British fintech company Revolut moved around 2,000 developers to the Chinese AI model GLM, making it the primary programming tool in place of Claude Code and OpenAI Codex. The company says the switch cut AI inference costs by around 90% and accelerated software development.
Revolut is not alone in choosing cheaper models alongside products from OpenAI and Anthropic. AT&T, Airbnb and others are pursuing similar strategies. A trend appears to be emerging: businesses do not necessarily need the most expensive and powerful models for everyday tasks.
How Revolut cut AI costs tenfold
David Abitbol, Revolut's head of AI infrastructure, described the results at AI Engineer in New York on October 13, 2026. The story gained wider attention in November after The Register reported it.
Revolut initially used commercial AI tools including Gemini, Claude Code and OpenAI Codex. Costs grew quickly: Abitbol said some developers spent $1,000–$10,000 a month on AI without always showing a comparable productivity gain.
The company imposed strict token limits, controlling budgets but encouraging engineers to save requests for only the hardest tasks. In April 2026, Revolut began running Z.ai's GLM-5.1 on its own infrastructure, later moving to GLM-5.3 instead of paying a third-party API for requests.
According to Revolut's internal figures:
- Inference costs fell by 90% compared with the previous commercial solutions.
- Coding time improved by 60%.
- Lead time improved by 70%: this measures the time from starting a task to changes being ready for deployment.
- The estimated average processing cost was about $0.30 per million tokens.
Revolut uses 56 Nvidia B200 GPUs to serve roughly 2,000 developers. Infrastructure optimizations brought time to first token below three seconds and generation speed to around 200 tokens per second.
Revolut has not abandoned Claude and Codex entirely. More powerful commercial models remain available for certain difficult tasks, and image requests can be routed to other models.
There is an important qualification: Revolut has not published a detailed total cost of ownership calculation including GPU purchases and specialist staff. The claimed savings concern model usage, not necessarily all AI spending.
Beyond Revolut: other companies using cheaper models
Several large companies have reported using cheaper models in place of more expensive alternatives, although the scale of each change differs.
AT&T: 40% of requests moved to open models
AT&T uses around 45 billion tokens a day and routes requests intelligently among models to reduce costs. By August 2026, open-weight models, including Nvidia Nemotron and Google Gemma, handled about 40% of employees' internal AI requests.
AT&T says optimization can reduce AI costs by up to 90% in certain scenarios, saving millions of dollars. It plans to increase the open-model share to 60–70%. Expensive commercial models remain available where their extra capabilities justify the cost.
Sources: AT&T, The Information
Airbnb: China's Qwen for customer support
Airbnb uses Alibaba's Qwen in its own customer-support AI assistant. CEO Brian Chesky has said it handles tasks well, works quickly and costs less than many American alternatives.
Airbnb uses a mix of models from different providers, including OpenAI. In 2025 Chesky said it preferred cheaper solutions for many production tasks. US lawmakers raised data-security questions about the Chinese model; Chesky explained that Airbnb runs open models itself and does not send data to Chinese developers.
Sources: Bloomberg, Silicon UK
Cursor: its own model based on China's Kimi
Cursor took a different route, building Composer 2 on Kimi K2.5 from China's Moonshot AI rather than merely switching to a cheaper API. It continued training the model and optimized it for programming; the same open base model was later used for Composer 2.5.
Standard Composer 2.5 costs $0.50 per million input tokens and $2.50 per million output tokens, less than several flagship models. Cursor has not completely replaced Claude, but it can offer its own inexpensive alternative for everyday development.
Source: Cursor's official blog
Lindy: a 90% saving by moving from Claude to DeepSeek
In June 2026, AI assistant developer Lindy reported moving most of its managed model traffic from Claude Sonnet and Gemini to China's DeepSeek V4 Flash. Lindy uses AI for emails, calendars, meeting preparation and other office tasks. Inference costs on migrated requests fell by around 90%.
Migration required testing answer quality, adapting prompts and moving users gradually. Good test results did not always translate into real-world performance: Kimi K2.5 passed preliminary evaluation but proved less suitable in practice. Claude remained available for tasks requiring greater capability.
Source: Lindy's engineering blog
Tinder: rising AI costs prompted a search for alternatives
Tinder also began routing some internal employee requests to open-weight models. CTO Vinay Kuruvilla said its annualized AI spending rose from about $1 million in January to $10 million by July 2026. To contain costs, Tinder started sending some requests from nontechnical employees to cheaper models. Unlike Revolut, this is a partial redistribution of traffic, not a company-wide replacement.
Source: analysis of Financial Times reporting
How widespread is the move to cheaper AI?
Industry statistics accompany these corporate examples. According to Menlo Ventures' State of Enterprise AI 2026, open models grew as a share of the enterprise usage studied:
| Metric | 2025 | 2026 |
|---|---|---|
| Open models' share of tokens | 11% | 19% |
| Open models' share of spending | 6% | 4% |
| Flagship models' share of spending | 35% | 63% |
Open models process nearly one in five tokens but receive only 4% of model spending in the segment studied. Menlo says their average price per million tokens fell about 3.3-fold over the year to $0.13, versus around $0.70 for closed models.
Chinese developers account for about 89% of spending in the open-model segment, led by Kimi, DeepSeek and GLM. Meanwhile, flagship models' share of enterprise spending rose from 35% to 63%.
Businesses are buying more inexpensive models for high-volume tasks and more expensive ones for the hardest work. It is too early to speak of a mass abandonment of Claude or OpenAI.
Why cheaper models could change the AI market
Revolut's experience suggests token costs can affect both budgets and productivity. When requests are expensive, companies restrict usage and employees must decide which tasks deserve the available tokens. Lower costs can broaden use to debugging, documentation, emails, analysis and other everyday work.
A cheaper model need not beat Claude or GPT overall: it needs to handle a specific task reliably for substantially less. OpenAI and Anthropic may retain an advantage on the hardest tasks, but expensive models become harder to justify where cheaper alternatives suffice.
The central trend is hybrid AI use, rather than a complete replacement of American models with Chinese ones: inexpensive models handle routine volume, while flagship models are used when their added capabilities are needed. Revolut's experience suggests the economic effect can be substantial.