Sarvam AI's newly launched Sarvam Code, pitched as a cheaper India-hosted alternative to Claude Code and OpenAI's Codex, appears to be built directly on top of OpenAI's open-source Codex CLI. A technical analysis of the product's binary code found unmistakable fingerprints from Codex, including identical command structures, configuration schemas, and even an unscrubbed internal header string reading "x-openai-subagent." The discovery raises questions about transparency and attribution for a well-funded startup that has positioned itself as a sovereign AI alternative for Indian enterprises.

What Did the Technical Teardown Actually Find?

On August 27, 2026, independent researcher Ojasvi Yadav published a detailed analysis of Sarvam Code's 348-megabyte macOS binary using standard reverse-engineering tools. The teardown documented specific technical overlaps that go well beyond coincidence.

• Command Surface: Sarvam Code shares Codex's exact command set, including exec, resume, fork, and mcp, which are not standard naming conventions across the industry.
• Sandbox Modes: The product uses identical sandbox configurations, including read-only, workspace-write, and danger-full-access modes with the same approval-policy structure.
• Internal Strings: The binary contains the literal flag "--dangerously-bypass-approvals-and-sandbox" alongside Codex-linked components such as shell_environment_policy, apply_patch, and profile structure.
• OpenAI Header: An unscrubbed "x-openai-subagent" string remained in the compiled code, suggesting the binary was derived from or closely based on Codex's codebase rather than built independently.

Why Does This Matter for Developers and Enterprises?

Sarvam AI launched Sarvam Code at its Epoch 2026 developer conference in Bengaluru on July 30 and 31, marketing it as a planner-worker-verifier system that solves tasks at approximately $2 per completed task. The company positioned this as significantly cheaper than Claude Code and OpenAI Codex, which cost between $4.10 and $27.80 per task. The pitch emphasized that the product was India-hosted and built on Sarvam's own models, implying a ground-up engineering effort.

According to Pratyush Kumar, one of Sarvam's founders, the company framed the product as an alternative to licensing arrangements that Indian IT services firms currently maintain with Anthropic. Kumar stated that Sarvam Code was "hosted in India with our own models" and "very competitive," with the same number of tasks getting solved at lower cost. Once the Codex fingerprints became public, that messaging took on a different character.

Kumar

The concern isn't that Sarvam used open-source code. OpenAI's Codex repository is licensed under Apache 2.0, which explicitly permits forking, rebranding, and commercial use. The issue is one of transparency. Developers and enterprises making purchasing decisions based on the narrative of a homegrown Indian alternative may feel misled if the foundation is an existing open-source project that wasn't clearly disclosed.

What Has Sarvam Actually Built on Top of Codex?

The teardown is not an indictment of Sarvam's entire product. The company has added real engineering on top of the Codex base. According to the analysis, Sarvam Code includes document optical character recognition (OCR) for 22 scheduled Indian languages, a Guardian approval system, RLM mode, durable memory, and a self-improvement governor. These are substantive additions that reflect genuine product development work.

Sarvam AI itself is not a shoestring operation. In June 2026, the company announced the first close of a planned $300 million Series B funding round at a $1.5 billion valuation, with HCLTech investing $150 million for a 10.46% stake. The company has been selected under India's IndiaAI Mission and has scaled training across more than 4,096 Nvidia H100 GPUs on Yotta infrastructure. That level of resources and government backing means Sarvam operates under different expectations than a weekend GitHub fork.

What Does Apache 2.0 Licensing Actually Require?

OpenAI's Codex repository has accumulated over 100,000 stars on GitHub and has hundreds of contributors. Under Apache 2.0, licensees must preserve the original license and required notices, but they can use, fork, rebrand, and sell the software. The legal question is not whether Sarvam had the right to use Codex; it almost certainly did.

The practical question is whether Sarvam met the spirit of open-source attribution. The public launch materials describe Sarvam Code as a product built to route work across 100-billion, 700-billion, and 3-trillion parameter models, manage context, and optimize cost per completed task. There is no mention of OpenAI or Codex in the official positioning. For a company backed by $234 million in Series B funding and positioned as a sovereign AI alternative, that silence reads as a choice rather than an oversight.

How Should Developers Interpret This Finding?

As of August 27, 2026, neither Sarvam AI nor OpenAI has issued a public response to the teardown. The finding comes from a single independent researcher, not from OpenAI, a court filing, or a security firm. Major technology outlets have not yet covered the story. That context matters for proportionality.

However, the specific technical evidence is difficult to dismiss. The command names, sandbox modes, configuration schema, and internal header strings documented in the teardown are not generic patterns that multiple teams would independently converge on. They are specific design choices that point to a shared codebase.

The real test will come if Sarvam Code moves from early beta into production use within Indian IT services firms, banks, government projects, or regulated enterprise environments. Developers can live with open-source forks; they use them every day. What they typically object to is being sold independence and discovering inheritance later. The question Sarvam will need to answer is straightforward: what is Sarvam's work, and what came from OpenAI's open-source base? That transparency would have been easier to provide upfront.