GPT-5.6 Luna output drops from $6 to $1.20 per million

OpenAI has cut its mid-tier GPT-5.6 Luna from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens, according to a Financial Times report published by Ars Technica.

At a glance

• Customers pay separately for input tokens, which measure data fed into a model, and output tokens, which measure what it generates, with costs varying by model version and effort setting.
• Anthropic launched Opus 5 at $5 per million input tokens and $25 per million output, half the price of Fable 5, and called off a Sonnet 5 rise due from September.
• Artificial Analysis measured Opus 5 at medium effort as delivering similar performance and cost per task to Moonshot's Kimi K3 at max effort, a comparison drawn across math, science, coding and reasoning benchmarks.

Why it matters. The reductions land on mid-tier models rather than on the top of each line-up, which reads as a decision to compete on volume pricing while holding margin at the flagship level. For buyers, headline token rates are becoming a weaker guide to actual spend, since effort settings and token efficiency can reverse the ranking between two models. Whether flagship pricing stays insulated as mid-tier rates converge with Chinese offerings is likely the open question.

GPT-5.6 Luna's input and output rates both fall to a fifth of their previous levels

Both of OpenAI's new rates for GPT-5.6 Luna work out at a fifth of the previous ones. The latest price cuts from US labs apply to mid-tier products, where they make American models more competitive with Chinese offerings from vendors including DeepSeek and Moonshot.

AI labs sell a range of models at different capabilities and prices, with costs varying further by the version of a model and the effort setting used. Customers are charged for input tokens, which measure the data fed into a model, and for output tokens, which measure what it generates in response.

Most models can also run at different effort settings, which vary the computing power used to answer a question and affect both performance and the eventual cost of completing a task. Headline token prices therefore do not offer a straightforward comparison between models.

Anthropic priced Opus 5 at $5 per million input tokens, half the level of flagship Fable 5

Anthropic launched Opus 5 at $5 per million input tokens and $25 per million output tokens, which the Financial Times reported as half the price of its existing Fable 5 model. Fable 5 is the flagship of Anthropic's line-up.

This week Anthropic called off a planned rise in prices for its Sonnet 5 model, which had been due to take effect from September. A person close to Anthropic said pricing Opus 5 below the flagship Fable 5 was how the startup's "family of models is built, so there's no connection to competitors," according to the Financial Times, which reported the pricing changes at both labs.

Artificial Analysis put GPT-5.6 Luna at max effort at just under twice the cost per task of DeepSeek's V4 Flash

More capable models can sometimes complete a task using fewer tokens or with fewer attempts, so a model that looks more expensive on headline token prices can end up costing less. Artificial Analysis benchmarks models across math, science, coding and reasoning.

Artificial Analysis found Opus 5 at "medium" effort delivered similar performance and cost per task to Moonshot's Kimi K3 at "max" effort. GPT-5.6 Luna at "max" effort performed similarly to DeepSeek's V4 Flash at "max" but cost just under twice as much per task.

Mantas Lukauskas, AI tech lead at the website hosting provider Hostinger, which has used large language models since 2020, said prices for the very best models were "flat to rising" and called the recent changes the "first real test" of whether Anthropic and OpenAI can protect the cost of their most advanced offerings.

The US labs have cut the middle and are defending the top.

What's next. The Artificial Analysis comparisons put cost per completed task, rather than the posted token rate, at the centre of vendor selection as the mid-tier fills with cheaper options from both US and Chinese labs. Lukauskas frames the coming period as a test of whether the most advanced models hold their price while the middle of the market is cut.