What does FP8 mean?
FP8 is an 8-bit floating-point format. Providers can use it for parts of model inference to reduce memory use and, with suitable hardware and implementation, improve throughput. FP16 and BF16 are 16-bit formats.
Serving systems can combine FP8 with other precisions for different operations.
Does it change answer quality?
Lower precision can introduce numerical differences. Their effect depends on the model, quantization method and serving implementation.
Try the same representative tasks on each route. For coding, compare passing tests and usable changes alongside cost and latency. Keep the model version, prompts and generation settings consistent.
How to read our labels
Our comparison shows FP8 when the listed provider publishes that precision. “Precision not disclosed” means the linked provider source does not specify its serving precision.
For OpenRouter, the label and quote belong to the named route. Other routes may use different precision and prices. Follow each provider’s source link for its current listing.
What about cached input?
Cache reads concern reuse of input computation. FP8 concerns how numbers are represented. A request may use both. Compare the regular input, cache-read and output prices separately. Read About Cache →
Further reading
NVIDIA TensorRT · Working with Quantized Types ↗
Independent explainer · Reviewed November 5, 2026.