Cheapest model per token can cost 8x more per right answer

AWS ran gpt-5.6-luna, terra and sol against gpt-5.4-mini and nano on Bedrock and measured cost per correct answer instead of cost per token. Anyone picking a model by list price alone is the target here, because the results flip that logic. On AIME math problems, Luna cost $0.0021 for each passing answer. Mini, despite a lower price per token, cost $0.0139 per passing answer, per AWS's benchmark post.

The gap grows with conversation length. On DeepSearchQA, a multi-turn agent benchmark, Luna cost $0.05 per passing answer against Mini's $0.40, while also scoring higher (F1 0.50 versus 0.39).

AWS ties the widening gap to how multi-turn agents work. Every turn re-sends the full conversation history. Token spend then accumulates roughly quadratically as a task runs longer, so a model that looks cheap on a single-turn benchmark can rack up cost fast once it's chaining tool calls or asking follow-up questions.

GDPval, which scores real deliverables, told the same story. Luna passed 56% of tasks at $0.010 each. Mini passed 42% at $0.030 each: lower quality and higher cost per usable output, from the model with the lower list price.

AWS doesn't publish the turn count where the crossover happens, only that the curve bends upward with each additional turn. That's the number worth pulling out of your own agent logs before the next model swap, not the AIME gap. If your workflow chains tool calls or asks clarifying questions, benchmark cost per correct output on your actual trajectories before trusting the per-token sticker price.