OpenAI cuts prices on new GPT-6 Sol and Luna models

Sam Altman wrote on X on Tuesday that the two models are "also half the price per token, and even less per task." ^[1] The second half is where a bill moves. OpenAI's AutomationBench figure for Sol at extra-high effort is $0.27 a task, on a test covering business workflows across 47 tools ^[10]. That $0.27 buys about 27,000 Sol output tokens, or 135,000 input tokens ^[3]. OpenAI puts Claude Opus 5 at maximum effort on the same test at 11.1 times that cost, roughly $3.00 a task, scoring 26.9% against Sol's 33.2% ^[11][1].

The comparison that decides a roadmap is inside OpenAI's own line-up. Sol costs 20 times Luna per token, on input and on output ^[4]. On the DeepSWE software engineering test, Sol scored 68.8% and Luna 66.6% ^[14]. OpenAI says Luna's result is comparable to Opus 5 and Fable 5 at medium effort, at 93% and 96% less cost ^[15].

OpenAI's own table argues for setting effort before picking tier: Sol at extra-high effort scored 33.2% on AutomationBench against Astra at low effort on 30.3%, and Astra at that setting cost 3.9 times as much per task ^[12][7].

At its best measured setting on the agent test, Sol did not complete about two-thirds of the workflows ^[8]. Read the score as a completion rate and a finished task costs about $0.81, three times the headline figure ^[9][11].

For a product with a long fixed prefix, the caching change may matter more than the token price. The discount brings Sol's cached input down to about $0.20 per million ^[10]. Developers can set explicit breakpoints to choose where the cached prefix ends, and can change reasoning effort or switch tools on and off without losing the cached context ^[8]. GitHub told OpenAI that the changes cut the share of prompt tokens needing fresh processing by more than half, measured over recent months across billions of requests ^[9].

OpenAI attaches its own caveats to the competitor numbers. Rival scores came from published reports rather than in-house runs, and Fable 5 results stood in where Fable 5.1 figures were unavailable ^[17]. The company also says its Fable 5.1 chart point understates the real cost, because it leaves out the Opus 5 fallbacks that fired on about 40% of tasks ^[18]. Its factuality claim, that Sol makes roughly half as many mistakes as GPT-5.6 Sol, rests on de-identified ChatGPT conversations in which a user flagged an earlier model's error ^[19], and OpenAI says those conversations were chosen for being error-prone and do not represent ordinary use ^[20].

Every cost comparison in the announcement sets one hosted model against another ^[12]. Anthropic shipped Claude Opus 5.5 the same afternoon at around 40% less to run than Opus 5 ^[22]. Dianne Penn, who heads product management, research and labs at Anthropic, told CNBC the company keeps working on making the model's thinking more efficient, so it uses fewer tokens depending on the effort setting ^[23].

Two questions sort the calls in a product. Does a test or a person check the output before a user acts on it, and does the prompt prefix hold steady across calls? Checked output with a stable prefix is Luna work, and the cache discount compounds the saving. Unchecked output with a prefix rebuilt per request is where the Sol rate gets paid twice over. Then measure cost per completed task on your own traffic; OpenAI's own benchmark puts Sol's at about $0.81 ^[9].