// pricing

You pay for the GPU your model needs.

Every model runs on its own GPU. A bigger model needs a bigger GPU, and a bigger GPU costs more per hour. Pilot picks the smallest one that fits, and your per-token price follows from that GPU’s hourly cost. No subscription, no minimum plan.

// free pre-flight

Try any model for free

Every launch is a free Pre-flight today: the model stays up for 15 minutes or 25 requests, whichever comes first, then it is torn down. Up to 2 Pre-flights per account. Paid, always-on launches are opening soon.

// dedicated launches

Price by model size

A dedicated launch is billed from your wallet: a one-time start fee, then per-token prices set by the GPU. A model that fits a small GPU costs far less per token than one that needs a large GPU.

MODEL SIZE	GPU MEMORY	TYPICAL GPUS	GPU $/HR	INPUT / 1M TOKENS	OUTPUT / 1M TOKENS	START FEE
up to ~4B	16 GB	RTX A4000 / A4500 / 4000 Ada	$0.58	$0.30	$1.21	$0.021
~4B – 9B	24 GB	L4 / RTX A5000 / RTX 4090	$0.69 – $1.10	$0.36 – $0.57	$1.44 – $2.29	$0.025 – $0.040
~9B – 10B	32 GB	RTX 5090 / RTX PRO 4500	$1.15 – $1.58	$0.60 – $0.82	$2.40 – $3.29	$0.042 – $0.057
10B, long context	48 GB	A40 / L40S / RTX 6000 Ada	$1.22 – $1.75	$0.64 – $0.91	$2.54 – $3.65	$0.044 – $0.063

// estimates for 16-bit weights; quantized models can fit a smaller GPU. Pilot makes the final pick at launch time.

// how it is calculated

From GPU hour to token price

Output tokens = GPU $/hr ÷ (200 tokens/s × 3600) × 1,000,000 × 1.5.
Input tokens cost a quarter of output tokens, since reading a prompt is much cheaper than generating a reply.
Start fee is what the GPU costs while your model cold-starts (about 130 seconds of its hourly price). It is charged once, when the launch succeeds.
GPUs scale to zero when idle, so you are not billed for hours nobody calls the model. A model unused for 7 days is deleted.

Launch your first model free

get started →