You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.
Billing: cap what a single chat, report run or agent call may spend
New Feature. Only Company Admins.
Billing has a new Session limits page, next to Budgets. Set a default credit limit per session, and give any agent its own limit, higher or lower. A session is one chat, one report run or one agent invocation.
When a session reaches its limit, the turn in progress wraps up with what it has gathered so far, and the session takes no further turns. A chat locks its message box and offers to export the conversation. A report run ends incomplete and still delivers a partial report when it got far enough. An agent invocation fails with a clear reason. Sessions that hit their limit are marked in Top sessions.
New companies start with a default of 2,500 credits per session; existing companies have no limit until you set one. See the Credit budgets guide for how limits apply.
Answer an agent's questions in Chats with one click
New Feature.
When an agent needs you to choose between alternatives, it now asks with a card of numbered options above the message box: click one or press its number, type your own answer when the agent offers that, or skip. Several questions are asked one at a time and answered together. Reports and scheduled runs never stop to ask.
Learn more in Using chats.
An agent's question as a card of numbered options above the message box.
Company admins can now see their credit allowance in the Billing Center
New Feature. Only Company Admins.
If your TripBuilder contract includes a monthly credit allowance, the Billing Center Overview now shows it as a small note beside the Billable credits: this month's usage against the allowance. Hover over it for the reset date and any upcoming change. The monthly budget on the Budgets page shows the allowance beside its limit. Allowance credits are their own pool: included usage never counts against them. Learn more.
The allowance note in the Billable tile of the Billing Center Overview.
Agents: turn chat off per agent with the Streaming API switch
Improvement.
The Streaming API switch in an agent's Channels tab now really turns chat on or off. Every existing agent stays on.
With chat off, the agent leaves the chat agent pickers, Test Chat is greyed out, and an open conversation stops at its next message.
REST API calls and reports keep working.
The unused Max History Turns and Heartbeat Seconds settings are gone.
Learn more in Using agents.
Agents: Claude Sonnet 5.5 is now the default model
Improvement.
TAI's built-in agents now run on Claude Sonnet 5.5 — Anthropic's latest Sonnet — and so does any agent that follows the company default, unless your company chose a different model tier. Nothing to configure: existing agents pick it up on their next conversation. Agents pinned to a specific model keep it.
Company admins can now build connectors with Claude Opus 5.5
Improvement. Only Company Admins.
The Connector Builder now builds and self-tests connectors with Claude Opus 5.5. Nothing to configure: ongoing Builder conversations switch on their next message. Builder turns are now metered at the Most capable credit tier.
Agents: a clearer Channels tab with on/off switches per channel
Improvement.
Each channel now has an on/off switch with its own description; its settings appear below only while it is on.
Report settings are grouped into Output, Input parameters and Generation, and the output format is a choice with a short explanation per option.
Billing Center: connector calls made in batch steps are now counted
Fix. Only Company Admins.
Connector calls an agent made inside a batch step were missing from Usage Stats and the Billing Center. They are now counted and billed like any other connector call, from this release on. Responses served from the connector cache still get the cache discount.
Agents drop a thought cut off mid-step instead of repeating it later
Fix.
When an agent ran out of room while reasoning, the unfinished thought was sent back to the model as plain text on the later steps of the same answer. It is now discarded once the step is retried.
Connectors: agents see images returned by MCP tools again
Fix.
A connector tool that answers with text and an image now hands the agent the image itself, not a text placeholder.
Agents you can't run no longer show Test Chat or Test API
Fix.
Users below an agent's Minimum Workbench role no longer see Test Chat or Test API, which used to fail when clicked.
Chats: follow-ups now remember an answer that hit a limit
Fix.
A follow-up after an incomplete (limit) answer used to forget that answer. It now continues from it, so "continue" or "go on" works.