A 1.5x per-token price still bought a 25 percent cheaper correct answer in AWS's benchmark

Cost per correct answer is two numbers multiplied: the price per token, and the billed tokens spent to get one right answer, wrong attempts included ^[6]. In the AIME run, luna sat at roughly 1.5 times mini's list price per token and came out 25 percent cheaper per correct answer ^[10]. Divide 0.75 by 1.5. Luna used about half as many billed tokens per correct answer as mini did in that sample ^[18]. The reason is in the setup: the Bedrock models ran with reasoning disabled, and the OpenAI API baselines ran at their defaults ^[4].

That confound is disclosed, not buried. "This is a comparison of practical deployment configurations, not a controlled estimate of intrinsic model capability," AWS wrote ^[5]. Disabling reasoning is a legitimate deployment choice, and it is also the cheapest way to win a token-count comparison. For the 25 percent to mean anything on your workload, luna with reasoning off has to be accurate enough on your tasks, and you have to be running mini at its defaults today.

Then the list price moved. The July 30, 2026 reduction took luna down 80 percent and terra down 20 percent on Amazon Bedrock ^[11]. Stacked on the earlier gap, luna's cost per correct AIME answer comes to about 15 percent of mini's ^[19]. The recorded figure is $0.0021 for luna ^[12]. If mini's pricing did not change, that puts mini near $0.014 per correct answer in the same sample ^[20].

The harness itself is careful work. One Responses API client drives all five models, with the backend and model ID swapped while the evaluation logic is held constant ^[2]. Grading combines deterministic checks with a judge model, gpt-5.5, which is not one of the five under test, and the judge prompts are frozen with their hashes written into every result file ^[13]. Each run writes a timestamped JSON, and every number and chart in the post is generated from those files at build time ^[14].

The agentic claim is argued from how turns bill: every turn re-sends the growing conversation, so turn count multiplies the invoice ^[17]. Multi-turn trajectories on live web-research tasks were measured alongside the single-call benchmarks and the rubric-graded deliverables ^[16]. The figure carrying the ranking inversion is a single-call mathematics benchmark ^[10]. A per-turn curve depends on your tool loop, your context management and how often your agent re-fetches a document it already read. The post asks for that directly: reproduce the evaluation on your own workload before selecting a model ^[15].

The retry estimate needs the same treatment. At mini's AIME accuracy, independent attempts would average about 2.7 tries per success ^[8]. "Real retries are correlated, so measure your actual retry strategy rather than assuming that estimate transfers directly," AWS wrote ^[9].