Developers are using more models and moving to their own hardware

The single-model trend has reversed

According to data published by OpenRouter on November 2, customers spending at least $100 a month use a median of six model families. That is twice the figure recorded a year earlier. Among the top 10 percent of spenders, the number of model families rose from 7 to 14.

The change followed an earlier period of consolidation. The metric was still declining in the first half of 2025, reaching a low of 3.8 in June, before rising in every quarter and reaching 7.3 in October 2026. OpenRouter said this was the highest value it had recorded to date.

Higher usage comes with a broader choice of models. Based on October data, customers spending between $100 and $1,000 a month used models from a median of five labs, those spending between $1,000 and $10,000 used models from eight, and those spending more than $10,000 used models from 12 labs. Ninety percent of the highest-spending group used models from at least five labs. OpenRouter noted, however, that its data may be affected by selection bias, since users turn to a model aggregator specifically to access multiple models.

The leading model varies by task

Kilo's own usage dashboard shows a similar distribution within a single developer tool. The leading model depends on the task, and no model accounts for more than 30 percent of usage in any role.

Nvidia's Nemotron models lead in coding, planning and debugging. Review differs from the other tasks, with Anthropic and OpenAI's flagship models appearing in larger proportions. Claude Opus alone accounts for 11.5 percent of usage, while GPT models represent a further 14.4 percent.

Only one flagship model appears among the 10 most frequently used models for coding, gpt-6-luna, with a 1.4 percent share. Kilo's interpretation is that developers use cheaper models to write and fix code, then choose a more powerful model for identifying errors.

Revolut runs its default model on its own hardware

Revolut has set its self-hosted open model, GLM 5.3, as the default coding model for around 2,000 engineers. David Abitbol, the bank's head of AI operations, discussed the solution at the AI Engineer NYC event on October 13, and the presentation spread on X during the first week of November.

According to Abitbol, the initiative was driven by cost. Some engineers were spending between $1,000 and $10,000 a month on flagship models, with “not a very high return”. Revolut began running its own model with GLM 5.1 in April 2026 and currently uses GLM 5.3 on approximately 56 Nvidia B200 GPUs. The blended cost is around 30 cents per million tokens, which Abitbol said is one-tenth of the amount spent on Anthropic and OpenAI models.

In Revolut's measurements, self-hosting resulted in a 70 percent improvement in development cycle time and a 60 percent improvement in coding time after the switch to GLM 5.1. Abitbol said that because the cost was fixed, engineers were less likely to economize on tokens and also began using AI for Jira, Confluence and inbox tasks.

Kilo offers multiple configurations

According to Kilo, the two trends lead to a common operating model: a controlled, cheaper model handles routine tasks, while flagship models are used for steps where they deliver greater value. The company makes this available in several ways.

Users can choose a model for each agent role, selecting an open model for the Code and Debug roles and a flagship model for review. The settings can be configured on the Settings → Models page or in the kilo.jsonc file. Kilo Desktop also supports connecting local and self-hosted models.

Auto Efficient, the kilo-auto/efficient router, sends each request to the cheapest model that has proven sufficiently accurate for the task. Kilo's live usage data is available at kilo.ai/models, and the company said it changes continuously as new open models are released.