On September 22, 2026, OpenAI and Anthropic announced price cuts on the same day. OpenAI released GPT-6 Sol and Luna. The former was priced at $2/$10 per million input/output tokens, half the $4/$20 of GPT-5.6 Sol during the promotional period; Luna launched at $0.10/$0.50, also halved. Anthropic introduced Claude Opus 5.5 the same day; according to the company, under typical workloads its total cost of use fell about 40% compared with Opus 5, while output speed increased by more than 30%.
80%: An Experiment Testing the Limits of Elasticity
On July 30, 2026, OpenAI cut the API price of GPT-5.6 Luna by 80%, from $1/$6 to $0.20/$1.20 (input/output per million tokens).
OpenAI CFO Sarah Friar disclosed at the Goldman Sachs Communacopia technology conference that Luna usage grew about 10x. Enterprise customers are demanding quantifiable returns on AI investments, and OpenAI is also exploring pricing models tied to business outcomes.
The 10x growth shows that when prices fall below a tipping point, demand elasticity changes qualitatively. Use cases previously shelved because costs were too high enter a feasible range after the price cut.
Cache Reads: The Ignored 96%
In typical workloads such as coding agents and multi-turn task agents, cache reads account for 96% to 99% of total token spending, with cache hit rates usually between 90% and 100%. Cutting cache-read prices alone has a far greater impact on actual usage costs than the headline reduction suggests.
Anthropic released Fable 5.1 on September 1, 2026, pressing the cache-read price from $1/MTok to $0.25/MTok, a 75% reduction; other pricing was unchanged from the previous-generation Fable 5. According to calculations by minimaxir.com, for coding tasks run in agent mode, total cost fell by 70%–75%.
Anthropic's Claude Opus 5.5 continues this approach: input/output prices were each cut by 20%, and cache reads fell from $0.50/MTok to $0.20/MTok, a 60% reduction. Combined, the two bring typical workload costs down by about 40%.
The Ghost of DeepSeek
According to Axios, DeepSeek's new-generation low-cost coding model released in August 2026 pushed per-task unit prices for some tasks down to a few cents. When OpenAI announced the 80% price cut for GPT-5.6 Luna, its official cost-efficiency comparison chart showed that without the cut, Luna would lose to DeepSeek V4 Pro on cost-effectiveness.
A more fundamental driver is the maturation of scale economies at U.S. labs. As inference infrastructure is optimized, marginal inference costs fall; price cuts both respond to competition and pass cost benefits on to developers.
According to data cited by edwardconard.com, an 80% annual decline in LLM API prices means labs need usage growth of more than 400% to keep revenue unchanged. OpenAI's 10x usage growth proves this path works at the Luna tier.
Flagship Models: The Market Signal Has Changed
GPT-6 Astra (launched September 3) is priced at $10/$50/MTok, comparable to Fable 5.1 and the previous-generation Fable 5, and did not follow mid- and low-end models in cutting prices. OpenAI's GPT-5.6 Sol only offered a staged discount of about 20% on August 21, marked "valid for three months."
According to payment data cited by edwardconard.com (covering 70,000 companies), spending on Anthropic's most expensive flagship model held steady at about 11% of total spending, with growth flattening. Anthropic claims Opus 5.5 performs "close to Fable 5.1" on most tasks while priced at less than half of the latter, directly capturing migration demand.
Holding the line on flagship pricing maintains brand premium and signals R&D investment; on the other hand, enterprise users consuming flagship compute are less price-sensitive and care more about stability and the capability ceiling.
"Cost per Task": The Next Pricing Framework
When launching GPT-6 Sol/Luna, OpenAI disclosed "cost-per-task" data alongside benchmark results, rather than only giving token unit prices.
According to TechWire Asia, research published by GitHub in September pointed out that merely reducing the number of tokens in a single model call does not necessarily lower the total cost of completing a coding task, because agent systems may add extra calls to make up for the compressed context. In a Copilot experiment, removing non-essential information from model inputs reduced inference costs by only about 5% in offline agent benchmarks.
Token unit prices are no longer sufficient to describe the real economics of AI inference. Developers need to measure "how much it costs to complete a business objective." OpenAI's introduction of this framework, on the one hand, prevents unfavorable pure price comparisons; on the other, it foreshadows pricing models evolving from usage-based billing to outcome-based billing—the "pricing experiments tied to business outcomes" mentioned by Sarah Friar are an early sign of this trend.
Independent Judgment
This round of price cuts is a systemic reset of the industry's pricing benchmark.
Within three months, OpenAI's mid-tier model fell from $1/MTok to $0.10/MTok, a 90% cut; Anthropic's flagship model saw total cost of use fall 40%, with even larger cuts in cache-intensive scenarios. These changes have moved beyond "fine-tuning pricing" into "redefining the market's expected price band."
For developers, the window of opportunity lies in agent use cases previously shelved because costs were unviable—especially coding agents and document-processing agents that require long context and multiple iterations—where the economics are now fundamentally different. The selection logic shifts from "which model is strongest" to "which model is most cost-effective for a specific task."
For labs, 10x usage growth proves elasticity exists, but scaling volume from Luna's low-price entry tier and generating revenue from the flagship tier to support tens of billions of dollars in capital expenditure are fundamentally different things. Whether the price-and-usage flywheel can reach the flagship tier is the key test of sustainability.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接