Falling AI prices do not tell a single economic story. In an October 25 analysis, data scientist Max Woolf argues that competition from open-weight models and independent inference providers constrains what the largest laboratories can charge. His account links a sequence of price reductions to the availability of alternatives, while questioning claims that today’s low prices must eventually reverse.
This is a newly published interpretation of earlier changes, not a new round of price cuts announced this weekend. The underlying vendor disclosures support the reductions; they do not establish every explanation for why the companies made them.
Several Different Reductions Share One Headline
OpenAI’s July 30 announcement cut GPT-5.6 Luna pricing by 80 percent and Terra pricing by 20 percent. Luna’s published API rates became $0.20 per million input tokens and $1.20 per million output tokens. OpenAI attributed the changes to efficiency improvements across models, serving systems and workflows. That attribution is the company’s explanation, not a public audit of its costs.
A separate, earlier launch announcement for GPT-6.1 Sol listed $2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens. These are offers for different models, not interchangeable measurements of one service becoming cheaper.
That distinction is important when comparing the cumulative headlines. A lower-priced successor, a discount to an existing model and cheaper reuse of stored context can all reduce spending, but they operate through different parts of a customer’s workload. None automatically produces the same saving for every application.
Token Prices Are Not the Whole Task Budget
Anthropic’s September 22 Opus 5.5 release illustrates the distinction. It listed input and output prices 20 percent below Opus 5, at $4 and $20 per million tokens, and a cached-read price of $0.20, down from $0.50. The company said typical workloads cost 40 percent less at default settings, combining price changes with greater token efficiency. That larger reduction is a vendor workload claim, not a universal discount.
For a deliberately simplified example, one million input tokens, one million output tokens and one million cached-read tokens would cost $24.20 at those Opus 5.5 rates, versus $30.50 at the listed Opus 5 rates. Our arithmetic gives a reduction of roughly 21 percent, before other charges or differences in how many tokens a task needs. It is an illustration, not a representative production workload.
The practical measurement is therefore spending for an accepted result. A comparison that omits retries, different output lengths or failed tasks may show a cheaper unit while missing the cost of completing the job. Conversely, a model that finishes with fewer calls can improve task economics beyond its posted token discount.
More Suppliers Do Not Reveal Their Margins
There is a concrete mechanism behind Woolf’s competition argument. OpenRouter’s provider-routing documentation allows customers to prioritize price, latency or throughput when selecting among providers. Its default routing balances requests across leading providers for availability; explicitly sorting by price changes that preference. The documentation also provides controls for fallbacks and supported parameters.
Those options make supplier comparison operational rather than purely theoretical. They do not mean every endpoint supplies identical performance, supports every setting or meets the same operational requirements. Changing a routing preference is not evidence that switching has no cost for an organization.
OpenRouter also documents provider-specific quantization choices and warns that reduced precision can change performance on some prompts. Consequently, comparing offers for the same named open-weight model still requires checking the serving configuration, rather than assuming a shared model label makes every endpoint equivalent.
Woolf’s explanation remains an economic argument, not proof of laboratory profitability or a guarantee of future pricing. Published prices and routing options are observable. The private cost of serving a model, the commercial terms of large contracts and the durability of particular margins are separate questions. The price decline is documented; its eventual floor is not.










