Varion

Varion is an OpenAI-compatible proxy that sits in front of LLM providers and reduces token usage before a request goes out. It compacts conversation history, prunes unused tool schemas, and caches exact-match duplicate requests, while passing through required context and instructions unchanged.

Integration is a base URL and key swap for any OpenAI SDK client. Varion adds a benchmark header on responses so teams can verify actual token savings rather than take the reduction on faith, plus a browser-based Test Lab and CLI/Python/Node clients for trying it before wiring it into an app.

The company also runs Varion Compute, a metered inference-efficiency service in commercial early access: it runs a workload on NVIDIA A40 GPUs and reports measured energy and throughput deltas against an unoptimized baseline, with validated results published per model (36.4% lower energy per token and 131.4% higher throughput on Qwen 2.5 1.5B, smaller models tested separately). 10 free GPU-hours for metered testing, no card required.

Pricing for the proxy is prepaid token-processing packages from EUR 19/month (Starter, 10M tokens) through Growth, Business, and Scale tiers. Provider costs are billed separately by the underlying LLM provider. New verified users get a free processing allowance to test with.