GPT-5.6 arrived earlier this month with three pricing tiers, a killed-off Codex, and a 7.8% score on ARC-AGI-3 that beat the previous state of the art by a factor of five. The benchmark number got headlines. The pricing tiers are what will matter in retrospect.
When a technology starts being sold in tiers named after planetary bodies and lunar concepts, you are watching commoditization happen in real time. Sol is the premium tier for people who need the smartest possible answer. Terra is the workhorse tier for everyday coding and analysis. Luna is the cheap tier where quality degrades but the price drops so far that it changes the calculus of whether to use AI at all.
The interesting thing is not the spread between these tiers. The interesting thing is that people are actually running the comparison. A developer I know switched from Codex (which he loved) to GPT-5.6 Luna tier last week. His code completion quality dropped about 10 percent. His monthly cost dropped 90 percent. He made the decision in five minutes and says he hasn't regretted it once. This is what commoditization looks like as a felt experience: not a trend to be analyzed, but a procurement decision that you feel slightly dumb for not having made sooner.
A few months ago, the conversation about AI was dominated by capability comparisons. Which model scores highest on MATH? Which one passes the most SWE-bench tests? The fixation on benchmarks was a symptom of a market that had not yet found its price equilibrium. People were still trying to figure out what the technology could do, because the answer was different every week.
GPT-5.6 changed that. The benchmark conversation has not stopped - ARC-AGI-3 is genuinely impressive - but the center of gravity has shifted to cost. On forums, on social media, in engineering standups, the question is no longer "which model is best?" It is "which tier makes sense for this workload?" The shift from capability question to cost question is the inflection point of commoditization. Once people stop benchmarking and start budgeting, the technology has crossed a threshold.
OpenAI killed Codex as a standalone product because specialized models no longer make economic sense when a general model at Terra tier is good enough for most coding tasks and Luna tier is cheap enough for the rest. The specialist intelligence market is being commoditized by generalists that are cheap enough to waste. When you can dump entire codebases into context at Luna pricing, the skill of prompt craftsmanship - the delicate art of saying exactly the right thing to extract maximum quality from a stingy model - becomes a hobby rather than a professional necessity.
The prompt engineers I know are feeling this most acutely. They spent the last two years developing a skill that looked like a superpower: the ability to coax brilliant answers out of models that were smart but brittle. Now the models are robust enough that careful prompting and sloppy prompting converge on the same acceptable answer. The mystique is gone. The arcane knowledge has been devalued. There is a real psychological dimension to watching your hard-won expertise become irrelevant not because you were wrong but because the technology moved past the regime where your skill mattered.
This has happened before. The AWS precedent is instructive: when cloud infrastructure became cheap and reliable enough, the value chain shifted from owning servers to building on top of them. The people who were good at managing physical racks found themselves competing with a credit card form. The people who built applications on top of the infrastructure captured the value. The same pattern is playing out now. The model itself is becoming like a cloud instance - interchangeable, price-comparable, something you provision rather than something you tune.
Edge commoditization reinforces this. GLM 5.2 can run frontier-grade models on a 16GB laptop, further shifting the question from capability to deployment cost. When a model can run on consumer hardware, the question is no longer "how smart is it" but "where can it run." Intelligence becomes a resource constraint rather than a capability question.
The real signal of commoditization is not when people compare prices. It is when people stop talking about the model entirely. Nobody blogs about TCP/IP. Nobody gets excited about the specific brand of electricity flowing into their house. Infrastructure becomes invisible when it becomes cheap and reliable enough. If the trajectory holds, we are approaching a world where the model name fades into the background and the conversation shifts entirely to what is being built on top of commoditized intelligence.
The moat after the model is not a better model. It is proprietary data that APIs cannot reach, domain expertise that cannot be reconstructed from public training data, distribution that makes switching costly even when the underlying intelligence is fungible. The defensibility moves up the stack to the application layer, the data layer, the relationship layer. The model itself becomes a cost center to be minimized rather than a capability to be maximized.
This is not good or bad. It is a structural transition that happens to every transformative technology eventually. Electricity was magical until it was infrastructure. The internet was a frontier until it was a utility. Intelligence will follow the same path. The question is not whether commoditization is happening. It is what you build on the other side of it.