Mythos and Fable lesson for India: Sovereign AI is no longer optional

On June 12, 2026, the US government sent Anthropic a letter that changed the terms of the debate around sovereign AI. Claude’s Fable 5 and Mythos 5 were suspended for foreign nationals everywhere in the world, including Anthropic’s own non-American employees. The stated reason was national security. Anthropic called it a misunderstanding and said it was working to restore access. Within hours, the models were unavailable globally.

That single action exposed a hard fact about modern AI dependence. The most advanced systems in the world can be turned off for most of the planet by one government, with no warning and no appeal. For countries still treating sovereign AI as a long-term ambition rather than a near-term necessity, the argument has already been settled by events.

More from Opinion

India’s Response is Sarvam AI:

A promising foundational-model company founded in 2023 by Vivek Raghavan and Pratyush Kumar, Sarvam has developed models trained from scratch in India and designed to support all 22 scheduled Indian languages. That achievement goes beyond language localisation. Indic languages remain under-represented across much of frontier AI, from training data and evaluation benchmarks to real-world deployment. Building language diversity into a foundational model is therefore a meaningful form of inclusion. A national AI capability that cannot effectively serve large sections of its own population falls short of its purpose. By investing in broad linguistic coverage, Sarvam is addressing one of the most important gaps in contemporary AI and aligning technological capability with India's social and cultural realities.

Sarvam’s most ambitious release, a 105-billion-parameter model, uses a Mixture of Experts architecture that activates roughly 9 billion parameters during inference. The design improves compute efficiency and reflects a pragmatic approach to building large-scale models under resource constraints. The company has also reported encouraging results in mathematical reasoning and strong performance across Indic language benchmarks. The architecture appears technically sound, and the ambition is aligned with India’s broader objective of building indigenous AI capability.

As Sarvam evolves from a promising AI company into a component of national digital infrastructure, greater independent validation would help strengthen confidence in its achievements. Most benchmark results available today are company-reported, which is common for early-stage model releases. Participation in widely recognised public leaderboards and third-party evaluations would provide broader visibility into performance and allow the wider research community to assess progress using shared standards. Such transparency would not diminish Sarvam’s accomplishments; it would reinforce them and help build deeper public trust in a project that carries national significance.

There is also a governance problem hidden inside the word “sovereign”. Sarvam is a private company. Public funding flows into the project, but public accountability does not automatically flow back out. That creates a familiar risk: a national capability with private control, private discretion, and limited institutional oversight. Sovereignty in that form is incomplete. It gives the country access to the system, but not necessarily authority over it. National capability held behind private keys is not the same thing as national control.

None of this diminishes what Sarvam is building. It is doing serious technical work under tighter compute constraints than the frontier labs enjoy. Its language agenda is relevant to India’s linguistic reality. Its engineering choices suggest a practical understanding of efficiency rather than a race for scale alone. But the next phase cannot be only about model launches. It has to be about institutions.

India needs independent benchmarking, transparent governance, and legal frameworks that define data rights, audit rights, and state oversight. A sovereign AI programme without those elements risks reproducing the same dependency it is meant to escape, only with local ownership replacing foreign ownership. The label changes. The structure does not.

The Fable 5 shutdown showed how dependency behaves in practice. India has noticed. The real question is whether Sarvam AI becomes a foundation for sovereignty or merely another layer of dependence with a different flag on the door.

AI sovereignty ultimately depends as much on governance as it does on technology.

(Sreejith Sreedharan is a technology analyst and author. The views expressed in this piece are personal and solely those of the author. They do not necessarily reflect Firstpost’s views.)