跳到主内容
@wquguru
精选88Microsoft Foundry 博客(RSS)模型发布/更新多源精选 ×10

GPT-6.2 Astra与Sol在Microsoft Foundry正式可用

GPT-6.2 Astra and GPT-6.2 Sol now generally available in Microsoft Foundry

原文
发到 X
推荐理由

GPT-6.2系列正式发布且含详细定价与部署选项,直接影响企业选型与成本核算,建议开发者关注缓存配置策略。

Better answers for less: the newest Astra reasoning model delivers higher accuracy with fewer output tokens in Foundry testing, bringing down cost per task.

以更少成本获得更优答案:最新的 Astra 推理模型在 Foundry 测试中实现了更高的准确率,同时输出 token 更少,从而降低了每项任务的成本。

Today, OpenAI's GPT-6.2 Astra and GPT-6.2 Sol are generally available in Microsoft Foundry. Both models build on their predecessors: GPT-6.2 Astra is the next reasoning model in the Astra family, built for the most demanding work, and GPT-6.2 Sol advances the model teams rely on for production agents that run all day, every day. Both are built for where enterprise AI is heading: beyond answering questions and into getting real work done, inside the tools your business already runs on.

今天,OpenAI 的 GPT-6.2 Astra 和 GPT-6.2 Sol 已在 Microsoft Foundry 中全面可用。这两款模型均基于其前身构建:GPT-6.2 Astra 是 Astra 系列的下一代推理模型,专为最具挑战性的工作而设计;GPT-6.2 Sol 则进一步提升了团队赖以支撑全天候运行生产代理的核心模型能力。两者均面向企业 AI 的未来发展方向:超越问答,深入实际工作,在您企业现有运行的工具内部完成真正的业务处理。

Model capability is only part of the story. AI projects often slow down on identity, networking, data handling, governance, and compliance. Microsoft Foundry brings those fundamentals together in Azure, so you can take the newest GPT models from prototype to production with speed and trust.

模型能力只是故事的一部分。AI 项目常在身份验证、网络、数据处理、治理和合规方面遇到瓶颈。Microsoft Foundry 将这些基础要素整合于 Azure 之中,使您能够以速度和信任度将最新的 GPT 模型从原型快速推向生产环境。

What's new in GPT-6.2 Astra and GPT-6.2 Sol

GPT-6.2 Astra 和 GPT-6.2 Sol 的新特性

GPT-6.2 Astra is an upgrade to GPT-6 Astra, designed to improve reasoning, writing, coding, computer use, and the creation of documents, slides, and spreadsheets. It also does more with less: in Foundry benchmark testing, it achieved higher average accuracy than GPT-6 Astra with substantially fewer output tokens, delivering better results at a lower cost per task for the same per-token price.

GPT-6.2 Astra 是对 GPT-6 Astra 的升级,旨在提升推理、写作、编码、计算机使用以及文档、幻灯片和电子表格的创建能力。它同样做到“以更少实现更多”:在 Foundry 基准测试中,它以显著更少的输出 token 实现了比 GPT-6 Astra 更高的平均准确率,在相同单 token 价格下,以更低的每项任务成本交付更优结果。

Put together, those upgrades let GPT-6.2 Astra take on bigger pieces of work. Give it an open-ended goal and it can map out a plan, weigh the trade-offs, and follow through across your applications with tool use and computer use. What comes back is ready for review: documents, slides, spreadsheets, and analyses built to your templates and standards.

综合这些升级,GPT-6.2 Astra 能够承担更庞大的工作任务。赋予它一个开放式目标,它便能规划方案、权衡利弊,并通过工具和计算机使用功能跨应用执行到底。返回的结果已准备好供审查:包括按照您的模板和标准生成的文档、幻灯片、电子表格和分析报告。

Built on GPT-6.1, GPT-6.2 Sol is designed for more concise technical writing, more focused coding and debugging, and better decisions about when and how to use tools. It also brings improved front-end design when building websites and a more natural interaction style.

基于 GPT-6.1 构建,GPT-6.2 Sol 旨在提供更简洁的技术写作、更聚焦的编码与调试,以及更明智的工具使用时机与方式决策。它在构建网站时还能带来改进的前端设计体验,并采用更自然的交互风格。

Both models accept text and image inputs, produce text output, and work with the Foundry APIs and tools you already use.

两款模型均接受文本和图像输入,生成文本输出,并与您已使用的 Foundry API 和工具兼容。

The right intelligence behind every agent

为每位代理提供恰到好处的智能支持

The right model for each job should be determined through evaluations. Start with GPT-6.2 Astra for complex reasoning and high-stakes decisions, and evaluate GPT-6.2 Sol for production agents, coding workflows, and high-frequency tasks where quality and cost both matter. Look beyond price per token to cost per task, alongside quality and reliability. Foundry brings evaluation, tracing, and monitoring together so teams can make that call with evidence, then switch models without re-platforming.

每个任务应通过评估来确定最合适的模型。从 GPT-6.2 Astra 开始,用于复杂推理和高危决策;并评估 GPT-6.2 Sol,适用于生产环境中的智能体、编码工作流以及质量和成本都至关重要的高频任务。不要只看每 token 的价格,还要看每项任务的总成本,同时兼顾质量与可靠性。Foundry 将评估、追踪和监控功能整合在一起,使团队能够基于证据做出决策,然后在不更换平台的情况下切换模型。

Do more with every token

让每个 token 发挥更大价值

Cost per task depends on token efficiency on both sides of a request. GPT-6.2 Astra already uses fewer output tokens, and prompt caching brings that efficiency to input. Both models support 12-hour cache writes on Standard deployments, alongside the 30-minute default, so agents that go quiet overnight or between sessions can still reuse cached instructions, documents, and tool definitions at discounted cache-read rates instead of writing them again. Learn more about 12-hour cached writes here.

每项任务的成本取决于请求两端(输入和输出)的 token 效率。GPT-6.2 Astra 已经减少了输出 token 的使用量,而提示缓存(prompt caching)则将这种效率延伸至输入端。两种模型在标准部署中均支持 12 小时缓存写入,同时也保留默认的 30 分钟设置,因此那些在夜间或会话之间处于空闲状态的智能体,仍可以折扣的缓存读取费率复用已缓存的指令、文档和工具定义,而无需重新写入。在此了解更多关于 12 小时缓存写入的信息。

The 12-hour option is opt-in. Storing content in the cache for the first time is billed at a write rate, and a 12-hour write costs more than a 30-minute one, while later reuse is billed at the lower cache-read rate. Choose 12 hours when the same prefix would otherwise expire and be stored again several times a day, and keep the 30-minute default where steady traffic already keeps the cache warm. To opt in, set prompt_cache_options.ttl to "12h" in a Responses API request:

12 小时选项为可选配置。首次将内容存储到缓存中时,按写入费率计费,且 12 小时写入的费用高于 30 分钟写入,但后续的重复使用则按较低的缓存读取费率计费。当相同的提示前缀若不选择 12 小时就会过期,导致每天被多次重新存储时,请选择 12 小时;而在流量稳定、缓存始终保持活跃的场景下,请保留 30 分钟的默认设置。要启用此选项,请在 Responses API 请求中将 prompt_cache_options.ttl 设置为 "12h":

{

"model": "<your GPT-6.2 deployment>",

"model": "<你的 GPT-6.2 部署实例>",

"prompt_cache_key": "support-assistant-v1",

"prompt_cache_options": { "mode": "explicit", "ttl": "12h" },

"input": [

{ "role": "system", "content": [{

"type": "input_text",

"text": "<1,024+ tokens of reusable instructions>",

"text": "<1,024+ 个可复用的指令 token>",

"prompt_cache_breakpoint": { "mode": "explicit" } }] },

{ "role": "user", "content": "<variable user input>" }

{ "role": "user", "content": "<可变的用户输入>" }

]

}

Smarter token use is one lever on cost per task. The other is where your agents run, matching each workload to the capacity, residency, and pricing model it needs.

更智能地使用 token 是降低每项任务成本的一个杠杆。另一个杠杆在于智能体的运行位置,即将每种工作负载与其所需的容量、数据驻留地和定价模型相匹配。

Deploy where your data needs to be

在数据需要的位置进行部署

At launch, both models are available in three deployment types:

在发布之初,这两种模型均提供三种部署类型:

  • Global Standard: the broadest capacity and efficient throughput across all 28 supported Azure regions, billed by usage.
  • Data Zone Standard: data processing stays within the selected data zone (US, EU, or APAC), helping organizations meet data residency and compliance requirements without giving up access to the latest frontier models.
  • Global Provisioned Throughput: dedicated capacity, reserved in provisioned throughput units (PTUs), for high-volume workloads that need predictable latency and cost.
  • 全球标准版(Global Standard):在所有 28 个支持的 Azure 区域提供广泛的容量和高效的吞吐量,按使用量计费。
  • 数据区标准版(Data Zone Standard):数据处理保持在所选的数据区内(美国、欧盟或亚太),帮助组织在不妨碍访问最新前沿模型的同时,满足数据驻留和合规要求。
  • 全局预置吞吐量:专用容量,以预置吞吐量单位(PTU)预留,适用于需要可预测延迟和成本的高负载工作负载。

Same model, same code, only the deployment changes. When residency, scale, or performance requirements shift, you change the deployment type, not the architecture.

相同的模型,相同的代码,仅部署方式发生变化。当驻留、扩展或性能要求发生变化时,您只需更改部署类型,而无需更改架构。

GPT-6.2 Astra Pricing

GPT-6.2 Astra 定价

ModelDeploymentContext LengthPricing (USD $/million tokens)
InputCached Input30m Cached Writes12h Cached WritesOutput
GPT-6.2 AstraGlobal StandardShort context$8.00$0.40$10.00$20.00$40.00
Long context$16.00$0.80$20.00$40.00$60.00
Data Zone Standard (US)Short context$8.80$0.44$11.00$22.00$44.00
Long context$17.60$0.88$22.00$44.00$66.00
Data Zone Standard (EU)Short context$9.60$0.48$12.00$24.00$48.00
Long context$19.20$0.96$24.00$48.00$72.00
Data Zone Standard (APAC)Short context$9.60$0.48$12.00$24.00$48.00
Long context$19.20$0.96$24.00$48.00$72.00
模型部署上下文长度定价(美元/百万代币)
输入缓存输入30分钟缓存写入12小时缓存写入输出
GPT-6.2 Astra全球标准短上下文$8.00$0.40$10.00$20.00$40.00
长上下文$16.00$0.80$20.00$40.00$60.00
数据区域标准(美国)短上下文$8.80$0.44$11.00$22.00$44.00
长上下文$17.60$0.88$22.00$44.00$66.00
数据区域标准(欧盟)短上下文$9.60$0.48$12.00$24.00$48.00
长上下文$19.20$0.96$24.00$48.00$72.00
数据区域标准(亚太)短上下文$9.60$0.48$12.00$24.00$48.00
长上下文$19.20$0.96$24.00$48.00$72.00

GPT-6.2 Sol Pricing

GPT-6.2 Sol 定价

ModelDeploymentContext LengthPricing (USD $/million tokens)
InputCached Input30m Cached Writes12h Cached WritesOutput
GPT-6.2 SolGlobal StandardShort context$2.00$0.10$2.50$5.00$10.00
Long context$4.00$0.20$5.00$10.00$15.00
Data Zone Standard (US)Short context$2.20$0.11$2.75$5.50$11.00
Long context$4.40$0.22$5.50$11.00$16.50
Data Zone Standard (EU)Short context$2.40$0.12$3.00$6.00$12.00
Long context$4.80$0.24$6.00$12.00$18.00
Data Zone Standard (APAC)Short context$2.40$0.12$3.00$6.00$12.00
Long context$4.80$0.24$6.00$12.00$18.00
模型部署上下文长度定价(美元/百万代币)
输入缓存输入30分钟缓存写入12小时缓存写入输出
GPT-6.2 Sol全球标准短上下文$2.00$0.10$2.50$5.00$10.00
长上下文$4.00$0.20$5.00$10.00$15.00
数据区域标准(美国)短上下文$2.20$0.11$2.75$5.50$11.00
长上下文$4.40$0.22$5.50$11.00$16.50
数据区域标准(欧盟)短上下文$2.40$0.12$3.00$6.00$12.00
长上下文$4.80$0.24$6.00$12.00$18.00
数据区域标准(亚太)短上下文$2.40$0.12$3.00$6.00$12.00
长上下文$4.80$0.24$6.00$12.00$18.00

Prices above are for Standard deployments, in USD per million tokens. GPT-6.2 Astra and GPT-6.2 Sol are also available on Global Provisioned Throughput at $1 per PTU per hour. To learn more, visit the Azure OpenAI pricing page.

上述价格为标准部署的价格,单位为美元/百万代币。GPT-6.2 Astra 和 GPT-6.2 Sol 也可通过全局预置吞吐量获取,价格为每小时每 PTU 1 美元。欲了解更多信息,请访问 Azure OpenAI 定价页面。

Build safer agents on Foundry

在 Foundry 上构建更安全的智能体

GPT-6.2 Astra and GPT-6.2 Sol run with the same layered protections as the rest of the GPT-6 series in Foundry:

GPT-6.2 Astra 和 GPT-6.2 Sol 在 Foundry 中与 GPT-6 系列的其他模型一样,运行着相同的多层防护机制:

  • In the model: alignment and safety training.
  • On prompts and outputs: content filters and guardrails.
  • On tool calls and responses: prompt injection mitigations.
  • 模型层面:对齐与安全训练。
  • 提示词与输出层面:内容过滤器和安全护栏。
  • 工具调用与响应层面:提示注入缓解措施。

Security extends to the platform underneath, so your data stays your data:

安全性延伸至底层平台,确保您的数据始终属于您:

  • Identity and access: Microsoft Entra identity and access controls.
  • Network: isolation through private endpoints.
  • Encryption: in transit and at rest.
  • Data privacy: prompts and outputs are not used to train the models.
  • Compliance: workloads inherit Azure's extensive compliance portfolio.
  • Governance: Microsoft Purview applies data policies, with human checkpoints where the workflow needs them.
  • 身份与访问控制:Microsoft Entra 身份与访问控制。
  • 网络:通过私有端点实现隔离。
  • 加密:传输中和静态数据均进行加密。
  • 数据隐私:提示词和输出结果不会用于模型训练。
  • 合规性:工作负载继承 Azure 广泛的合规组合。
  • 治理:Microsoft Purview 应用数据策略,并在工作流程需要时设置人工检查点。

These controls make it easier to put safeguards in place and keep people in the loop, but no set of controls removes risk entirely. Each organization remains responsible for choosing the controls that fit its own use cases and regulatory requirements.

这些控制措施有助于更容易地建立保障措施并保持人员参与,但没有任何一套控制措施能完全消除风险。每个组织仍需负责选择适合其自身用例和监管要求的控制措施。

Start building

开始构建

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

关联讨论

同一事件的更多信源

相似阅读

关联信息,但可能不是同一事件