Eight LLM Price Moves in Nine Weeks, Up to 80% Each. Stop Prepaying for Tokens and Build the Swap Layer Instead.
Max Woolf (minimaxir) tallied the price changes from OpenAI and Anthropic since the end of July, and the list has eight entries in about nine weeks, with the largest being an 80% cut to GPT-5.6 Luna on July 30. His post reached the Hacker News front page this week with over 400 points. If you signed a prepaid or annual commitment at any point since August, this list is the reason to read your contract again.
The eight moves
All of this is from minimaxir's table, prices per million tokens as input, output and cache read. On July 30, GPT-5.6 Luna dropped 80% from $1, $6 and $0.10 to $0.20, $1.20 and $0.02. On August 10, Anthropic made Claude Sonnet 5's launch price permanent at $2, $10 and $0.20, avoiding a planned increase. On August 21, GPT-5.6 Sol got a cut of about 20% for three months, from $5, $30 and $0.50 to $4, $20 and $0.40.
On September 1, Fable 5.1 cut cache reads 75% versus Fable 5, to $10, $50 and $0.25. On September 22, three things happened at once: GPT-6 Sol at $2, $10 and $0.20, GPT-6 Luna at $0.10, $0.50 and $0.01, and Opus 5.5 at $4, $20 and $0.20, which the author puts at 40 to 60% lower total cost than Opus 5. On September 29, GPT-6.1 Sol halved its cache price to $0.10 at the same $2 and $10 base.
Not every row is a straight cut. The Sonnet 5 item is a price that did not rise, and the Sol cut is temporary. But the direction is clear, and the author says the GPT-5.6 family became 75% cheaper cumulatively by late September.
Why this keeps happening
The author's argument is that nobody cuts prices 80% without outside pressure, and the pressure is competition. Models are easy to swap, switching costs are low, and there are many substitutes, especially open-weight models served by many providers. He points to DeepSeek V4.1 Flash 1017 having 29 hosting providers on OpenRouter. He also argues against the idea that prices must eventually rise to the "true cost of a token," on the reasoning that providers only serve models that are profitable, so prices may keep drifting toward marginal cost.
I am less sure about the second half. He notes himself that the cuts are odd with IPOs coming, since lower revenue per token looks worse in a prospectus. His guess is that inference margins are healthy and efficiency is improving. That is a guess, and nothing in the filings I have read settles it. What I am confident about is the first half: the market is acting like a commodity market, whatever the cause.
What it does to your decisions
Three things change if you accept that prices are falling and noisy. First, prepaying is a bet that prices stay put. A prepaid credit balance at the August Sol price of $5 and $30 looks poor next to $2 and $10 in September. Second, optimizing prompts to squeeze a specific model's cost is a depreciating skill, since the same savings arrive for free every few weeks. Third, the thing worth owning is how fast you can move.
I mean this concretely. The model name should live in one config value. Your prompts should not rely on one vendor's quirks without a test that tells you when you changed behavior. And you should be able to run the same twenty example inputs through a new model on the day it launches and see cost, speed and quality side by side.
Where this could be wrong
Prices could also go up. Introductory discounts end, as the Sol cut's three-month term shows, and the Sonnet 5 note says a planned increase was avoided, which means someone planned one. If you have a volume discount that really is below list price, a commitment can still make sense. I am arguing against prepaying at list price by reflex, not against all negotiated deals.
What I'd actually do
Stop buying tokens in bulk unless the discount is larger than a typical month of price movement, which this autumn means more than 20%. Put the model ID in one place. Build a twenty-input regression file for your main prompts, and set a monthly reminder to run it against the current cheapest model from each of two vendors. If the result is acceptable and cheaper, switch. It takes an hour a month, and it is a better return than any prompt trimming.
Author
Lukas
@lukcombinator