AI Practices AI实践 2h ago Updated 44m ago 更新于 44分钟前 44

Beyond the price per token: Choosing the right OpenAI model on Amazon Bedrock for your workload 超越每token价格:为你的工作负载在Amazon Bedrock上选择合适的OpenAI模型

Organizations should evaluate AI models by cost per correct outcome rather than cost per token, as production workloads buy results, not tokens OpenAI's GPT-5.6 Luna on Amazon Bedrock achieved the lowest observed cost per correct answer ($0.0021 on AIME) after an 80% price reduction, outperforming even cheaper-per-token models like nano Agent trajectory costs are dominated by turn count due to quadratic growth in billed input tokens as conversation context accumulates across turns The benchmarki 生产工作负载购买的是"结果"而非"token",需综合评估准确率、token消耗量和智能体对话轮数等隐性成本 通过开源基准测试框架比较了Amazon Bedrock上的GPT-5.6系列(luna/terra/sol)与OpenAI API的gpt-5.4-mini/nano 核心评估维度:单次调用的正确回答成本、多轮智能体轨迹成本、专业交付物质量 Luna模型在2026年7月30日降价80%后成为成本最低选择,AIME任务每正确回答仅需$0.0021,远低于mini的$0.0139 智能体工作负载中每轮对话都会重新发送完整上下文,累积输入成本随轮数近似二次方增长

62
Hot 热度
70
Quality 质量
58
Impact 影响力

Analysis 深度分析

TL;DR

  • Organizations should evaluate AI models by cost per correct outcome rather than cost per token, as production workloads buy results, not tokens
  • OpenAI's GPT-5.6 Luna on Amazon Bedrock achieved the lowest observed cost per correct answer ($0.0021 on AIME) after an 80% price reduction, outperforming even cheaper-per-token models like nano
  • Agent trajectory costs are dominated by turn count due to quadratic growth in billed input tokens as conversation context accumulates across turns
  • The benchmarking harness (openai-on-aws/benchmarks-openai) evaluates models across three dimensions: single-call accuracy, multi-turn agent trajectories, and professional deliverable quality
  • Capability tiers are clearly visible: Sol solves 75% of AIME problems versus 37% for mini, with similar gaps on GPQA Diamond (68% vs 43%) and MMLU-Pro (82% vs 59%)

Why It Matters

This article provides a practical framework for AI practitioners to move beyond simplistic token-based pricing comparisons and evaluate models based on real-world outcome costs, which is critical for production deployments where accuracy, token efficiency, and agent turn counts compound to determine true expenditure. The open-source benchmarking harness enables organizations to reproduce results on their own workloads before making model selection decisions, reducing the risk of costly misallocations.

Technical Details

  • The benchmarking harness evaluates five models (gpt-5.6-luna, gpt-5.6-terra, gpt-5.6-sol on Amazon Bedrock; gpt-5.4-mini and gpt-5.4-nano on OpenAI API) through a single identical code path using the OpenAI Responses API, ensuring consistent evaluation logic
  • Three benchmark categories were used: single-call accuracy on AIME (mathematics), GPQA Diamond (graduate-level science), and MMLU-Pro; multi-turn agent trajectories on DeepSearchQA with live web_search and fetch_page tools; and rubric-graded professional deliverables scored via deterministic checks plus an LLM judge (gpt-5.5)
  • Cost per correct answer was calculated by dividing total spend across all attempts by the number of correct answers, revealing that Luna's token efficiency and reduced billing (with reasoning disabled) made it 25% cheaper per correct answer even before the July 30, 2026 price reduction
  • Agent trajectory measurement used client-managed history with store:false, causing every turn to re-send the full conversation context, resulting in approximately quadratic growth in cumulative billed input tokens relative to turn count
  • Sample sizes ranged from 48–198 items, with results written as timestamped JSON files and frozen prompt hashes recorded for reproducibility

Industry Insight

  • Model selection should be driven by outcome economics rather than sticker price; organizations running cost-optimized models like mini or nano should evaluate whether upgrading to newer Bedrock models delivers sufficient accuracy and token efficiency gains to justify the switch
  • For agentic workloads, minimizing turn count is as critical as minimizing per-turn cost, since context accumulation creates quadratic billing growth—architecting agents for fewer, more decisive turns can dramatically reduce total expenditure
  • The open-source benchmarking approach should be adopted as a standard practice, with teams running the harness against their own task distributions before committing to model procurement decisions, as small performance gaps may be directional rather than statistically significant

TL;DR

  • 生产工作负载购买的是"结果"而非"token",需综合评估准确率、token消耗量和智能体对话轮数等隐性成本
  • 通过开源基准测试框架比较了Amazon Bedrock上的GPT-5.6系列(luna/terra/sol)与OpenAI API的gpt-5.4-mini/nano
  • 核心评估维度:单次调用的正确回答成本、多轮智能体轨迹成本、专业交付物质量
  • Luna模型在2026年7月30日降价80%后成为成本最低选择,AIME任务每正确回答仅需$0.0021,远低于mini的$0.0139
  • 智能体工作负载中每轮对话都会重新发送完整上下文,累积输入成本随轮数近似二次方增长

为什么值得看

这篇文章为AI从业者提供了超越"每token价格"的模型选型方法论,强调以业务结果为导向的成本评估框架。对于正在Amazon Bedrock和OpenAI API之间做选择的团队,提供了可复现的基准测试工具和实际数据支撑。

技术解析

  • 基准测试框架:开源工具openai-on-aws/benchmarks-openai,通过统一的OpenAI Responses API客户端评估5个模型,确保评估逻辑一致,仅切换后端和模型ID
  • 单次调用评估:使用AIME(数学竞赛)、GPQA Diamond(研究生科学)、MMLU-Pro等基准测试,计算"正确回答成本"=总花费÷正确回答数
  • 智能体轨迹评估:使用DeepSearchQA数据集(50题分层采样),通过真实web_searchfetch_page工具进行多步网络研究任务,测量多轮对话成本
  • 质量评估:专业交付物使用rubric评分,结合确定性检查与LLM judge(gpt-5.5,非被测模型),使用冻结提示词并记录哈希值
  • 关键数据:Sol在AIME上准确率达75%(mini为37%),但Luna因token效率更高且降价后成为最经济选择;样本量48-198项,小差距需视为方向性指标

行业启示

  • 模型选型范式转变:从"每token成本"转向"每结果成本",需结合业务可接受的质量门槛综合评估,而非单纯比较定价页数字
  • 智能体成本陷阱:多轮对话场景下上下文线性增长导致输入成本近似二次方增长,减少轮数比降低单轮成本更重要
  • 实践建议:团队应在自身工作负载上复现基准测试,使用开源harness验证模型表现,而非直接依赖公开数据做决策

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

GPT GPT LLM 大模型 Agent Agent Deployment 部署 Inference 推理