Sol's 27-cent benchmark task undercuts Opus 5 by more than eleven times

Luna's two legs fell by different proportions. The input price halved, the output price came down 58%, and a job that reads a million tokens and writes a million now bills 60 cents where it billed $1.40 ^[6][7]. Sol's cut is the clean one: $2 per million input tokens and $10 per million output, against $4 and $20 before ^[4]. That is exactly half on each leg. The same token volume costs half what a plan written on the older rates assumed ^[8].

Per-token prices are comparable across vendors, but the per-task comparisons on the record come from OpenAI. AutomationBench is Zapier's test of whether an agent can carry a business workflow through 47 tools across sales, marketing, operations, support, finance and HR. On it, GPT-6 Sol at its highest reasoning setting passed 33.2% for $0.27 a task ^[9]. Claude Opus 5 at its own top setting passed 26.9% and cost more than 11 times as much per task, by OpenAI's numbers ^[10]. Eleven times 27 cents is $2.97, which works out at about 11 cents for each point of pass rate against Sol's 0.8 cents ^[11].

Anthropic counts its saving in tokens consumed, not in the price of a token ^[31]. Dianne Penn, head of product management, research and labs at Anthropic, told CNBC: "One of the things we're continuing to innovate on is how to make that thinking, how to make the answering more efficient, so it uses less tokens depending on your effort setting" ^[12]. The company said it is "passing these efficiency savings on to our customers in the form of price cuts and rate limit increases" ^[13]. Fortune, setting the two announcements side by side, found OpenAI cheaper per tier ^[26].

Randall Hunt, CTO at the AI consulting company Caylent, told Fortune that "CFOs have seen some of the sticker shock, and they haven't seen some of the gains that were promised in the initial investments." He added: "So they're coming back to their planning for 2027 and beyond, and saying, 'How can we optimize our costs here?'" ^[17] Caylent's method is to time how long employees take to finish tasks with AI tools and then compute a cost per task. OpenAI encouraged that approach in a blog post in mid-July 2026 ^[18].

OpenAI's 50% is measured against a promotion it is currently running on GPT-5.6, so the comparison is to a discounted rate ^[2]. Reasoning effort is an adjustable setting too: turn it up and the model spends more time and computing power double-checking itself, and accuracy and cost rise together ^[27]. A cheaper token at a higher effort setting can still produce a bigger invoice. "OpenAI and Anthropic are engaged in a price war that is driving down the price of AI," Ara Kharazian, lead economist at Ramp, told Fortune, "and therefore driving down their ability to profit from it and grow the price of models" ^[15]. On the bull case, he said: "AI bulls assume that there will be highly performant models that provide more and more value, and therefore they should be more expensive," and that "that is not how normal technology makes it to market" ^[16].

Neither release is a major step up in capabilities ^[20], and OpenAI credits its reductions to "improvements in caching and inference" ^[14]. In my view posted prices per token keep falling from here while total spend per customer rises, because the effort setting is where the volume goes. Anthropic says Sonnet 5.5 and Haiku 5.5 are expected over the coming weeks ^[25]; if either arrives at a higher posted price than the model it replaces, that view is wrong.