OpenAI's 272K surcharge erases most of Sol's price advantage over Opus 5.5

The line that decides whether OpenAI's cut reaches your invoice is the long-context surcharge. Requests over 272K input tokens pay 2x on input and cache and 1.5x on output, and the multiplier applies to the whole request ^[10]. Over that threshold Sol's $2 input becomes $4 per million tokens, its $10 output becomes $15, and its cache reads go from $0.20 to $0.40 ^[21]. Opus 5.5 bills $4 in, $20 out and $0.20 cached ^[2][8]. Fresh input is level at that point, and Sol's cache reads cost twice Anthropic's. Batch and Flex run at half price, Fast mode at double ^[11].

The dev.to writeup prices an agent session at 10M input tokens with 8M of them served from cache, plus 1M output: $15.60 on Sol against $29.60 on Opus 5.5, a 47 percent saving ^[9]. The surcharge is per request. Hold that token mix, assume each request in the session crosses 272K, and Sol bills 2M x $4 + 8M x $0.40 + 1M x $15, or $26.20, about 11 percent under Opus 5.5 ^[22]. That comparison assumes Anthropic charges one rate at any context length; the writeup does not report a long-context rule for Opus 5.5 ^[24].

AutomationBench is the row where the two vendors can be set side by side, because both print the same numbers for Opus 5 at 26.9 percent and Fable 5.1 at 31.4 percent ^[7]. Opus 5.5 lands at 40.0 percent, 6.8 points above Sol and just under GPT-6 Astra's 41.4 ^[6]. OpenAI's charts run against Opus 5 and Fable 5/5.1 ^[5], which the writeup puts down to the clock: nobody benchmarks a new model in ninety minutes ^[25]. TechCrunch, cited in the same piece, read the gap between the two launches as a sign of how intense the competition has become ^[17].

OpenAI's pitch is cost per task. On AutomationBench Sol spends $0.27 per task, and OpenAI published multiples of 11.1x for Opus 5 and 3.9x for Astra ^[18], which works out to about $3.00 and $1.05 ^[23]. On DeepSWE Sol lands within 1.1 points of Fable 5's best score for roughly 80 percent less per task, and on OSWorld it matches Opus 5 at medium effort for about 80 percent less ^[19]. Those multiples price Opus 5, the model Anthropic had just repriced from $5 / $25 to $4 / $20 ^[3].

For Sol's numbers to carry, requests have to stay under 272K input tokens and the cache hit rate has to sit near the 80 percent the worked example assumes ^[26]. AutomationBench also has to resemble your task mix. Sol is aimed at coding and agentic work, Luna at high-volume jobs with a clear goal such as summarising documents or extracting fields ^[13]. OpenAI credits caching and inference improvements for the cut and says the GPT-5.6 rates it halved were promotional, so the new list price is half of a discount ^[12].