The AI price war: OpenAI cuts GPT-5.6 and turns the frontier into a perishable asset

THE CUT, IN CONTEXT. On July 30, 2026, OpenAI trimmed the prices of its GPT-5.6 family. Terra, the mid-range model, went from $2.5 to $2 per million input tokens and from $15 to $12 for output (a token is the smallest unit of text the model processes, roughly three-quarters of an English word). That is a 20% cut. Luna, the budget model, dropped 80%, to $0.20/$1.20. Sol, the flagship, did not move: it stays at $5/$30. OpenAI attributes this to two things: it serves each query at 20% lower cost and generates each answer 15% more efficiently. Translation: it is not giving away margin. It is passing on savings it already had. The cut is real, but it is defensive.

THE NUMBER THAT MATTERS IS NOT THE TOKEN. Here is our first reading. Comparing models by token price is like comparing cars by the price of a liter of gasoline. What a company pays for is not a million tokens: it is a finished task. A support ticket resolved. An invoice extracted. A pull request reviewed. And the same task consumes very different amounts of tokens depending on the model. "Agentic" workflows—where the model plans, uses tools, reads the result, and retries in a loop—burn 5 to 30 times more tokens than a simple chatbot question, according to DoiT. In coding, the gap can reach a thousandfold. That is why the number that matters is cost per solved task. On the Intelligence Index for August 2026, GPT-5.6 Sol costs $1.04 per task, almost half of Claude Opus 5 ($2.03) and less than Fable 5 ($2.75). That is the ground OpenAI wants to fight on, and it is no accident.

CHEAP CAN GET EXPENSIVE. But cost per task hides a trap. A model that is cheaper per token can end up more expensive per job done right. The key is reliability. If a model is correct less often, you have to retry. And if it makes mistakes someone must fix, that fix costs money. DoiT illustrates this with a customer-support agent: a model at $0.22 per attempt with a 58% success rate ended up costing 79% more per correct task than one at $0.30 with a 64% success rate. The reason? The cheap one left more errors to clean up. Our reading: the list price is the easy part of the equation. The expensive part is reliability. And there, the quality frontier still rules. In our live comparison (zendoric.com/models), Fable 5 leads on quality with 90 points, ahead of GPT-5.6 Sol (79) and the best Chinese models, GLM-5.2 (77) and Qwen3.7-Max (73).

CHINA SETS THE PRICE FLOOR. OpenAI is not cutting in a vacuum. It is cutting with China pushing the floor down. Chinese labs lowered their API prices six times in the first half of 2026, and three of those cuts were declared permanent, according to a DEV community tally. DeepSeek V4-Pro charges $0.435 for input and $0.87 for output, with the cache—repeated answers that don't need recomputing—almost free, at $0.0036. Moonshot's Kimi holds a $0.07 cache floor. Here we should attribute, not swallow the narrative: according to KraneShares, DeepSeek sustains a 60-80% gross margin at those prices, comparable to Anthropic's 2028 targets. In other words, it is not selling at a loss to bury the rival. It is a different cost structure. The pressure is real and it is here to stay. What we must not confuse is "cheap" with "given away at a loss."

THE FRONTIER IS A PERISHABLE ASSET. This brings us to the underlying thesis, which ties into our Kimi K3 deep dive. Inference—the act of running the model to get an answer—is becoming a raw material. The price of a token no longer reflects "intelligence": it reflects hardware, KraneShares argues. The data backs this up. GPT-4 launched in March 2023 at $30/$60; by 2026 there are lightweight models below $0.40 for output, a drop of about 99.7% in three years, and the venture firm a16z speaks of costs falling roughly tenfold a year. Every frontier model is now a perishable asset: it leads for a few months, then a rival—often open-weight—matches it on the common tasks and sinks its price. Kimi K3 showed it: it launched at $3/$15, but with a 90% cache-hit rate its effective input drops to about $0.30. The question is no longer which model is the smartest today. It is what remains when they are all "smart enough" for the routine. Two things, per the analysis: the hardest 10% of tasks, where quality still commands a premium, and the residue a deployment accumulates through use—its evaluation sets, its error library, its workflows. The model depreciates. That residue does not.

OUR LONG-TERM READING. We close with the implications, without euphoria and without doom. In the short term there is real strain. The cut compresses margins across the whole industry and squeezes startups that resell inference with little added value: if your product is just an API call, your price floor is now set by DeepSeek. And inference this cheap also makes fraud and disinformation cheaper at scale; we do not lose sight of that. But the underlying balance is positive, and it is worth saying calmly. When running a capable model costs almost nothing, applications get built that yesterday did not pay off: a tutor for every student, a first medical screening for those without a doctor nearby, tools for people who never learned to code. The price war is not the end of the AI business. It is the beginning of its diffusion. Value stops living in the model and moves to what we do with it. And that, in the long run, is good news.