The cheapest AI in the world is Indian. It is also the weakest

Of the 257 artificial intelligence models the world can buy today, two were built in India. Almost all of the country's AI work is answered by machines made in America and China.

Ask your bank's chatbot why a payment failed. Ask a homework app to explain photosynthesis to a Class 8 student. Something answers in a second, in decent English.

Almost none of that “something” was built here. Of all the AI work Indian developers sent out between January and July through OpenRouter, a service that routes requests to hundreds of models at once, 94 per cent went to machines built in America or China.

Artificial Analysis, a benchmarking firm, runs every major model through the same nine tests. It tracked 610 of them. Of the 257 still on sale, two were built in India, both by the same Bengaluru company, Sarvam.

NINTH OUT OF ELEVEN

India's best model is called Sarvam 105B. It scores 11.9 out of 100 on that scoreboard, according to Artificial Analysis data read on August 20, 2026. Anthropic's Claude Opus 5, the world's best, scores 63.1. China's best, Kimi K3, scores 59.7.

Line the countries up by their strongest models, and India comes ninth of the 11 that have built anything at all, behind South Korea, Singapore, France, Canada, Spain, and the United Arab Emirates.

One caveat: artificial Analysis marks both Sarvam scores as estimates rather than full test runs. It does this when a model has not been through the whole suite. The ranking holds, though.

TWO YEARS BEHIND

Points out of 100 mean little on their own. Time is easier to feel. The first AI model ever to score higher than 11.9 was o1-preview, which OpenAI released on Sept 12, 2024, according to the same scoreboard. India's best model today is roughly where the world's best stood two years ago. Of the 257 models a developer anywhere can buy this month, 171 score above it.

The gap is not closing quickly, because India is barely in the queue. Since January, the US has put out 113 new models and China 91. India has put out two.

Nor is the frontier the only thing ahead of it. Gpt-oss-120b, a model OpenAI released — free for anyone to download and run on their own machines — scores 24.1. That is more than double Sarvam 105B.

WHAT INDIA GOT RIGHT

Sarvam is close to the cheapest thing on the board. A batch of work that costs $10 to run through Claude Opus 5 costs about seven cents on Sarvam 105B, a 135th of the price. And Sarvam gives its models free to download and run, which neither OpenAI nor Anthropic allows with their strongest models.

That is the opening China took. Chinese labs went cheap and open, and in June their models overtook American ones on OpenRouter. More than 60 per cent of the work routed through it last month went to Chinese models.

It has not worked that way at home.

NINETY-FOUR OUT OF EVERY HUNDRED

That traffic is not small. India pushed nine trillion tokens through OpenRouter over those seven months. A token is the word fragment an AI reads and writes. That makes India the 10th-largest market in the world. Fifty-one per cent went to American models, 43 per cent to Chinese ones.

The six squares left over cover every other country on earth, India's two models among them. It is a remainder rather than a measured Indian share, and still the closest thing anyone has to a figure for how much of India's AI work runs on Indian AI.

THE TEST NOBODY SETS IN TAMIL

The nine tests measure coding, scientific reasoning, general knowledge, and how well a model works on its own. Not one measures an Indian language. Nothing in the score says whether a model can take a complaint in Bhojpuri, read a land record in Marathi, or answer a pension query in Tamil. That is the problem Sarvam was set up to solve, and the one a global leaderboard cannot judge.

NOT EVERYONE THINKS THIS IS A PROBLEM

India's most quoted technologist says it should not be. At the NCAER India Policy Forum in New Delhi on August 6, Infosys co-founder Nandan Nilekani argued the country should stop chasing the frontier.

"The Western focus on AI is on infrastructure like larger models, more chips, more data centres. But infrastructure eventually gets commoditised," he said. "The value shifts to the application layer, and India can become the AI use-case capital of the world."

The numbers give him something to stand on. Stanford University's AI Index found that 58 per cent of employees worldwide used AI at work regularly in 2025. But in India, China, Nigeria, the UAE, Egypt, and Saudi Arabia, the share passed 80 per cent. India is among the heaviest users of a technology it does not build.

The same report found India registered the sharpest rise in concern about AI of any country surveyed between 2024 and 2025, up 14 percentage points, against two points of extra excitement. Being the use-case capital also means being the most exposed to decisions taken in San Francisco and Hangzhou.

THE PART THAT IS NOT ABOUT TECHNOLOGY

"India does not need to build every layer of AI itself, nor does it need to win every benchmark. But as AI begins mediating banking, education, courts, and public services, dependence on foreign models becomes more than a technology or trade issue. It becomes a question of accountability, geopolitical risk and institutional sovereignty," Dr Sanjay Sharma, a Supreme Court lawyer who works on AI, told India Today.

That is the gap no scoreboard shows. A benchmark can say how well a model writes code. It cannot say who answers for the decision when a model in a bank, school, or courtroom gets one wrong.

Chips are necessary. They are not the same as a model.

- Ends