Sarvam AI unveils plans for a trillion-plus parameter model built from scratch in India, priced five times cheaper than global rivals, alongside new voice tools and an India-hosted inference platform.

Bengaluru-based Sarvam AI has announced plans to build a trillion-plus parameter foundation model entirely from scratch in India, positioning it to compete directly with frontier systems from OpenAI, Google, and Anthropic while undercutting them sharply on price.

The announcement was made at Epoch 2026, Sarvam's first developer conference held in Bengaluru, where the company unveiled its biggest product roadmap to date. Cofounder Pratyush Kumar told the audience at the event that the company was building the trillion-plus parameter model right in India, adding that it was being trained from scratch to be competitive in coding, cybersecurity, simulation, and scientific research. He said the model is expected to arrive within six months, according to the company.

Alongside the announcement of the upcoming frontier model, Sarvam also revealed new pricing for its existing Sarvam 105B model. The upgraded model will now cost $0.80 per million blended tokens, which the company said makes it roughly five and a half times cheaper than OpenAI's GPT-5.4 Mini, priced at $4.50, and more than eleven times cheaper than Google's Gemini 3.5 Flash, priced at 9 dollars per million tokens.

New voice tools take centre stage