AI Skills AI技能 4h ago Updated 3h ago 更新于 3小时前 49

Is Grok 4.5 Really More Token Efficient Than Claude Opus 4.8? I Checked the Numbers Grok 4.5真的比Claude Opus 4.8更节省Token吗?我核实了数据

Grok 4.5 demonstrates significantly higher token efficiency than Claude Opus 4.8, using approximately 4.2x fewer output tokens on SWE-Bench Pro tasks. Independent verification by Artificial Analysis confirms Grok 4.5 uses over 60% fewer tokens on general capability benchmarks and drastically fewer tokens in agentic coding loops compared to competitors. The combination of lower per-token pricing and superior efficiency results in Grok 4.5 being roughly 17x cheaper per task for software engineerin Grok 4.5 宣称在保持与 Claude Opus 4.8 相当能力的同时,显著提高了代币效率,独立第三方验证证实了这一说法。 在 SWE-Bench Pro 基准测试中,Grok 4.5 的输出代币量约为 Opus 4.8 的四分之一(4.2倍差异),结合更低的价格,单任务成本降低约17倍。 独立机构 Artificial Analysis 的测试显示,Grok 4.5 在通用智能指数上比 Opus 4.8 少使用60%以上的输出代币,且在编码代理任务中总消耗远低于竞品。 Grok 4.5 的高效率部分归因于其独特的训练数据,包括来自 Cursor 的真实开发者会话数据、调试痕迹和多文件

75
Hot 热度
70
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Grok 4.5 demonstrates significantly higher token efficiency than Claude Opus 4.8, using approximately 4.2x fewer output tokens on SWE-Bench Pro tasks.
  • Independent verification by Artificial Analysis confirms Grok 4.5 uses over 60% fewer tokens on general capability benchmarks and drastically fewer tokens in agentic coding loops compared to competitors.
  • The combination of lower per-token pricing and superior efficiency results in Grok 4.5 being roughly 17x cheaper per task for software engineering workloads compared to Opus 4.8.
  • This efficiency is attributed to unique training methodologies involving real-time developer session data from Cursor, teaching the model to prioritize direct action over verbose explanation.

Why It Matters

Token efficiency is becoming a critical economic factor in AI deployment, directly impacting operational costs for high-volume tasks like agentic coding. This case study validates that "brevity as a feature" is a viable strategy for reducing inference costs without sacrificing capability, offering a concrete alternative to the prevailing trend of increasingly verbose model outputs.

Technical Details

  • Benchmark Performance: On SWE-Bench Pro, Grok 4.5 averaged ~15,954 output tokens per task versus ~67,020 for Claude Opus 4.8. Artificial Analysis recorded ~14,000 tokens for Grok on its Intelligence Index.
  • Agentic Efficiency: In multi-step coding agent evaluations, Grok 4.5 consumed ~1.9 million total tokens, significantly less than Anthropic's Fable 5 (~7.2 million) and OpenAI's leading model (~6.2 million).
  • Pricing Structure: Grok 4.5 is priced at $2/M input and $6/M output tokens, whereas Claude Opus 4.8 is priced at $5/M input and $25/M output, compounding the cost advantage.
  • Training Data: The model was trained on real developer session data from Cursor, including debugging traces and iterative code edits, rather than solely on static public code repositories.

Industry Insight

  • Cost Optimization Strategy: Organizations should evaluate models not just on raw benchmark scores but on token consumption rates, particularly for agentic workflows where token counts can spiral.
  • Training Paradigm Shift: The success of Grok 4.5 suggests that training on dynamic, process-oriented data (like live coding sessions) may yield more efficient and practical models than static dataset training alone.
  • Vendor Transparency: The availability of independent verification for marketing claims sets a new standard for credibility in AI launches, encouraging buyers to demand measurable efficiency metrics alongside performance benchmarks.

TL;DR

  • Grok 4.5 宣称在保持与 Claude Opus 4.8 相当能力的同时,显著提高了代币效率,独立第三方验证证实了这一说法。
  • 在 SWE-Bench Pro 基准测试中,Grok 4.5 的输出代币量约为 Opus 4.8 的四分之一(4.2倍差异),结合更低的价格,单任务成本降低约17倍。
  • 独立机构 Artificial Analysis 的测试显示,Grok 4.5 在通用智能指数上比 Opus 4.8 少使用60%以上的输出代币,且在编码代理任务中总消耗远低于竞品。
  • Grok 4.5 的高效率部分归因于其独特的训练数据,包括来自 Cursor 的真实开发者会话数据、调试痕迹和多文件代码编辑记录。
  • 这种“简洁即价值”的特性使得 Grok 4.5 在处理需要多步循环的代理编程任务时,能显著降低账单成本,具有极高的商业吸引力。

为什么值得看

对于AI从业者和企业决策者而言,这篇文章揭示了从单纯追求模型能力上限转向关注“单位任务成本”的关键趋势,特别是在编程和代理工作流中,代币效率直接决定了规模化应用的可行性。它提供了具体的独立验证数据,打破了厂商营销的迷雾,证明了通过特定领域数据(如真实开发会话)微调可以显著提升模型的实用性和经济性。

技术解析

  • 基准测试数据:在 SWE-Bench Pro 上,Grok 4.5 平均每个任务消耗约 15,954 个输出代币,而 Claude Opus 4.8 消耗约 67,020 个,差距达 4.2 倍。
  • 定价与成本优势:Grok 4.5 输入/输出价格为 $2/$6 per million tokens,Opus 4.8 为 $5/$25。结合效率优势,执行相同任务的总成本 Grok 4.5 约为 Opus 4.8 的 1/17。
  • 独立验证:Artificial Analysis 的 Intelligence Index 显示 Grok 4.5 每任务输出约 14,000 代币,比 Opus 4.8 少 60% 以上;在 Coding Agent Index 中,Grok 4.5 总消耗 1.9M 代币,远低于 Fable 5 (7.2M) 和 OpenAI 领先模型 (6.2M)。
  • 训练数据创新:Grok 4.5 不仅使用静态公共代码库,还融入了来自 Cursor 的真实开发者会话数据,包括调试追踪、多文件编辑和修正过程,使其学习到了更高效、更直接的交互模式。

行业启示

  • 代币效率成为核心竞争力:在模型能力趋同的背景下,降低推理成本和提升资源利用率将成为区分模型商业价值的关键指标,企业应优先评估模型在长链路任务中的实际代币消耗。
  • 真实世界数据的重要性:利用包含人类纠错、迭代过程的动态交互数据进行训练,能显著提升模型在特定垂直领域(如编程)的效率和实用性,这为后续模型训练提供了新的数据策略方向。
  • 代理工作流的成本优化:对于依赖多步自我调用的 AI 代理应用,选择高代币效率的模型不仅能降低单次任务成本,更能避免因代币累积导致的账单失控,是构建大规模自动化系统的前提。

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。

LLM 大模型 Inference 推理 Evaluation 评测