Cast AI launches Kimchi Coding with production support

Cast AI has made its Kimchi Coding software generally available, moving the coding agent out of Early Access with production support.

The product is designed as an autonomous multi-model coding agent that routes work across different AI models rather than relying on a single commercial model for every task.

Cast AI is launching Kimchi Coding as companies weigh the cost of AI-assisted software development against concerns over data handling and budget control. Kimchi targets that pressure point: customers can run it inside their own virtual private cloud or use dedicated Nvidia B300 graphics processors on Cast AI's infrastructure.

The product includes hard spending caps, from individual application programming interface keys to entire organisations. It also automatically shuts down what Cast AI described as runaway agentic loops and provides a real-time dashboard that breaks down spending by developer, team and project.

Central to the launch is Cast AI's claim that a multi-model approach can cut the cost of coding assistants without reducing output quality. Kimchi routes each task based on complexity and cost, while feedback loops assess generated code and reduce token use.

In shadow-mode testing against a baseline that used only commercial models, Cast AI said Kimchi was 2.5 times cheaper and matched or exceeded quality on specification match and test pass rates. The company attributed the lower cost to routing more work to open-weight models and to self-hosted inference managed on its own infrastructure.

Customer evidence

Cast AI also used the launch to highlight Akamai as a production customer.

"Kimchi is fundamentally changing how our engineering team thinks about AI-assisted development," said Dekel Shavit, Senior Director of Engineering at Akamai.

"The multi-model approach keeps the quality of a single frontier model while cutting token costs dramatically. And the governance suite is the best we've seen in the market. Data sovereignty, cost observability, and budget control in one place. For an organization like Akamai, that makes a real difference," Shavit said.

The reference to governance reflects a wider issue in enterprise AI adoption: companies want more visibility into how coding tools consume computing resources and where data is processed. Cast AI said Kimchi was built with data sovereignty controls and can operate as a standalone deployment inside a customer environment.

Cast AI also said the service is ISO 27001 and SOC 2 Type II certified and complies with GDPR requirements. Those features are likely to matter to larger organisations in regulated sectors that have been cautious about placing source code and internal data into external AI systems.

Multi-model strategy

Laurent Gil, President and Co-Founder of Cast AI, said the economics of relying on a single top-tier model had become difficult for businesses trying to roll out AI coding tools across larger teams.

"We built Kimchi to give every developer frontier-quality AI coding without frontier-sized bills or data risk," Gil said.

"General availability means that enterprises are now using Kimchi Coding at scale. The economics of AI coding are broken when you rely on a single commercial model. Kimchi fixes that, and does it without asking teams to compromise on quality, security, or control," he said.

The launch also shows how infrastructure providers are moving further up the AI software stack. Cast AI built its business around Kubernetes automation and cloud cost management, and Kimchi runs on the same optimisation engine it uses for Kubernetes and graphics processor infrastructure.

Cast AI said it reached unicorn status after an investment from Pacific Alliance Ventures, the US corporate venture arm of Shinsegae Group. The company has named BMW, Cisco, FICO, HuggingFace and Swisscom among customers for its broader infrastructure business.

With Kimchi now generally available, Cast AI is entering a crowded market for AI coding tools with a proposition centred less on raw model access and more on cost management, deployment control and the ability to mix commercial and open-weight systems within one workflow.

In its shadow-mode evaluation, the company said that approach delivered a 2.5-times cost reduction while matching or beating a commercial-models-only setup on specification match and test pass rates.