Mumbai/New Delhi: Two of the country’s feted artificial intelligence (AI) firms, including the not-for-profit BharatGen and unicorn Sarvam AI, have rolled out access to their foundation models at a price that is a fraction of what Anthropic, OpenAI and China’s DeepSeek are offering.
Industry experts attributed part of the pricing advantage to government subsidies for graphics processing units (GPUs) used to train and run AI models. They cautioned, however, that India's AI firms remain far behind global rivals in capability and may not trigger the kind of worldwide price disruption that China's DeepSeek did.
Internal pricing data from a government-backed AI firm, reviewed by Mint, show that BharatGen's Param-2 model costs ₹5 per million output tokens, while Sarvam charges ₹10 for its 30-billion-parameter model and ₹16 for its 105-billion-parameter model.
The cheapest US-based rival is GPT-5 Mini by OpenAI, which falls in the same class as Sarvam’s 30-billion-parameter AI model—and charges ₹191 for responses of the same size. The larger Gemini 3.5 Flash by Google costs ₹858.
The Chinese model, DeepSeek, is more expensive than its Indian counterparts: DeepSeek’s 284-billion-parameter V4 Flash AI model costs ₹27 per 1 million tokens. To be sure, V4 Flash is nearly 17 times as large as BharatGen’s Param-2, which is built on only 17 billion data parameters.
On the input side, BharatGen charges ₹1 per million tokens, while Sarvam charges ₹2.5 for its 30-billion-parameter model and ₹4 for its 105-billion-parameter model. GPT-5 Mini costs ₹23.8 per million input tokens, while Gemini 3.5 Flash charges ₹143.
Output tokens refer to the AI-generated response, while input tokens refer to the text or prompt submitted by the user. Prices of AI models are typically measured per 1 million ‘tokens’, which equates to 750,000 words on average.
Since the advent of generative AI in the mainstream, the cost of running AI has remained one of its biggest challenges to global adoption, with most companies deeming it too high.
AI models are built using "parameters" of data—bits and bytes of information that the algorithms learn when being ‘trained’.
Industry stakeholders said the comparison comes with a caveat. While OpenAI, Google and DeepSeek develop large, general-purpose AI systems designed to perform a broad range of tasks across multiple languages and reasoning benchmarks, BharatGen and Sarvam's current models are smaller, specific to Indian languages, and yet to be proven among consumers and enterprises.
Foundational AI models are algorithms trained by tech companies from scratch, capable of human-like writing, speech and even video production. While more parameters typically correlate to larger models with more information and ability to process a wide range of subjects, they’re not an absolute gauge of an AI model’s quality.
Can the pricing last?
GPUs, specialized electronic circuits, which India’s IT ministry has so far subsidized for both BharatGen and Sarvam, typically account for just over half the net expenditure of a company solely working on developing AI models.
The current pricing of Indian foundational models is important, as low-cost AI access for users is one of the key tenets of India’s $1.2-billion AI Mission, run under the ministry of electronics and IT (Meity).
Global leaders have also pushed for AI to become affordable. On 18 June, Jonathan Ross, inventor of Google’s custom processing chip and chief software architect at Nvidia, told Mint that running AI for under ₹100 could make the technology accessible to everyone in India.
Industry stakeholders, however, are not yet convinced—underlining whether such pricing by India’s state-backed AI startups may be sustainable once government-subsidized access to GPUs runs out.
"Some of the cost advantages are structural and will remain because of engineering efficiencies. But part of the current advantage comes from policy intervention designed to accelerate India's AI ecosystem," said Jibu Elias, country lead for responsible computing at non-profit internet services body Mozilla Foundation. “Over time, pricing will need to reflect commercial realities such as GPU depreciation, electricity costs, data-centre operations and ongoing model development.”
AI analyst Kashyap Kompella, said comparing Indian AI startups with US giants may not be a direct and fair comparison. So far, BharatGen’s first model is available for developers through public platforms. Sarvam has models and a chatbot named Indus—but access to it is limited.
In an interview with Mint in February, Sarvam co-founder Vivek Raghavan said the company “is not competing with ChatGPT and Gemini,” underlining that US giants have “much more capital available to offer consumer-end AI.”
While Indian firms, for now, appear to be undercutting the price by OpenAI, Anthropic, Google and DeepSeek, they’re not direct competitors. This, in turn, could also impact the kind of enterprise deals they may win.
Especially for Sarvam, enterprises are proceeding cautiously before signing on Indian AI models at scale, said Anushree Verma, senior director analyst at Gartner. "Enterprises are still in the early stages of deploying projects with companies like Sarvam and others. They simply can’t price themselves at OpenAI or Anthropic levels, so they’ll need to be far more strategic on pricing to stay competitive."
Winning enterprise customers
Sarvam has made four client announcements so far—with financial services firms Razorpay, SBI Life, and Tata Capital, and e-commerce platform Swiggy. On 15 June, the company announced a $234 million funding round, valuing it at $1.5 billion, to build foundational models in Indian languages.
A representative of BharatGen declined to comment. An email sent to Sarvam seeking comment did not elicit a response.
Pricing won’t let Indian firms target the diversified uses that Big Tech is eyeing, Kompella added. "A model built for Indian languages, speech, translation and enterprise workflows can be smaller, more specialized and cheaper to serve than a general-purpose frontier model trying to be best-in-class across every global benchmark. A large part of the global AI cost base sits in dollar-denominated research talent, hyperscale cloud compute, and very large frontier-model training runs,” he said.
“Indian sovereign AI companies have lower local operating cost and can narrow their products to Indian languages, and local enterprise or government use cases. That does not make AI cheap in an absolute sense, but it does change the economics," he said.