The new GPT-5.6 family: Luna, Terra, Sol 全新的GPT-5.6系列：Luna、Terra和Sol

TL;DR

• OpenAI released the GPT-5.6 family (Luna, Terra, Sol), featuring a 1M token context window and a February 2026 knowledge cutoff.
• The models emphasize long-running agentic performance, with Sol leading on Agents’ Last Exam benchmarks while claiming superior cost-efficiency compared to competitors.
• New API capabilities include Programmatic Tool Calling, native multi-agent support, and explicit prompt cache breakpoints.
• OpenAI contested SWE-Bench Pro validity, estimating 30% of tasks are broken, likely explaining lower scores relative to Anthropic's models.

Why It Matters

This release signals a strategic shift toward optimizing for complex, multi-step agentic workflows rather than just static benchmark scores, directly impacting how developers design autonomous AI systems. The introduction of granular control over reasoning effort and caching mechanisms offers practitioners new levers for balancing latency, cost, and accuracy in production environments. Furthermore, the public dispute over benchmark integrity highlights the growing importance of evaluating models on real-world, long-horizon tasks rather than isolated coding challenges.

Technical Details

• Model Architecture & Specs: Three variants (Luna, Terra, Sol) with a 1,000,000 token context window and 128,000 maximum output tokens. Pricing ranges from $1/$6 (Luna) to $5/$30 (Sol) per 1M input/output tokens.
• Agentic Performance: Optimized for "long-running professional workflows," achieving a score of 53.6 on Agents’ Last Exam, outperforming Claude Fable 5 by 13.1 points.
• API Innovations:
  • Programmatic Tool Calling: Allows JavaScript composition to orchestrate tool calls, bridging gaps between MCPs and terminal sessions.
  • Multi-Agent Support: Native ability to spin up subagents for parallel, focused work within the core API.
  • Prompt Cache Breakpoints: Explicit control over caching locations to optimize costs, moving beyond automatic detection.
  • Image Handling: New detail: original parameter to bypass automatic image resizing.
• Reasoning Control: Supports variable reasoning efforts (none, low, medium, high, xhigh, max), allowing users to trade off cost and performance dynamically.

Industry Insight

Developers should prioritize integrating Programmatic Tool Calling and multi-agent patterns to leverage the new agentic strengths of GPT-5.6, particularly for complex, multi-step professional workflows. The explicit prompt cache breakpoints offer a significant opportunity for cost optimization in high-volume applications, requiring architectural adjustments to define cache boundaries effectively. Additionally, the controversy surrounding SWE-Bench Pro suggests that reliance on single-dataset metrics for model selection is becoming increasingly risky; organizations should adopt broader, workflow-based evaluations to assess true capability.

TL;DR

• OpenAI发布GPT-5.6系列旗舰模型，包含Luna、Terra、Sol三种规格，知识截止至2026年2月16日。
• 模型主打极致效率与长程智能体性能，在Agents’ Last Exam基准测试中大幅超越Claude Fable 5，且成本显著更低。
• 引入多项API新功能，包括程序化工具调用、原生多智能体支持及显式提示词缓存断点。
• OpenAI质疑竞争对手使用的SWE-Bench Pro基准测试有效性，指出约30%任务存在缺陷。

为什么值得看

本文揭示了大模型竞争焦点从单纯的性能指标转向“单位Token效用”与“长期智能体工作流”的效率比拼。对于开发者而言，新的API特性（如原生多智能体和程序化调用）将深刻影响复杂应用架构的设计方式。

技术解析

• 模型规格与定价：GPT-5.6分为Luna ($1/$6)、Terra ($2.50/$15)、Sol ($5/$30)三档，均支持百万Token上下文窗口和128,000最大输出Token。
• 基准测试表现：在评估跨55个领域长程专业工作流的Agents’ Last Exam中，Sol得分53.6，比Claude Fable 5高出13.1分；中小模型以约1/16的成本实现超越。
• API新特性：新增“程序化工具调用”允许模型编写JavaScript编排工具；“多智能体”功能内置子代理并行处理机制；支持手动设置“提示词缓存断点”以优化成本。
• 图像与推理控制：支持detail: original参数避免图片预处理缩放；提供从None到Max的六级推理力度控制，显著影响Token消耗与成本。

行业启示

• 智能体工程成为核心竞争力：模型能力评估标准已从单点任务转向长程、多步骤的专业工作流执行能力，企业应优先关注模型在复杂场景下的稳定性与效率。
• 成本控制策略升级：随着推理Token数量的差异变得关键，简单的每百万Token价格对比已失效，开发者需结合推理力度控制和缓存策略进行精细化的成本优化。
• 基准测试可信度危机：OpenAI公开质疑主流代码基准的有效性，表明行业对标准化评估体系的信任正在动摇，未来可能出现更多去中心化或定制化的评估方案。