{"id":2101879,"date":"2026-09-30T16:44:58","date_gmt":"2026-09-30T14:44:58","guid":{"rendered":"https:\/\/www.luxsure.fr\/?p=2101879"},"modified":"2026-10-02T05:54:15","modified_gmt":"2026-10-02T03:54:15","slug":"gpt-6-sol-and-luna","status":"publish","type":"post","link":"https:\/\/www.luxsure.fr\/en\/2026\/09\/30\/gpt-6-sol-and-luna\/","title":{"rendered":"GPT-6 Sol and Luna: what the price cut changes"},"content":{"rendered":"\n

On 22 September, OpenAI announced prices cut by half for Sol and Luna, its everyday working models. The percentage is calculated on promotional prices, and the cost of a processed file remains to be measured.<\/em><\/p>\n\n\n\n

Reading the amounts before the percentage<\/strong><\/p>\n\n\n\n

Per million tokens, Sol goes from 4 to 2 dollars for input and from 20 to 10 dollars for output. Luna goes from 0.20 to 0.10 dollar for input and from 1.20 to 0.50 dollar for output. OpenAI announces \u00ab50% cheaper\u00bb for both. The calculation is correct for Sol and for Luna’s input, but Luna’s output falls by about 58%.<\/p>\n\n\n\n

The starting prices are those of GPT-5.6’s promotional offers, not its list prices, which OpenAI states in its own text. The company attributes the drop to improvements in caching and inference, with a 90% discount on reading inputs already held in memory.<\/p>\n\n\n\n

Since launch, the page carries an update dated 29 September that points to a more recent model, GPT-6.1 Sol. The announcement says nothing more about it in the text I read: it will therefore be necessary to check whether the prices and comparisons in this article still hold for this new model.<\/p>\n\n\n\n

The cost of a task, as measured by the provider<\/strong><\/p>\n\n\n\n

OpenAI prefers to talk about cost per task. On AutomationBench, an enterprise-workflow benchmark, Sol at the \u00abxhigh\u00bb setting scores 33.2% for 0.27 dollar per task, against 26.9% for Claude Opus 5 at maximum, at a cost the company puts at more than eleven times that of Sol. On the DeepSWE benchmark, devoted to software engineering, Sol is presented at 68.8%, a little over one point from Claude Fable 5’s best score.<\/p>\n\n\n\n

These figures come from OpenAI, which chose the tests and the settings compared. It states that it takes competitors’ scores from public reports, and that it used those of Claude Fable 5 when those of version 5.1 were not available. The family’s scores on this enterprise benchmark remain, moreover, modest in absolute terms.<\/p>\n\n\n\n

A tier for each task, to be tested rather than assumed<\/strong><\/p>\n\n\n\n

What follows is the author’s reading. Triage work, such as spotting an incomplete file, does not call for the same spend as a provenance analysis or a text committing the House. One can therefore imagine an economical model for the former and a more demanding model for the others. A model’s name does not prove that it suits the task.<\/p>\n\n\n\n

A team would do well to set what an acceptable result contains before choosing: the indispensable information, the tolerable errors, the cases that go up to a person. The handover matters as much as the pricing. A cheap first pass that erases its sources or masks its doubts passes on a problem that the next step will not see.<\/p>\n\n\n\n

The unit that counts<\/strong><\/p>\n\n\n\n

The token price is an input, and a task’s consumption depends on the level of effort chosen. To decide, ten real files run through three model tiers, with a grid of the errors that matter, would give the useful figure: the cost per file that a person accepts without redoing it. No launch page can supply it.<\/p>\n","protected":false},"excerpt":{"rendered":"

OpenAI is lowering the prices of Sol and Luna. This development invites comparing several tiers of models on specific tasks, while keeping the same level of requirement.<\/p>\n","protected":false},"author":214,"featured_media":2094203,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_lmt_disableupdate":"","_lmt_disable":"","footnotes":""},"categories":[77972],"tags":[81224,57600],"class_list":["post-2101879","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-luxury-and-ai","tag-openai","tag-technology"],"_links":{"self":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts\/2101879","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/users\/214"}],"replies":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/comments?post=2101879"}],"version-history":[{"count":0,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts\/2101879\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/media\/2094203"}],"wp:attachment":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/media?parent=2101879"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/categories?post=2101879"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/tags?post=2101879"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}