Sarvam Responses API Goes GA: Models, Pricing and Migration Limits

Sarvam made its Responses API and Chat Completion V2 generally available on October 22, alongside access to GLM-5.3, Gemma 4 31B and DeepSeek V4.1 Flash. Developers no longer need beta whitelisting, but moving an existing agent requires more attention than changing its model name.

The useful change is access. These are models available through Sarvam's platform, not three new models developed by Sarvam. Its own sarvam-105b is available on the same endpoints, so a team can compare different models through one provider without treating this announcement as a new benchmark result.

What became generally available

The October 22 changelog names two endpoints: POST /v2/chat/completions and POST /v2/responses. Both support sarvam-105b, glm5.3, gemma4 and deepseekv4.1-flash.

There is a small documentation mismatch worth noticing. The pricing page still contains a sentence calling the three externally developed models beta. However, the dated announcement and current beta-status documentation explicitly say the endpoints and models are generally available. The latter says no model on either endpoint is currently in beta.

For an application that already manages its own conversation history, this makes the service easier to evaluate. It removes an access requirement. It does not establish that every model is equally suitable for the application, or that the two API formats behave identically.

Sarvam API pricing

Sarvam's published prices are in Indian rupees per million tokens:

Model ID	Input	Cached input	Output
sarvam-105b	₹15	₹5	₹60
gemma4	₹35	₹10	₹90
glm5.3	₹126	₹20	₹390
deepseekv4.1-flash	₹25	₹0.50	₹105

GLM-5.3 and DeepSeek V4.1 Flash reasoning tokens are billed as output tokens. That matters when estimating an agent's cost: the visible answer alone does not tell you how much output was billed.

These numbers are useful for budgeting, not enough to choose a model. A cheaper call that needs repeated attempts can cost more per completed task. Compare the same workflow, including failures and retries, before choosing the default.

Responses compatibility does not include stored conversations

Sarvam's Responses API guide describes an OpenAI-compatible interface for generation, reasoning, structured output and tool calling. Its conversation handling is the important boundary: the endpoint is stateless.

Send the complete conversation in input on each request. Omit store or set it to false; true returns a 400 error. Continuing through previous_response_id or conversation is also unsupported. Your application needs to retain the messages and replay the relevant history itself.

This is manageable if that is already how the agent works. If the application assumes the provider remembers prior turns, changing the base URL alone will leave a gap in the workflow. The migration needs a deliberate decision about what history to retain, what to send again and how to keep it within budget.

Set an output budget before running an agent

The endpoint reference says max_output_tokens has no default cap. Set it explicitly. On reasoning models, that limit includes reasoning tokens, so a cap also affects how much room remains for the final answer.

There are model-specific details too. On the Responses endpoint, GLM-5.3 accepts low or high reasoning effort; omitting the setting selects max-level reasoning, while the literal max value returns 400. DeepSeek V4.1 Flash accepts all three values directly.

Another detail matters for tool orchestration: parallel_tool_calls: false is accepted but does not enforce a single tool call. An application that needs sequential execution should enforce that in its own execution layer rather than rely on the flag.

General availability makes evaluation easier; it does not remove integration work. For teams comparing models through Sarvam, the useful next test is a complete task with conversation history, tool execution and a spending limit. That reveals more than a successful first response.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI founders turn technical launches and product updates into clear stories that developers and founders read and share.

Book a 30-minute call.