes nothing, which is exactly what you want from a routing policy. If a 30-day promo can flip your architecture, your architecture was already coupled to something it should not care about.
There is a second-order problem too. Migrating traffic to a gateway has switching costs that are real but back-loaded: you learn its rate-limit behaviour, its error semantics, its failover quirks, its billing granularity. None of that is visible in month one. Doing that work to capture a discount that expires inside the learning period is how you end up on a platform you did not evaluate properly, paying full price.
What I'd actually do
Three things, none of them structural.
Take the money. If you already route through OpenRouter or Vercel AI Gateway, you are getting it automatically and there is no reason to refuse. Do not move traffic there just to capture it.
Put September 18 in your calendar, with the current and post-promo rates written in the entry. Because the discount is invisible at the call site, the reversal will show up as an unexplained doubling on a bill four weeks from now, and you want a note from your past self rather than an afternoon of confusion.
Write down your current cost per completed task, not per token. That is the number that actually moves your margin, and it is the one that will tell you whether the promo mattered at all. My guess for most solo operators is that it barely registers, because token spend is rarely the binding constraint. If half off your frontier model materially changes your unit economics, that is worth knowing about your business independently of anything OpenAI does.
Where this take could be wrong
If you run genuinely high-volume inference, high enough that model spend is a top-three line item, then a month at half price on your most expensive model is real money and my "it barely registers" line does not apply to you. Capturing it might justify the migration work even accounting for the switching cost, especially if you were already considering a gateway.
I would also concede that the market-share hypothesis, which I find plausible, is not required for my conclusion. Even if this is a straightforward customer-acquisition promo with no cleverness behind it, it still expires on September 18 and it still should not drive an architecture decision. The motive is interesting. The expiry date is what matters.
And there is a version of the future where this becomes permanent, prices continue falling across the board, and people who moved early look prescient. I do not think a time-boxed discount on two gateways is evidence for that, but I have been wrong about the direction of inference pricing before.
Author
Lukas
@lukcombinator