跳转至

GPT-5.6 Sol/Terra/Luna 分层定价,Codex 合并入 ChatGPT,ChatGPT Work 发布

Ch01.1042 GPT-5.6 Sol/Terra/Luna 分层定价,Codex 合并入 ChatGPT,ChatGPT Work 发布

📊 Level ⭐⭐ | 6.0KB | entities/gpt-56-sol-terra-luna-tiered-pricing-codex-merge-2026.md

GPT-5.6 Sol/Terra/Luna 分层定价,Codex 合并入 ChatGPT,ChatGPT Work 发布

OpenAI 于 2026年7月10日正式发布 GPT-5.6 系列,同时将 Codex 整合进 ChatGPT 桌面应用,并推出新智能体工具 ChatGPT Work。三件事同日发生,核心变化不是新模型跑分,而是 OpenAI 将顶级 Agent 能力拆成了可按任务购买的多个价格档位。

模型层级与定价

GPT-5.6 系列分三个能力层级:旗舰 Sol、均衡 Terra 和轻量 Luna。API 标价分别为每百万 token 5 / 2.5 / 1 美元(输入)和 30 / 15 / 6 美元(输出)——Sol 的标价约为 Claude Fable 5 的一半。Sol、Terra、Luna 三档模型加上 medium / high / max / ultra 多级推理深度,同一模型体系下出现多个独立的价格-能力组合。显式缓存断点和多 Agent 并行进一步丰富了定价维度——单次任务的总成本不再由模型名称决定,而是由所选档位组合决定。

Benchmark 表现

启用最高推理强度的 Sol 在 Agents' Last Exam(55 个领域)上取得 53.6%,在 Coding Agent Index v1.1 上取得 80 分,在 Terminal-Bench 2.1 Ultra 模式下达到 91.9%——三个数字均为各自评测的当前最高或并列最高分,且均以低于 Fable 5 的 API 标价实现。SWE-Bench Pro 上 Sol 得分为 64.6%,而 Fable 5 为 80%,差距显著。OpenAI 官方对 SWE-Bench Pro 结果提出异议,声称约 30% 的评测实例存在结构性缺陷,但这一主张目前无独立裁决。

GPT-5.6 在任务路径明确、步骤可拆分的场景(命令行操作、终端测试、浏览器自动化)表现突出,在开放式的仓库级代码修改上出现缺口。Fable 5 则相反——SWE-Bench Pro 80 分一骑绝尘,但 Terminal-Bench 2.1 上被 Sol 甩开近 9 个百分点。

架构变化

Codex 被整合进 ChatGPT 桌面应用,不再作为独立产品存在。ChatGPT Work 作为新的生产力工具推出,与 Cursor Composer 等 AI 编码工具形成竞争关系。顶级 Agent 能力的按任务定价模式可能影响整个 AI 工具定价生态。

Amazon Bedrock Availability — Mantle Inference Engine & Enterprise Features

GPT-5.6 Sol, Terra, and Luna are generally available on Amazon Bedrock, running on Mantle, the next-generation inference engine. Pricing matches OpenAI first-party rates, and usage counts toward existing AWS commitments.

Bedrock Inference Engine

  • Durable state capture: Every request captures its full state continuously; if hardware fails or a node restarts mid-call, the request picks back up where it left off instead of starting over
  • Isolated queue: Each customer gets their own isolated queue with automated capacity management, ensuring predictable performance under heavy load
  • In-Region inference: Requests stay within the specified AWS Region for data residency

Prompt Caching

GPT-5.6 on Bedrock introduces prompt caching with explicit cache breakpoints. Reusable parts of a prompt (system instructions, tool definitions, reference files) are cached. Cached input is billed at a 90% discount and stays available for at least 30 minutes — long enough to cover the burst of calls a single agent run generates.

Security

  • Zero-Operator Access (ZOA): Enforced at the chip level — no AWS operators can access prompts or completions
  • Every model call runs under IAM policies, inside VPC, logged in CloudTrail
  • Data perimeter policies prevent exfiltration across account and network boundaries
  • Classifier-flagged traffic data retained for up to 30 days for automated abuse detection

Regional Availability

  • Sol: US East (N. Virginia), US East (Ohio)
  • Terra and Luna: US East (N. Virginia), US East (Ohio), US West (Oregon)

ChatGPT Work & Codex

Alongside GPT-5.6 on Bedrock, OpenAI launched ChatGPT Work — an agent for larger, multi-step tasks in the ChatGPT desktop app. The updated desktop app brings Chat, Work, and Codex together. Users can configure the app to use GPT-5.6 through the Responses API on Amazon Bedrock.

量子位报道机器之心报道AWS Bedrock GA announcement


关联