OpenAI Models, Now Side by Side with Anthropic

July 2, 2026 · InfraScout Team

InfraScout has always run its assessments on Anthropic's Claude models. As of today you have a second choice: you can point an AI provider connection at OpenAI and run the exact same assessment conversations on OpenAI's GPT-5 models. Anthropic stays the default and nothing changes for your existing connections — this is an option you opt into, one connection at a time. When you do, chat behaves the way you already expect: the AI streams its answer, works through the same InfraScout tools, shows its reasoning, searches the web, and keeps long conversations from overflowing.

Choosing OpenAI on a connection

Provider is a property of each AI provider connection, so you decide model vendor the same place you already manage models. When you configure a connection, you pick whether it runs on Anthropic or OpenAI, and the connection carries that choice from then on. Because it lives on the connection, you can run both side by side: one connection on Anthropic for the work you already trust it with, another on OpenAI for the reviews where you want it — without touching a single global setting.

An OpenAI connection brings its own API key. You enter it in the AI Providers admin view, and InfraScout stores it encrypted. There is no server-wide shared OpenAI key: each connection authenticates with the key you gave it, so different teams or tenants can bring their own OpenAI accounts and keep them isolated from one another.

The GPT-5 models you can pick

An OpenAI connection opens up the GPT-5 reasoning family — GPT-5, GPT-5 mini, and GPT-5 nano — alongside the GPT-5.4 family. The larger models are built for the deep, multi-step reviews where sustained reasoning pays off; the mini and nano tiers are the faster, lighter options for routine checks and quick questions, the same way you already choose between heavier and lighter Claude models.

Cost estimation works for every one of them. Just as with Anthropic models, InfraScout attaches an estimated cost to each run, so you can see what an assessment spent regardless of which vendor's model produced it. Switching a connection to OpenAI doesn't leave you guessing about spend.

Full parity in chat

An OpenAI connection isn't a stripped-down mode. It runs the same chat experience you already know, feature for feature. Answers stream token by token as the model works. The agentic tool loop is intact, with access to the full set of InfraScout assessment tools — the model can size up your fleet, open a session against an agent, run commands, query Microsoft Graph and Azure, and save findings, exactly as it does on Anthropic. When the model reasons through a problem, its reasoning summaries appear as thinking, so you can follow how it reached a conclusion. Hosted web search is available so it can consult the outside world mid-review. And long conversations are kept in bounds by automatic context compaction, so a lengthy investigation doesn't run into a wall.

In practice, that means you can hand an OpenAI connection the same substantial work you'd give Anthropic — a tenant security review, a fleet-wide sweep, a playbook run — and it behaves the same way from the first token to the final finding.

Web answers carry inline citations

When an OpenAI model draws on the web, it doesn't just fold the information in silently — it cites where each claim came from, inline, as numbered source references you can open and verify. This matters for assessments specifically: a finding backed by an external advisory or a CVE entry is only as trustworthy as the source behind it, and the numbered citations let a reviewer trace a web-derived statement straight back to the page it rests on. The references stay with the message, so when you scroll back or hand a session off, the sources are still sitting next to the conclusions they support.

What differs on an OpenAI connection

A few controls behave differently on OpenAI connections, and InfraScout makes the differences visible rather than leaving you to discover them.

Web search on OpenAI honors an allowed-domains list — name the domains you want the model to consult and it stays within them. It does not apply blocked-domain lists or per-turn web usage caps, and there is no separate web fetch tool for reading a specific URL in full; those remain Anthropic-only capabilities. The model dropdown is scoped to the connection's provider, so you only ever see the models you can actually pick, and the settings that apply only to Anthropic are labeled as such. You always know which knobs are live for the connection in front of you.

Agentic Workflows run in sync mode

Agentic Workflows work on OpenAI connections too — you can point a headless, automated assessment at an OpenAI connection and it runs like any other. The one difference is execution mode: OpenAI workflows always run in sync mode, finishing inline in seconds, rather than on the batch engine. You don't have to remember this. If a scheduled run or a webhook fires a workflow that would normally batch, InfraScout falls back to sync automatically, so the run still happens — it just runs inline instead of queuing on the batch engine. Everything else about a workflow is unchanged: the instruction, the playbook, the tool-group scoping, the limits, and the full run history all behave the same.

Try it

Have an admin add an OpenAI key to an AI provider connection, choose OpenAI as its provider, and pick one of the GPT-5 models. Open a chat against that connection and give it something real — a tenant review or a fleet check — and watch it stream, call tools, reason, and cite its web sources exactly as an Anthropic connection would. Anthropic remains the default, so your other connections keep running unchanged.

See the AI provider connections guide for setting up and managing connections, and the chat guide for running assessment conversations. Questions or feedback? Reach us at info@infrascout.cloud.