Cheap tokens, expensive decisions: what an 80% AI price cut in three weeks really means
OpenAI and AWS cut GPT 5.6 Luna prices by 80% just 21 days after launch. The lesson is not that AI is cheap now: it is that your architecture cannot afford to marry any single model.
On July 30, OpenAI and AWS cut the price of GPT 5.6 Luna by 80%: from $1.00 and $6.00 per million input and output tokens down to $0.20 and $1.20. Terra dropped 20%. The cut landed 21 days after the GPT 5.6 family launched, and on Amazon Bedrock it applied automatically, with no action required from customers.
An 80% cut in three weeks is not a promotion. It is a signal about how this market now works, and it has direct consequences for any company building with AI, or deciding not to.
Why this is happening
Two forces are at play, and neither is temporary.
The first is competition. According to OpenRouter data reported by industry press, since February 2026 US companies have been routing over 30% of their weekly tokens to Chinese models, which are substantially cheaper. Add to that the fact that promotional pricing from other Western providers was already competing head to head with Terra. When a customer can switch models by editing one line of configuration, pricing stops being a marketing decision and becomes a survival decision.
The second force is more interesting: the cost of serving these models is falling because of software, not hardware. OpenAI claims that GPT 5.6 Sol autonomously rewrote parts of its own inference stack (GPU kernels and the speculative decoding model), reporting a 20% reduction in serving costs. That is a vendor claim and deserves the usual skepticism, but the direction matches what has been happening for two years: every few months, serving the same level of capability costs less.
The practical conclusion: there is no reason to believe today’s prices are a floor. The sensible assumption is that the capability you buy today at price X will cost a fraction of that within a year.
What this changes if you are building with AI
In most projects we see, cost per token stopped being the real bottleneck a while ago, and this cut confirms it. At $0.20 per million input tokens, processing a 50 page document costs fractions of a cent. Classifying an email, even less. The math that justified shelving a use case a year ago (“running every ticket through the model is too expensive”) probably no longer holds.
There is a less comfortable side to this: once cost is no longer the excuse, the hard work is exposed. Integrating the model with your systems, defining what an acceptable answer looks like, measuring quality continuously, and deciding what happens when the model gets it wrong. None of that gets cheaper when tokens do. Quite the opposite: that is where all the value, and all the risk, of an AI project now concentrates.
Three practical decisions
First: treat the model as configuration, not architecture. If switching providers or model versions forces you to touch business logic, you have a design problem. A thin abstraction layer over the model API (or a service like Bedrock, which exposes several providers behind one surface) lets you capture every price drop or quality jump with a configuration change. Teams that hard coupled to a specific model in 2024 have been paying that toll for two years.
Second: reopen the use cases you shelved on cost grounds. Any cost estimate older than six months is obsolete. The usual candidates: classifying and routing emails or tickets, extracting data from documents, generating internal summaries, drafting first responses in customer support. These are high volume workloads with reasonable error tolerance, exactly the profile that fast models like Luna are built for.
Third: do not sign long commitments at today’s prices. Reserving capacity or locking in annual consumption contracts at fixed rates makes sense for stable infrastructure. In a market where list prices fall 80% in three weeks, a long commitment is a bet against the trend. Keep contracts short and flexibility high, at least while the curve keeps behaving like this.
One caveat so this does not tip into carelessness: cheap tokens do not mean design stops mattering. Prompt caching, picking the small model when it is enough, and keeping context windows tight still separate a reasonable bill from an absurd one once volume grows. The difference is that these optimizations now come after a use case has proven its value, not before.
Where we stand
At Evoltix we design our clients’ AI integrations under one premise: the model is the component you will replace most often over the life of the system. So we put the effort into what stays: the integration with the client’s systems, the data that feeds the model and, above all, continuous quality evaluation. A test suite that tells you within hours whether the new model (or the same model, now cheaper) still meets the bar is what turns every price cut into margin instead of risk.
The question is no longer whether your company can afford to use AI. It is whether your architecture lets you benefit from it getting cheaper every quarter.
Sources: AWS pricing announcement, AWS Weekly Roundup, August 3, 2026, Tech Times analysis of the Luna price cut.