Celeris-1 - Fastest general-purpose AI, delivering 2,000+ tokens per second

Celeris-1 uses a new diffusion-based inference architecture instead of traditional autoregressive generation. Conventional models generate one token at a time where each token depends on the previous one, creating inherently sequential latency. Celeris's diffusion architecture enables parallel token generation, resulting in significantly lower latency.

Celeris provides an OpenAI-compatible API, so you only need to change the endpoint to inference.celeris.ai and the rest of your code stays the same. It also delivers lower latency and higher throughput—Celeris generates up to 24x more tokens per second than GPT-5 while maintaining comparable accuracy.

Celeris-1 achieves 75.9% accuracy on the MMLU-Pro benchmark, placing it near the frontier—only a few points below GPT-5 (81.9%) and other top models. It achieves this with a very short latency of 158 ms, while GPT-5 takes 2 seconds.

Celeris-1 is billed per token generated. Users pay only for the tokens the model outputs, so increased speed has no impact on pricing—generating 1,000 tokens costs the same whether it takes 100 ms or 1 second.