Cost per million tokens
Cost per million tokens is the price an LLM provider charges to process one million tokens, with separate rates for input and output tokens.
On this page
Your total bill depends on these rates and the tokens each request uses. Teams that rent GPUs by the hour convert their bill into cost per million tokens to compare it with API prices.
Inference providers list several costs per million tokens per model
Providers quote rates for uncached input (the prompt), output (generated tokens) and cache reads. Output usually costs more because input processing can run in parallel, while generation is typically sequential. For more on why input processing and generation differ, read Parasail's guide to prefill vs. decode in LLM inference. Cache reads are usually discounted and require a matching prefix served from cache.
Parasail's pricing page lists Kimi K3 at a cost per million tokens of $3.00 for input, $15.00 for output and $0.30 for cache reads. At those rates, a request with 10,000 uncached input tokens and 1,000 output tokens costs $0.045.
A blended price combines the separate rates into one cost per million tokens by assuming a token mix. Artificial Analysis assumes seven cache-hit tokens and two uncached input tokens per output token, which puts Kimi K3 at $2.31. A 2:1 input-to-output mix with no cache hits blends the same rates to $7.00.
Cost per task is cost per million tokens multiplied by millions of tokens used
To calculate cost per task, multiply each billed token category's usage in millions by its rate, then sum across the task. Lower rates can still cost more when a model generates more reasoning tokens.
Parasail's comparison of Kimi K3 and GPT-5.6 Sol quotes K3 at $3 per million input tokens and $15 per million output tokens as of its October 2, 2026 update. Sol's promotional rates for prompts up to 272,000 input tokens are $4 per million input tokens and $20 per million output tokens.
At maximum reasoning effort, the Artificial Analysis Intelligence Index results average about 48,000 output tokens per task for K3 and 29,000 for Sol at the time the comparison was updated. At those output rates, K3's output charges come to about $0.72 per task and Sol's to about $0.58. Including input and cache charges, the comparison reports weighted total costs of $2.00 per task for K3 and $1.99 for Sol.
For ticket tagging or extraction, compare each model's billed tokens of each type on representative tasks before you compare cost per million tokens.
On GPU-hour billing, throughput and utilization set cost per million tokens
Rented GPUs and dedicated endpoints are usually billed by the GPU-hour, whatever their token count. Their cost per million tokens is the hourly rate divided by the millions of tokens the GPU produces in that hour. The effective rate shows whether dedicated GPUs cost less than per-token pricing at your throughput and utilization.
Throughput sets how many tokens a busy GPU produces. A GPU rented at $3 per hour that produces 1,000 tokens per second generates 3.6 million tokens an hour, a cost per million tokens of $0.83. At 2,000 tokens per second, cost per million tokens falls to $0.42.
Utilization sets how much of each paid hour the GPU spends serving requests. In this example, 25% utilization means serving at 1,000 output tokens per second for a quarter of the paid hour. Over the paid hour, the GPU produces 0.9 million tokens, a cost per million tokens of $3.33.
For eligible models, Parasail’s Elastic Endpoints offer dedicated GPU performance with per-token billing and no charges for idle GPUs. To compare per-token and GPU-hour billing for your workload, read Parasail's post on serverless vs. dedicated inference.