Experiential Labs bets your inference spend should buy a model you own

Experiential Labs, a YC-backed AI infrastructure startup, launched in late August 2026 with an open-source model gateway that routes agent workloads across hosted, BYOK, self-hosted and marketplace models through one OpenAI-compatible API. Its Show HN on 2026-08-27 drew 222 points and 47 comments, and the company's pitch is deliberately contrarian: instead of charging markup on tokens like OpenRouter, it takes zero markup and uses customer traffic to train a specialized model the customer owns.

The co-founders are Kion, who earned an ML PhD at Georgia Tech and built simulation systems for autonomous vehicles at Waabi, and Silen, who grew AutoGPT to 160,000 GitHub stars and worked across continual learning, evaluations, interpretability and AI for science. The company shipped OpenClaas (Continual Learning as a Service) in March 2026 before open-sourcing the gateway, whose Rust-native engine handles streaming formats, tool calls, rate limits and provider quirks with under 1 ms overhead for BYOK requests.

The learning loop is the differentiator: standardized telemetry traces are mined into simulations of the customer's tasks, failures and edge cases; an LLM judge scores simulated rollouts; and a classifier trained on prompt embeddings picks the optimal model per request. Customers who opt in get a small model trained against that simulation, own the weights outright, and keep the learned behavior when a stronger base model appears — while hard requests still fall through to frontier models.

The company guarantees at least 50% lower cost at the same or better quality, backed by an SLA, and reports up to 97% cost reductions and +50% quality on some tasks (company claims; independent benchmarks have not been published). As of 2026-09-06 the gateway is open source with a hosted platform live, and no customers, revenue or usage figures have been disclosed.