August 21, 2026 | Respan Docs

New

Response formats: added a dedicated place to create, edit, and reuse saved response formats across model configurations.
Online evaluations: added alerts for online evaluation results.

Experiments: stabilized trace loading with clearer loading states, early row previews, and protection against stale results.
Experiment inference: reasoning models can now use reasoning and tool calls together during a run.
Models: updated model availability, pricing, context limits, recommended ordering, and BYOK filtering across the Platform and Enterprise catalogs.
Behaviors: improved custom behavior creation with better candidate selection and related reliability fixes.
Monitors and reports: standardized time, number, and unit formatting for clearer and more consistent results.
Agent: restored page context and skills, and corrected metric filter definitions.
Gateway: responses now preserve the exact model name provided in the original request.
Slack integration: improved channel caching and request handling to prevent rate-limit errors during setup.

Log privacy: requests with logging disabled no longer retain request or response content.
Log filters: fixed list-based filters causing server errors and prevented invalid filters from silently broadening results.
Log exports: large exports now complete more reliably, accurately report delivered rows, and fail clearly instead of returning incomplete files.
Datasets: fixed trace imports and large row-removal operations failing on high-traffic accounts.
SQL editor: prevented non-admin users from accessing internal system columns and metadata while preserving supported log filtering.
Gateway: fixed empty provider chunks ending streams early and ensured failed streamed responses still produce logs.
Gemini usage: fixed responses incorrectly reporting zero token usage or cost.
Monitors: fixed high-volume rolling monitors missing scheduled evaluations or firing late.
OAuth: fixed organization switching and authentication issues in MCP integrations.