Amazon Bedrock’s pricing for the GPT‑5.6 Luna and Terra models dropped dramatically in July 2026, and an open‑source benchmarking harness now shows that, when measured by cost per correct answer, Luna can be an order of magnitude cheaper than the OpenAI‑hosted GPT‑5.4 mini baseline. Practitioners need to move beyond the per‑token sticker price and evaluate models on the actual outcomes their workloads require.
Outcome‑Based Cost vs Token Price
The benchmark divides total spend (including both right and wrong attempts) by the number of correct answers to estimate the observed cost of a successful response. Across three academic‑level tests—AIME, GPQA Diamond, and MMLU‑Pro—Luna consistently delivered the lowest cost per correct answer. After the July price cut, Luna’s cost per correct AIME answer was recorded at $0.0021, compared with $0.0139 for the OpenAI mini model. Even the nano baseline, whose nominal token price is slightly lower, was out‑performed by Luna when token efficiency and reasoning‑disabled configuration are factored in.
Capability tiers remain evident: the Sol model solved 75 % of AIME problems versus 37 % for mini, and led GPQA Diamond (68 % vs 43 %) and MMLU‑Pro (82 % vs 59 %). Higher accuracy reduces the number of retries needed, directly influencing the per‑outcome cost.
Turn Count and Multi‑Turn Agents
For agentic workloads that involve multiple back‑and‑forth exchanges, each turn adds to the token bill. The harness measured live web‑research tasks and found that turn count can dominate total spend, especially when models generate verbose intermediate steps. Disabling reasoning on Bedrock models reduced token consumption per turn, contributing to Luna’s cost advantage.
Practical Deployment Configurations
The comparison used a single code path—the OpenAI Responses API—while swapping only the backend endpoint and model identifier. Bedrock runs were executed with reasoning disabled, whereas the OpenAI API baselines used default settings. This reflects real‑world deployment choices rather than a controlled capability test. Practitioners should therefore treat the results as a guide for configuration tuning: disabling optional reasoning can lower token usage without necessarily sacrificing correctness for many tasks.
Result files record pricing assumptions ($0.22/$1.32 per 1 M input/output tokens for Luna, $2.20/$13.20 for Terra). Because Bedrock pricing varies by inference tier and region, teams must verify the applicable rates before finalizing cost models.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
- Incorporate outcome‑based cost metrics (cost per correct answer, cost per successful turn) into model selection criteria rather than relying on raw token price.
- Run the open‑source harness—or a similar internal benchmark—against your own task set to validate that Luna’s token efficiency and pricing translate to your workload.
- Consider disabling optional reasoning features when they are not required, as this can reduce token consumption and overall spend.
- Instrument token usage and turn count in your monitoring stack to detect cost drift as model behavior or pricing changes.
- Stay aware of Bedrock pricing updates and region‑specific tiers; even an 80 % price cut can shift the cost‑effectiveness balance dramatically.
