Change Log

Agents can now be regression-tested with Tests: describe the conversations that matter as test cases, and the platform plays a simulated user against a committed version of your agent while an AI judge grades each expected behavior Pass or Fail. Suites carry shared variables and global guardrails, runs execute each case up to 20 times to expose flakiness, and every execution returns the full transcript with per-behavior reasoning. The same suites can be created, run, and read over REST — see the Tests API overview and Best practices.

Voice agents are now available over MCP, so Claude, Cursor, VS Code, and other MCP clients can build, deploy, test, and analyse voice agents directly from your assistant. It’s a hosted, remote server — nothing to install — reachable over streamable HTTP at https://mcp.sarvam.ai/voice-agents and authenticated through Sarvam SSO (OAuth 2.1). See the MCP Server page for setup, the full tool surface, and security guidance.

Added GLM-5.3 (glm5.3) as an open-source chat model on /v2/chat/completions. It offers a 1,048,576-token context window with tool calling and visible reasoning, priced at ₹126 / ₹23.4 / ₹396 per 1M tokens (input / cached input / output), with reasoning billed as output. Beta access is granted per API key; contact us to request whitelisting.

Boards can now be created, edited, and run over REST instead of only in the dashboard. Discover the tables available to your workspace with query-schema, develop SQL against query-preview, then save it as a widget and run a whole tab in one call. Recurring delivery of a widget’s result to email or Slack is configurable through notification rules. Authenticate with X-API-Key as with the rest of the Voice Agents API. See the Boards API overview, Writing board SQL, and Best practices.

Control how long Sarvam keeps your data, self-serve. Settings → Workspace → Data retention now lets an org Owner set a workspace-wide default retention period, plus a per-product override for Model APIs and Voice Agents — support for the rest of the product line is rolling out. Changing the period only affects data going forward; nothing already stored is deleted early or extended retroactively. See Data retention.

Assign a dedicated Text-to-Speech voice persona to each starting language directly from the agent’s Voice settings, replacing the need for hardcoded voice configuration. Enable Per-language voices (Overrides default voice) and define a conditional mapping for each language — for example, Mansi for Hindi, Cynthia for Odia, or Arjun Bhatt for English. At the start of a call, the agent selects the voice persona assigned to the starting language. The assigned speaker then remains consistent for the duration of the call, even if the caller switches languages mid-conversation.

Added Keyterm Prompting support for saaras:v4 on the REST and Batch APIs. Pass a JSON list of up to 50 domain-specific names, places, brands, or technical terms under keyterms to bias recognition toward them.

saaras:v4 is now available on the Realtime API alongside saaras:v3-realtime, with the same connection parameters and message protocol. saaras:v3-realtime remains the default.

saaras:v4 is now live on Voice Agents and generally available - routing happens automatically, with no configuration change needed. Saaras V4 extends English recognition beyond Indian English to Global English accents, stays state-of-the-art across all 22 Indian languages, and improves language identification and robustness to noisy, code-mixed audio. See the Saaras model page for the full feature list.

A voice agent can now hand a live phone call to a human, another number, or your own SIP trunk when the conversation needs it - an escalation, a specialist, or an explicit request to speak to someone. Configure it as Call forwarding under Settings → In Call Actions, with a fixed destination or an agent variable resolved at runtime, to a phone number (E.164) or a SIP URI. Supported end to end on Vobiz. Transferred calls show End Reason CALL_TRANSFER in Call Logs. See Call transfer.

Inbound deployments can now POST a call-completion event to your webhook endpoint after every inbound call, with the transcript, duration, agent variables, and the caller’s phone number - configured via the deployment’s webhook_config. See the Deployments webhook payload reference, List webhook deliveries, and Retry webhook deliveries.

Agent campaigns can now rotate caller IDs on retries. With Phone number rotation enabled, every subsequent call to the same user uses a different agent number until each number in the pool has been used once for that user - so a contact who didn’t answer one number gets the retry from another, not the same one again. Requires two or more agent numbers on the campaign. Enable it from the Phone number rotation toggle under Advanced on the Schedule step of the campaign wizard, or set phone_rotation.agent: true in app_config.

The product formerly called Content Agents is now Content Studio. Page URLs are unchanged.

The product formerly called Work Agents is now Cowork. Its docs moved from /work-agents/* to /cowork/*; the previous URLs redirect to the new ones.

API tools can now receive live call context as variables. In the tool editor, select User Identifier to pass the caller’s phone number, and Call Transcript, Interaction ID, and Call Length to pass the full conversation transcript, the unique interaction ID, and the call’s duration. These are available alongside agent variables - visible in the platform UI and fetchable via the APIs.

The Max call length limit in conversation settings was raised from 25 to 60 minutes, unlocking longer use cases like employee training, guided walkthroughs, and extended support conversations. When the limit is reached, the agent still wraps up and hangs up.

Campaigns can now POST an event to your webhook endpoint as each call finishes, whether or not the call connected. Set the Webhook URL in the campaign wizard’s Advanced step. The payload carries call outcome, connectivity status, retry metadata, timestamps, and output agent variables. Every delivery is recorded with a status and can be re-delivered via Retry webhook deliveries. See Webhooks under Campaigns.

sarvam-30b has been fully deprecated and moved to the Deprecated Models section, and is no longer available through the Chat Completions API. Sarvam-105B is now the only supported Sarvam chat model - migrate any remaining sarvam-30b integrations to sarvam-105b.

Introduced saaras:v4, available alongside saaras:v3 across the REST, WebSocket, and Batch Speech-to-Text endpoints. It supports the same five output modes (transcribe, translate, verbatim, translit, codemix) and adds Global English alongside Indian English, on top of the existing 22 Indic languages. saaras:v3 remains the default, recommended model.

Replaced the Document Digitization API with Document AI (/doc-ai/v1), powered by an upgraded Sarvam Vision 1.5 trained for both OCR and key-value extraction. Two endpoints: Digitise for full-document OCR with layout and tables preserved, and Extract for schema-based field extraction to JSON, CSV, or XLSX. The earlier Document Digitization API is now legacy.