Chat Update: Sonnet 5, Data Region, and Clearer Safety Refusals
July 2, 2026 · InfraScout Team
This is a short roundup of three improvements to assessment chat. You can now choose Claude Sonnet 5 for everyday reviews, pick where inference runs for each Anthropic connection to meet data-residency requirements, and see a clear card in the thread whenever a model declines to continue for safety reasons.
Claude Sonnet 5 joins the model picker
Claude Sonnet 5 is now selectable when you configure an Anthropic AI provider connection. Released on June 9, 2026, it is Anthropic's mid-tier model — a strong, balanced choice for routine assessments and the kind of steady, multi-step work that makes up most reviews. It ships with a 1 million token context window, so long playbook runs and conversations that range across many agents and systems have plenty of room before compaction steps in.
Its standard pricing flows straight into InfraScout's cost estimation, so the audit view attributes spend accurately from the first message. When you weigh Sonnet 5 against faster or more capable options, the cost you see reflects the model you actually picked.
Sonnet 5 is added as an option only. It is not the default, and existing connections are unchanged — nothing shifts unless an admin selects it in the AI provider configuration. Reach for it when you want a capable general-purpose model with a large context window; keep your faster models for quick questions and your top-tier model for the hardest investigations.
Choose where inference runs for each connection
Anthropic connections gain a data region setting that controls where inference runs for that connection. You can leave it on the workspace default, or pin a connection to the United States or to a global region. This is useful when a customer engagement or an internal policy requires assessment traffic to be processed in a specific region.
The setting is chosen per connection, so different connections can target different regions to match the requirements of the tenants they serve. InfraScout validates your choice before it runs the connection test, so an unsupported selection is caught up front rather than surfacing later as a confusing failure.
The region options apply to Anthropic connections; other providers ignore the setting. A European region is not offered yet, because not every Anthropic organization is entitled to EU residency — we will add it as an option once it is broadly available. Until then, choose the region that best fits your requirements and confirm it saves cleanly by reopening the connection.
Safety refusals now show as a clear card
When a model declines to continue for safety reasons, the chat now makes that outcome plain. Previously a refusal could end a turn with nothing to show — the conversation simply stopped, leaving you to guess why. Now the thread renders a warning card that states the reason the model declined, right where the answer would have appeared.
Any partial answer the model wrote before it stopped is kept and shown above the card, so you do not lose the work that was already produced. The refusal is treated as a real, final outcome rather than an error, so it never triggers the retry banner — there is nothing to retry, and the chat does not push you toward re-running the turn. The card is part of the saved conversation, so it survives a page reload and stays visible when you or a reviewer reopen the thread later.
This works the same way on both Anthropic and OpenAI connections, so however your chat is configured, a safety refusal reads as a clear, self-explanatory step in the conversation instead of a silent stop.
Try it
Have an admin add or edit an Anthropic connection to try Claude Sonnet 5 and set the data region that fits your requirements, then run your next review. Under the hood, the chat's web search and web fetch tools also moved to Anthropic's latest tool versions, so those steps keep working with no change on your side.
Questions or feedback? Reach us at info@infrascout.cloud.