Change Log

deepseekv4.1-flash-1017 now costs ₹14 input, ₹0.25 cached input, and ₹55 output per 1M tokens, down from ₹25, ₹0.5, and ₹105. See Pricing.

For faster output, the interactive variant, deepseekv4.1-flash-1017-interactive, generates more than 200 tokens per second (about 80 on the standard model) and costs ₹25 input, ₹0.5 cached input, and ₹105 output per 1M tokens.

Chat Completions now return one completion per request on every model, on both /v1 and /v2. Requests with n greater than 1 return 400 invalid_request_error. To get several completions, send separate requests. See V2 feature support.

Token Factory models are now labelled Serverless (any API key can call them, billed per token) or Provisioned Throughput. Models that were labelled beta are now on Provisioned Throughput: capacity set up for your workload and enabled for your organisation. Access works as before; contact us to request it.

• kimi-k3.1-pro replaces kimi-k3, on Provisioned Throughput at ₹200 input, ₹20 cached input, and ₹950 output per 1M tokens.
• kimi-k3, mimo-v2.6-flash, and mimo-v2.6-pro are no longer offered. For the lowest per-token price, use glm5.3-flash.

Three new Token Factory models are available on Provisioned Throughput. Contact us for access. Prices are per 1M tokens.

• qwen38-27b: a 27B reasoning model with text and image input, a 262,144-token context window, and tool calling, on Chat Completion V2 and Responses. ₹18 input, ₹5 cached input, ₹230 output.
• clef: a 27B decision model on the Decisions API. Like Sarvam Prism, it returns a calibrated probability for every option, and it also answers questions about images. ₹20 input, no output charge.
• embeddinggemma-2: 768-dimension text embeddings on the OpenAI-compatible POST /v2/embeddings. ₹2 input, no output charge.

You can now define your own metadata fields on a knowledge base, such as department, year or document type, and fill in the values for each file as you upload it. Mark a field as required, limit it to a list of allowed values, or make it filterable, then change the definitions later from the Fields tab. Every file’s values show up in the file table and in the preview pane. See Knowledge Bases.

You can now connect third-party tools to Sarvam and use them in your voice agents. Open Integrations in the sidebar to browse or search the catalogue, connect a tool once for the whole workspace by approving access on the vendor’s consent screen, or add any other MCP server by its endpoint URL. To use one, open Add a tool in the agent editor, choose MCP, and tick the actions that agent should be able to call — each parameter can be a fixed value, an agent variable, or decided from the conversation.

You can now add files up to 5 GB to a knowledge base, up from the 100 MB the upload dialog was quietly allowing. Files that are too large or the wrong type are now flagged with a clear message instead of disappearing without explanation. See Knowledge Bases.

deepseekv4.1-flash-1017 is now generally available on POST /v2/chat/completions and POST /v2/responses. Your API key no longer needs beta access. It reasons at high by default (reasoning_effort: "none" turns thinking off), accepts text and image input, and has a 1,048,576-token context window.

Pricing is unchanged: ₹25 input, ₹0.5 cached input, and ₹105 output per 1M tokens. See Pricing.

When you click Download CSV in Call Logs, we now email you a download link once your file is ready. You can also find your exports from the last 7 days under Recent exports.

You can now exclude contacts already on a campaign with Exclude cohort. Upload a CSV of users to remove — matched by user identifier or phone number — via the same cohort upload endpoint with cohort_category set to stop_file. An exclude cohort only removes contacts already on that campaign; it does not filter users added later. See Exclude cohort.

language_code / target_language_code is now optional on POST /voices/clone, POST /voices/clone/stream, and GET /voices/clone/ws. Omit it to detect the language from text. Pass the code when pronunciation has to be exact — detection can be wrong on short or code-mixed text, romanized Indic and Urdu return 400, and short Assamese without ৰ or ৱ is often detected as Bengali. On the WebSocket the first detection stays locked until you send a new config with an explicit code. See Supported languages.

You can now connect GitHub App as a Cowork connector and choose exactly which repositories your agent can reach. In our tests it finished tasks 26% faster and used 40% fewer tokens than the old connector. See Connectors.

Agent pages now have a Mine / Team toggle, so you can see your own recent sessions or those of your teammates in the same workspace. Any session opens as a read-only transcript you can review after the run finishes.

POST /v2/chat/completions and POST /v2/responses are now generally available with GLM-5.3, Gemma 4 31B, and DeepSeek V4.1 Flash, alongside sarvam-105b. No beta whitelisting is required for these endpoints or models. See Beta APIs and Models for anything still in beta.

New open-weight models are available on POST /v2/chat/completions and POST /v2/responses.

Generally available:

• glm5.3-flash: a faster GLM-5.3 with image input, log probabilities, and n > 1. Reasons by default. 1,048,576-token context.

In beta:

POST /v2/chat/completions and POST /v2/responses are now generally available with sarvam-105b, glm5.3, glm5.3-flash, and gemma4. Your API key no longer needs to be whitelisted. The new models in the entry above are in beta; see Beta models.

Call recordings are now captured in stereo, with the user on the left channel and the agent on the right. Mono recordings mixed both speakers onto a single channel, so overlapping speech was summed and could clip; in stereo each speaker stays isolated, which makes recordings easier to listen to and better suited for downstream processing (transcription, analytics). This is now enabled for all calls. See Call recordings.

The Analytics API now gives you more of what happened on each call. Get forwarded recording returns the recording of the part of a call after your agent transferred it, currently for transfers made through Vobiz, and Get trace returns a call’s step-by-step event trace: the speech recognition, LLM, tool-call, and text-to-speech events of every turn, plus the conversation’s start, end, and post-call analysis. List attempts and List interactions now include provider details for each call, channel_provider (the telephony provider that carried it) and provider_identifier (that provider’s own call ID, such as a call SID), so you can match calls to your provider’s records. Both also add deployment_id, app_version, bot_contact_identifier (your agent’s own number), and goal results in goal_overall_status and goal_criteria. You can also sort by attempted_at, and filter on app_version and goal_overall_status.

You can now save up to 50 cloned voices on the Starter plan, up from 5. Pro and Enterprise plans have higher limits. Voices are counted across your whole organization, in Content Studio and the API, and deleting a voice frees up a slot. See Managing Your Voice Library.

Genie, the AI assistant in Voice Agents, now goes beyond creating and editing agents. It is available from every Voice Agents page through Ask Genie, in a panel that stays open as you move between pages, so you can build an agent, deploy it, and check its calls in a single chat. Genie can now put an agent live on one of your connected phone numbers for inbound calls and place a test call to your phone. It can also create outbound campaigns, upload contact lists from a CSV, and pause, resume, edit, or cancel campaigns. To analyze calls, Genie can summarize performance and connect rates, explain why calls didn’t connect, read transcripts, check tool errors, and build Boards. Within an agent, it can also set up knowledge bases and test suites. Genie asks you to confirm before anything that places calls, affects live traffic, or can’t be undone. See Genie.