Latest in AI
DeepSeek V4.1 Flash 1017 on Sarvam: API Access, Pricing and Limits
Sarvam has removed the beta-access requirement, with reasoning defaults and image-input rules worth checking before switching an application.
Sarvam made DeepSeek V4.1 Flash 1017 generally available on October 30. Developers no longer need beta access to call it through Sarvam’s Chat Completions V2 or Responses API. This is a change in provider access, not a new DeepSeek model release or a price cut.
Sarvam API pricing and availability
The October 30 release note identifies the model as deepseekv4.1-flash-1017. It accepts text and images, produces text, and has a 1,048,576-token context window.
Sarvam lists these unchanged rates per million tokens:
| Token type | Price |
|---|---|
| Input | ₹25 |
| Cached input | ₹0.5 |
| Output | ₹105 |
The endpoints are POST /v2/chat/completions and POST /v2/responses. The availability change removes an approval step for developers who already want to use Sarvam’s infrastructure. It does not establish that this provider is faster or cheaper than every alternative.
Reasoning changes the cost of a request
The model documentation says reasoning defaults to high. Reasoning tokens are billed as output and share the output budget with the final answer. On Chat Completions, reasoning_effort: "none" disables thinking; Responses uses reasoning.effort: "none".
Budget for the reasoning, not just the answer. A short visible response can still involve more billed output than its length suggests. Conversely, a tight output cap can be consumed before a usable answer arrives. Sarvam documents that case as a successful HTTP response with null content and a length finish reason on chat.
For an existing application, the sensible comparison is cost per completed task under an explicit reasoning setting. Comparing only the price of a million tokens misses the effect of those defaults.
Image input has an integration constraint
Sarvam requires inline base64 image data rather than remote image URLs. Its request-body limit is 10 MiB, including that encoded data. Audio and video inputs are not supported.
That means an application passing hosted image links cannot assume migration is only a model-name change. It needs to prepare image payloads and check their size before sending the request. Long context does not remove the separate request-body constraint.
What to test before moving traffic
The useful next step is a small evaluation using the application’s actual documents, images and tool calls. Compare completion quality, total billed tokens and failure handling with the current provider. General availability removes the access gate; it does not replace that operational check.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders explain API changes and AI product launches in content that developers can actually use.