Sarvam’s Prism writes no text, so output is free; input costs ₹4 per million tokens

01What 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

02Two 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.²⁴

03Whose 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”.

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.