--- title: "Sarvam’s Prism writes no text, so output is free; input costs ₹4 per million tokens" summary: "How Sarvam’s Prism decision model works, what a call costs and whose benchmarks rank it, from Sarvam’s release, API docs and model card of 3 November." url: https://www.thetechbag.com/blog/sarvam-prism-decision-model-4-rupees-million-tokens format: Note desk: "AI in the enterprise" region: India published: 2026-11-04 facts_checked: 2026-11-04 author: TechBag (The TechBag Journal) tags: ["Sarvam","Sarvam Prism","decision model","open weights","Apache 2.0","token pricing","classification","routing","guardrails","agent evaluation","calibration","Indian languages","GST"] cover: https://www.thetechbag.com/blog/covers/sarvam-prism-decision-model-4-rupees-million-tokens.png --- # Sarvam’s Prism writes no text, so output is free; input costs ₹4 per million tokens *How Sarvam’s Prism decision model works, what a call costs and whose benchmarks rank it, from Sarvam’s release, API docs and model card of 3 November.* Note · AI in the enterprise · By TechBag · Published 2026-11-04 · Facts checked 2026-11-04 · Canonical: https://www.thetechbag.com/blog/sarvam-prism-decision-model-4-rupees-million-tokens Bengaluru’s Sarvam released Prism on 3 November: a model that answers without writing. Given a text and a question with a fixed set of answers, it returns a probability for every answer “in a single forward pass”.[1][4] Sarvam charges ₹4 per million input tokens and ₹0 for output, “on every plan”; by its own example, three short questions about a 1,000-token text cost about ₹0.004.[2] Prices exclude GST.[3] The weights are open, under the Apache 2.0 licence.[4] ₹4 per million input tokens, before GST Sarvam price list ₹0 per million output tokens: it returns probabilities, not text Sarvam API docs 44 ms median per request on one Nvidia B200; 95th percentile 420 ms Sarvam model card ₹0.004 three questions about a 1,000-token text, Sarvam's example Sarvam API docs ## What a decision model does Sarvam calls Prism a “System One” decision model, after Kahneman’s fast, intuitive mode of judgement, a category it says TypeSafe created a month earlier with its Jev model.[1] Where a chat model is told to “reply with one of these labels” and its reply is parsed, Prism takes a `state` and typed questions: yes or no, pick one of 2–255 options, or a scale of 2–10 levels. It returns probabilities that sum to 1 for each question, and `output_tokens` is always 0, which is why output is free.[2] Inside, a fine-tuned Sarvam 105B reads the text and the options, and a small decision head turns the end of each option into a probability; no option can see the others.[4] In Sarvam’s worked example, a ticket saying all checkout payments are failing comes back urgent at 0.98 and routed to billing at 0.80, for 153 input tokens.[2] Sarvam suggests acting automatically at a threshold such as 0.9 and sending the rest to a person: 68.5% of English answers clear that bar, and 38.4% in Indian languages.[2] TechBag countWhat a million calls cost At ₹4 per million tokens plus 18% GST, ₹4.72. A million calls of the size in Sarvam’s example, a little over 1,000 tokens each, come to just over ₹4,720 with GST, by our arithmetic. The input is counted once per request however many questions are asked.[2] On the same list Sarvam 105B costs ₹15 in and ₹60 out per million.[3] ## Two limits: context length and serving The hosted API takes 32,768 tokens per question, and one question over the limit fails the whole request.[2] The open-weights code takes 131,072,[4] the figure behind Sarvam’s claim that long agent traces fit “in one pass”.[1] Self-hosting needs one FP8 GPU with about 120 GB free, such as a B200 or H200, and the weights are in a custom layout that “general-purpose inference servers” cannot serve. The API reads text only and does not stream.[2][4] ## Whose benchmark Sarvam ran 25 decision models through three indices itself: the Jev Decision Index (37 public benchmarks), its own recreation of the Artificial Analysis Decision Index with curated hard questions in place of the private test, and an internal Indian-language benchmark.[1] Its claim is the highest average “of any model that answers in under 100 ms”. **Figure: On Sarvam's own tests, Prism ranks third overall and first among models answering in under 100 ms** *Average score across three indices, the 8 highest-scoring of 25 decision models, with median response time per request on one Nvidia B200* mammoth-1 284 ms 75.2 GLiDE 1.2 116 ms 73.2 Sarvam Prism 44 ms 71.8 elephant-1 145 ms 71.3 StartLux-27B 117 ms 70.8 KnowLine-35B 50 ms 70.8 d3-ultra 229 ms 70.3 KnowLine-26B 63 ms 69.1 | | Value | | --- | --- | | mammoth-1 | 75.2 | | GLiDE 1.2 | 73.2 | | Sarvam Prism | 71.8 | | elephant-1 | 71.3 | | StartLux-27B | 70.8 | | KnowLine-35B | 70.8 | | d3-ultra | 70.3 | | KnowLine-26B | 69.1 | Source: Sarvam, Introducing Sarvam Prism (Sarvam's benchmark) · as of 3 November 2026 · Run by Sarvam; two of the three indices are Sarvam's construction TechBag countRead across the table Prism leads no single index: 67.0 on Jev’s against mammoth-1’s 75.3, and 82.5 in Indian languages against KnowLine-35B’s 82.6. Its edge under 100 ms is one point over KnowLine-35B.[1] On Sarvam’s other tests, Prism judged 700 recorded agent runs with 78.5% accuracy against 70.7% for Claude Haiku 5.5, which Sarvam says took a median 2.5 seconds and US$0.64 per 1,000 runs. It scored 89.0% on Sarvam’s internal 5,517-question security set, which Sarvam says it will publish “in the coming weeks”.[1] Until it does, and until others rerun the indices, every score here is Sarvam’s. How agent products are billed today is in [our piece on AI agent pricing](https://www.thetechbag.com/blog/how-ai-agents-are-billed-credits-actions). How we did this Figures as of 4 November 2026, 7pm IST Prices and limits are from Sarvam’s API page and price list; speed, hardware and licence from the model card; every score from Sarvam’s release post and its data tables. The chart shows the eight highest averages of 25; response times are Sarvam’s medians on one B200. GST and the cost of a million calls are our arithmetic. We found no independent test of Prism. Spotted an error? Write to us; corrections are published at the foot of the piece. Sources 1. [Sarvam AI, Introducing Sarvam Prism](https://www.sarvam.ai/blogs/sarvam-prism) — 3 November 2026 2. [Sarvam API docs, Sarvam Prism (System-1 Decision Model)](https://docs.sarvam.ai/api/token-factory/models/sarvam-prism) — 3 November 2026 3. [Sarvam API docs, Pricing, Credits & Rate Limits](https://docs.sarvam.ai/api/getting-started/pricing) — 3 November 2026 Undated page; Prism launched 3 November. 4. [Sarvam AI, sarvamai/prism model card, Hugging Face](https://huggingface.co/sarvamai/prism) — 3 November 2026 **Disclosure.** TechBag is a software reseller: it earns a margin when a reader buys 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. **Read this next:** [An Agentforce action costs 10 cents, a conversation $2: how AI agents are billed now](https://www.thetechbag.com/blog/how-ai-agents-are-billed-credits-actions) — How five vendors price AI agents by action, conversation, credit, token and seat, from their own price pages, with a month’s bill worked through. Go deeper on TechBag - [Freshdesk](https://www.thetechbag.com/freshworks/freshdesk) — a help desk where ticket triage like Sarvam’s worked example happens today - [ITSM and service desk guide](https://www.thetechbag.com/browse/endpoint-management/itsm-service-desk) — the service desks that route and prioritise tickets, compared