{"id":29296,"date":"2026-09-16T21:07:39","date_gmt":"2026-09-16T19:07:39","guid":{"rendered":"https:\/\/whiteduck.de\/token-control-5-1-responses-api-tool-governance-and-more-models\/"},"modified":"2026-09-29T14:40:19","modified_gmt":"2026-09-29T12:40:19","slug":"token-control-5-1-responses-api-tool-governance-and-more-models","status":"publish","type":"post","link":"https:\/\/whiteduck.de\/en\/token-control-5-1-responses-api-tool-governance-and-more-models\/","title":{"rendered":"Token Control 5.1: Responses API, tool governance, and more models"},"content":{"rendered":"

With Token Control 5.1, we are expanding governance for the Responses API<\/a>. The focus is on a changing way of using AI: instead of single responses, multi-step tasks emerge in which models research, use tools, and develop results over longer periods. This is complemented by centralized approvals for these models\u2019 tools, additional models from Azure AI Foundry, and an API key that serves multiple approved deployments. <\/p>\n

\"\"<\/a><\/figure>\n

From a single response to a multi-step task<\/h2>\n

Research starts with a question. However, before producing an answer, the model may search multiple times, evaluate documents, and incorporate additional information. The Responses API supports such workflows, among other things, through Web Search, File Search, and specialized deep-research models. A single request can therefore trigger numerous processing steps. <\/p>\n

Longer-running tasks can also be executed in the background, independent of an application remaining open. The task continues to run, and the application retrieves the result later. Closing the interface does not stop it. <\/p>\n

For operations, this means: it is not only the final answer that matters. The ongoing task also needs a clear assignment, traceable access rights, and a governed way of handling status and cancellation. With Token Control 5.0<\/a>, we made visible in March what AI costs in the enterprise. Token Control 5.1 answers the question that has been added since then: what the models do on the way to the answer. <\/p>\n

\"\"<\/a>
Every tool call visible in the trace<\/figcaption><\/figure>\n

Responses API for modern AI applications and agents<\/h2>\n

Token Control now supports the Responses API for OpenAI and Azure-based deployments, including streaming and\u2014depending on provider and approval\u2014background processing. Your existing cost and access rules therefore also apply to these calls. <\/p>\n

Tasks can be started, retrieved, cancelled, and deleted. Token Control records the assignment of each task. Even a response that is retrieved or cancelled hours later remains assigned to the same API key, the same cost center, and the same access rights as at the start. A task does not lose its context just because no one is watching anymore. <\/p>\n

Token Control serves the endpoints that current SDKs and agent clients expect. If you work with the Microsoft Agent Framework, the .NET SDK, or a Copilot-like client, you do not need to change anything in your configuration.<\/p>\n

All approvals in this release are based on the same structure. A deployment is your configured model instance with a provider. You set approvals at the deployment level and, where more granular control is required, for individual models within it. In addition to Azure OpenAI and Azure AI Foundry, Token Control now also supports OpenAI and Anthropic directly as providers. <\/p>\n

Central approvals for Web Search, File Search, and MCP<\/h2>\n

You can now explicitly approve or block not only models, but also their tools. The rules are inherited along your organizational hierarchy; subordinate units cannot broaden them. You restrict external MCP servers via a URL allowlist. <\/p>\n

Why this matters to us is illustrated by an example we consider underestimated. Web Search in the Azure OpenAI Responses API uses Grounding with Bing Search. Microsoft states in its documentation that the Data Protection Addendum does not apply to data sent to Grounding with Bing Search, and that your data leaves your compliance and geo boundary<\/a>. If you have committed to the EU Data Boundary, you should not have this function enabled without noticing. <\/p>\n

\"\"<\/a>
Approve or block tools individually<\/figcaption><\/figure>\n

Specifically, in Token Control 5.1:<\/p>\n