Sarvam’s launch post and price pages of 16 November: open-weight models served from GPUs in India, what its cache does to the bill, and the rupee prices.
Sarvam serves more than 450bn tokens a day of DeepSeek V4.1 Flash, an open-weight model from the Chinese lab DeepSeek, from GPU clusters in India, and 95.7% of the prompt tokens it handles for that model are served from cache rather than computed again. It gave both figures on 16 November, opening its inference service, Token Factory, to everyone.1 It sells the model at ₹14 per million input tokens and ₹55 per million output tokens on a Standard endpoint, and ₹25 and ₹105 on an Interactive one built to hold about 200 tokens a second per user. Input that is reused from a cached prompt costs ₹0.25 per million on Standard and ₹0.50 on Interactive. All prices exclude GST.23
450bn+
tokens a day of DeepSeek V4.1 Flash, served from GPUs in India
Sarvam
95.7%
of that model’s prompt tokens served from cache
Sarvam
₹14 / ₹55
per million input / output tokens, Standard endpoint, introductory, excluding GST
Sarvam
₹0.25
per million cached input tokens, a fifty-sixth of the fresh rate
Sarvam
Token Factory serves Sarvam’s own Sarvam 105B alongside other labs’ open-weight models through one OpenAI-compatible API: GLM 5.3 and GLM 5.3 Flash from Z.ai, DeepSeek V4.1 Flash and V4 Pro, Kimi K3.1 Pro from Moonshot AI and Google’s Gemma 4 31B.1 Its catalogue also lists Alibaba’s Qwen 3.8 27B and Sarvam’s Prism decision model. Inference runs on Sarvam-operated GPUs in India, and requests “are not forwarded to the original model developer’s overseas API”.4 Sarvam says it runs “India’s largest B200 cluster” (the B200 is an Nvidia data-centre GPU); it gives no count of the chips.1 It calls DeepSeek V4.1 Flash “one of the most used models on OpenRouter”, the gateway where the Journal tracked Chinese models’ rising share of tokens.1
Every request runs in two phases: prefill reads the whole prompt in one compute-heavy pass, then decode writes the answer a token at a time. An AI agent calls a tool and comes back seconds or minutes later with almost the same prompt; Sarvam’s agent prompts on GLM 5.3 average 131,000 tokens. GPU memory held that context for about four seconds, two under load, so nearly every returning turn was read again from scratch. Sarvam added a shared store, built on Moonshot’s Mooncake, that every prefill server can read and that keeps prompts for hours or days. Over a week of production traffic its hit rate on DeepSeek rose from 91.4% to 95.7%, which halved the tokens that needed prefill; on GLM 5.3 at peak it cut the prefill servers needed from 58 to 34.1
Cached input costs ₹0.25 per million tokens on Sarvam’s DeepSeek endpoint, against ₹14 fresh
Sarvam Token Factory, DeepSeek V4.1 Flash 1017: price per million tokens by endpoint and token type, ₹, excluding GST
Source: Sarvam Token Factory pricing page and DeepSeek V4.1 Flash 1017 model page · as of 17 November 2026 · Standard prices are introductory; reasoning tokens are billed as output
Sarvam’s own comparison puts the Standard endpoint’s blended price at US$0.092 per million tokens, the same as DeepSeek’s own API, at 130 output tokens a second against 60, measured by Sarvam from Mumbai.1 The one OpenAI model its API processes in India, GPT-6 Luna, lists at US$0.11 in and US$0.55 out per million with its 10% regional premium, and prices cached input at a tenth of fresh input; Sarvam prices it at a fifty-sixth.5 Sarvam labels the Standard prices introductory and has not said when they end.3
Sources
Disclosure. TechBag is a software reseller: it earns when companies buy through it, including from vendors named here. The analysis is TechBag’s own; no vendor paid for or saw this piece before publication. Facts are labelled reported, TechBag count or our view.