OpenAI cut the API price of its flagship model by 20 percent on November 17, 2026, and paired the cut with a second, cheaper model built for everyday work. The headline reads like a routine discount, but it is better understood as a statement about what AI work actually costs. The unit that matters is not the token. It is the finished task. And on that measure, the cost of serious AI work is falling much faster than the posted price suggests, because the new models finish the same jobs with fewer tokens, and caching stops a business paying twice for text it already sent.
The Published Prices
GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. Its predecessor, GPT-6 Astra, cost $10 for input and $50 for output, with cached input at $1, so the cut is 20 percent on ordinary tokens and 60 percent on cached input. GPT-6.2 Sol keeps the rates of GPT-6.1 Sol: $2 for input, $0.10 for cached input, and $10 for output.
A token is a small chunk of text, roughly a few characters, that a model reads or writes; billing counts millions of them. Cached input is text the system has already processed and can reuse instead of reading again. OpenAI's rule is to keep the front of a prompt unchanged and append new instructions, because a stable prefix makes the cache work.
Why Frontier Prices Keep Falling
The price cut is not an isolated promotion. It is one data point in the fastest cost decline ever measured for a transformative technology. Researchers at Epoch AI estimate that the cost of reaching a given level of AI performance has fallen about 47 percent per quarter since 2023, or about thirteen times cheaper per year. That pace is four times faster than DNA sequencing, six times faster than compute, and eighteen times faster than lithium batteries.
The pattern is not uniform. Prices fall fastest right after a performance level first becomes state of the art, when the lab that reached it can charge a short-lived premium before competitors catch up. Averaging across the main benchmarks, Epoch found costs falling 66 percent per quarter at that debut moment, then slowing to about 32 percent per quarter two years later. Three forces do the work: competition among labs, steady hardware efficiency gains, and algorithmic improvements that let smaller models do what bigger ones used to do.
There is a counterweight. The cost of running frontier models has risen, not fallen, at an estimated three to eighteen times per year, because the models have grown larger and the reasoning they produce has grown longer. Per-token prices fall while the total bill for the most demanding work can climb. That tension is the real economics of AI, and it is why a rate-card price cut is only part of the picture.
The Token Efficiency Effect
OpenAI's claim for both new models is that they finish demanding tasks with fewer output tokens. That matters because a task bill is the number of tokens consumed multiplied by the price per token. If a model writes less to reach the same answer, the task gets cheaper even before the price changes.
A legal technology company reported that Astra used 34.5 percent fewer output tokens and completed more than three times as many legal benchmark tasks as the earlier flagship. A workspace software company said Astra edged out its production model while using about 40 percent fewer tokens and 30 percent fewer review rounds.
The benchmark claims follow the same logic. On Agents' Last Exam, a test of multistep professional work, GPT-6.2 Astra scored 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points above Claude Fable 5.2 at about 5 percent of the cost. On a science benchmark, it scored 9 points higher than its predecessor at less than a third of the cost. Sol came within half a point of Astra on a coding evaluation at about a fifth of the cost, and outscored the earlier flagship on an everyday-tasks evaluation at about a seventh of the cost.
What Caching Does to the Bill
Take a concrete job. Imagine a task that reads 20,000 input tokens and writes 2,000 output tokens. On the old flagship that bill was $0.30: $0.20 for input and $0.10 for output. At the new Astra price the same token counts cost $0.24. Cache the input and the bill drops to about $0.09, because the reused text costs $0.008 instead of $0.16. If the model also writes fewer tokens for the same work, as OpenAI reports on some professional tasks, the output cost falls and the whole task lands closer to $0.07. The same task on Sol costs about $0.06 uncached and about $0.02 with a cached prefix.
That is the arithmetic a business should do. A 20 percent price cut is pleasant; a fourfold reduction from caching plus fewer output tokens is a change in what the work is worth.
Cost per Task, Not Cost per Token
The lesson is that a rate card is not a budget. A cheaper model that needs three times as many tokens can cost more per task than the expensive one, and a premium model that finishes in one pass can be cheaper than a bargain model that wanders. Benchmark leaderboards measure quality on someone else's test set. The question a business should ask is what a completed task costs on its own workload, including the tokens spent on retries and review.
That reframing also settles the routing question. Routine volume, the drafting, summarizing, and classifying that fills a workday, belongs on the cheaper model. The hard cases, where a mistake costs real money or the output must match professional conventions, belong on the flagship. The price gap is now wide enough that splitting the work is a real decision rather than a rounding error.
What a Small Business Should Do
Start with your own task list rather than the announcement. Write down what you actually send to a model this month, sort it by the cost of a wrong answer, and measure the cost of a completed task, not the cost of a token. Move low-stakes volume to the efficient model and track it; keep high-stakes work on the stronger model and track that separately. Design prompts so the stable part stays at the front to unlock the cached-input discount. Then budget on the assumption that prices will keep falling, and treat any long-term price as temporary rather than a planning constant.
The cheapest token is not the goal. The cheapest finished task is.
If you want help working out which of your tasks belong on which model and what your real cost per task is, talk to us at Crowe Endeavors and we will build the measurement with you.
Sources
"AI News Brief – November 18, 2026" (Trend Match Touch (YouTube), November 18, 2026) — https://www.youtube.com/watch?v=4FUUyQuViJA
"OpenAI releases GPT-6.2 Astra and Sol models" (SCWorld, November 18, 2026) — https://www.scworld.com/brief/openai-releases-gpt-6-2-astra-and-sol-models
"Introducing GPT-6.2 Astra and Sol" (OpenAI, November 17, 2026) — https://openai.com/index/introducing-gpt-6-2-astra-sol/