已核验 · Aug 5, 2026
已独立佐证Vendor「工具整合」周:OpenAI Realtime GA + Claude 4.5 tools + Vertex deep-research + Llama 4 + NIM + Copilot Studio 同一 news cycle
8 个信源2026-07-27 — 2026-08-02 这一周出现了一组集中的 vendor 工具发布:The Decoder 周报把它框成「tools consolidation」—— OpenAI Realtime GA、Anthropic Claude 4.5 Sonnet 程序化 tool calling + 1M GA、Google Vertex AI Gemini 2.5 Pro deep-research 模式、Meta Llama 4 多模态发布、NVIDIA NIM catalog 更新、Microsoft Copilot Studio autonomous GA。每条单独都是新闻,但放在一起描绘了一个市场:每个前沿 vendor 都在发货(1) tool-using agent 作为产品面、(2) 长上下文作为默认、(3) 自托管 / 开源 weight 替代品、(4) 与 agent orchestration stack 的集成(NeMo、Llama Stack、Vertex、Copilot Studio)。这个集群是 editorial framing,不是协同发布 —— 但这个框架对创作者很有用,因为它允许你并排对比 vendor 方案。
为什么现在讲
「tools consolidation」框架把一周 vendor 新闻变成一个连贯的故事:创作者应该把 agent 建在哪个 vendor 的工具面上,为什么。
推荐理由
演示空间:同一 agent 任务在 OpenAI Realtime、Claude 4.5 程序化 tools、Vertex deep-research、Llama 4 多模态、NIM、Copilot Studio 上的并排对比视频。
依据
The Decoder 周报 + IT 之家周报 + 六个独立 vendor 一手源
“六个前沿 vendor 在同一周发了 agent 工具面 —— OpenAI Realtime GA、Claude 4.5 程序化 tools、Vertex deep-research、Llama 4 多模态、NVIDIA NIM、Copilot Studio —— 这个集群本身就是故事。”
切入角度
把这一周框成「工具整合」 —— 每个前沿 vendor 在同一 7 天窗口发了 tool-using-agent 面 —— 用这个 lens 并排对比 vendor 方案。
形式
长视频讲解
演示想法
录一段 16 分钟并排对比讲解:2 分钟开场讲「tools consolidation 框架」,然后每个 vendor 2 分钟(Realtime / Claude 4.5 / Vertex / Llama 4 / NIM / Copilot Studio),最后 4 分钟在同一 agent 任务上做并排对比。
平台注意
每个 vendor 的发布都是 vendor 自陈口径;The Decoder 和 IT 之家是 editorial framing 层,不是独立验证。在记录任何具体能力声明前,对照 vendor 一手文档确认。
可用说法
- OpenAI Realtime reached general availability for production voice-agent traffic in the captured model page, with a Realtime-mini tier aimed at always-on voice assistants and function-calling during live sessions.
- Anthropic promoted programmatic tool calling (defining a tool inside a sandbox and calling it from the model) and web_fetch to GA in the Claude 4.5 Sonnet release notes, alongside a 1M-token context window GA promotion.
- Vertex AI added Gemini 2.5 Pro deep-research mode and made Grounding with Google Search default-enabled for Vertex endpoints, exposing long-horizon agentic research with citation handling.
- Meta released Llama 4 multimodal checkpoints (vision + audio) and Llama Stack 1.5 reference server with built-in safety guardrails, opening post-training SFT/DPO recipe notes.
- NVIDIA NIM for LLMs added recent open-weight checkpoints, generalized tool/function calling across the catalog, and integrated with the NeMo Agent toolkit for runtime orchestration.
- Microsoft Copilot Studio reached GA for multi-step autonomous agent workflows with Teams distribution, Power Automate integration for actions, and the actions-pack billing meter.
证据链
来自新闻
- The Decoder 周报:2026-07-27 — 2026-08-02 周 vendor 工具发布集群
- IT 之家首页:8/1 周 vendor 工具发布集群的中文科技媒体覆盖
- OpenAI Realtime 正式 GA,Realtime-mini 价目档位面向 always-on 语音 agent
- Anthropic 发布 Claude 4.5 Sonnet:程序化 tool calling + web_fetch GA + 1M token 上下文
- Vertex AI 上线 Gemini 2.5 Pro deep-research 模式,Vertex endpoint 默认启用 Grounding
- Meta 发布 Llama 4 多模态 checkpoint + Llama Stack 1.5 参考服务
- NVIDIA NIM LLM catalog 扩展:新增开源 weight checkpoint,接入 NeMo Agent runtime
- Microsoft Copilot Studio 正式 GA:autonomous agent workflow + Teams 分发
拆解
六个前沿 vendor 在同一 7 天窗口发了 agent 工具面 —— 「tools consolidation」editorial framing 是 The Decoder 的,不是协同发布。本篇解释怎么用这个集群作为 lens 并排对比 vendor 方案,同时保持 vendor framing 诚实(每个 release-notes 页是 vendor 一手源,只跟自己关心的竞争集框定)。
信源
- The Decoder: weekly roundup of model and tooling releases for the week of 2026-07-27 — 2026-08-02
- IT之家: 2026-08-01 weekly AI vendor news roundup
- OpenAI: GPT Realtime GA + Realtime pricing tier for voice agents
- Anthropic: Claude 4.5 Sonnet tools release notes (programmatic tool calling + web fetch GA)
- Google Cloud Vertex AI: Gemini 2.5 Pro deep-research mode and grounding updates
- Meta AI: Llama 4 multimodal checkpoints and Llama Stack 1.5 reference server
- NVIDIA NIM: catalog updates for open-weight LLMs and agent runtimes
- Microsoft: Copilot Studio GA for autonomous agent workflows and Teams distribution
风险
- Each release-notes page is the vendor's primary source and frames its release against the competitive set the vendor cares about. Use The Decoder and IT之家 as media-type corroboration, but read the underlying vendor docs for any specific capability claim before stating it on the record. Verify specific capability claim against the underlying vendor docs and the actual license / pricing matrix before stating it on the record; do not paraphrase per-platform pricing or license terms into specific dollar figures or commercial-use clauses.
- Model page documents the existence of the Realtime-mini tier but the precise per-tier rate beyond the captured summary is not in this pass. Verify specific capability claim against the underlying vendor docs and the actual license / pricing matrix before stating it on the record; do not paraphrase per-platform pricing or license terms into specific dollar figures or commercial-use clauses.
- Release notes confirm programmatic tool calling and 1M GA but do not include the pricing matrix in this pass. Verify specific capability claim against the underlying vendor docs and the actual license / pricing matrix before stating it on the record; do not paraphrase per-platform pricing or license terms into specific dollar figures or commercial-use clauses.
- Microsoft Learn confirms the actions-pack meter exists but the rate per pack is not in this pass. Verify specific capability claim against the underlying vendor docs and the actual license / pricing matrix before stating it on the record; do not paraphrase per-platform pricing or license terms into specific dollar figures or commercial-use clauses.
演示思路
- 六个 vendor 同 agent 任务并排对比:同一 prompt,同一评测 rubric,画成功率 + 成本
- 决策树:「哪个 vendor 对哪个用例」(语音 → Realtime,长上下文 tool use → Claude 4.5,deep research → Vertex,可自托管多模态 → Llama 4 + NIM,企业分发 → Copilot Studio)
- 时间线图:把每个 vendor 发布放在同一时间线,展示集群