Revolut says it cut AI coding inference costs by ~90% in less than a year by making self-hosted GLM models its default over commercial tools like Claude Code

The cost of using AI coding assistants falls by about 90% if you switch to self-hosted open-weight models and optimize their deployment.

British fintech company Revolut has drastically reduced its AI coding inference costs by approximately 90% within a year. This was achieved by making self-hosted GLM models its default choice over commercial tools like Anthropic's Claude Code. Revolut's ops and infrastructure lead, David Abitbol, shared that this transition led to a 60% improvement in coding time and a 70% improvement in lead time for its roughly 2,000 engineers. Initially, the company experimented with AI tools like Gemini and Claude Code, but high costs and limited token usage prompted the investigation into self-hosting. By utilizing GLM 5.1 and later GLM 5.3 on 56 Nvidia B200 GPUs, and optimizing parameters, they found that cheaper tokens enabled engineers to leverage AI for tasks such as debugging and managing Jira tickets more effectively. Abitbol also noted that extremely long context windows, like one million tokens, are not as beneficial as a well-tuned, shorter window, finding 256,000 tokens to be the optimal range for engineer productivity.

Revolut's strategy highlights a growing trend for enterprises to move away from expensive commercial AI models towards self-hosted, open-weight alternatives to manage costs and gain control. This shift is enabled by advancements in GPU hardware like Nvidia's B200 and optimization techniques, which make running powerful models like GLM feasible and economical at scale. The company's success suggests that the perceived value of "long context" windows may be overstated, and that performance gains are more attainable through careful tuning of hyperparameters and quantization, impacting how organizations approach AI infrastructure investment.

This move by Revolut demonstrates that significant cost savings and efficiency gains are achievable, potentially setting a new benchmark for AI adoption in software development. It implies that companies relying on commercial AI coding tools might be overpaying and could explore similar self-hosting strategies. The focus on optimizing inference, particularly for tasks beyond pure code generation like debugging and ticket management, indicates a broader application of AI in streamlining development workflows and warrants attention from AI providers and infrastructure vendors.