Modal pricing: every GPU rate, free credits, and worked costs

The GPU rate card

Read from the Modal pricing page in August 2026. The per-second column is Modal's published figure; the hourly column is that number times 3,600 and is ours, provided because every competitor quotes hourly and comparison is otherwise impossible.

CPU, memory, and storage

These are billed in addition to the GPU, and forgetting them is the most common way a Modal estimate comes in low.

Plans and free credits

The Starter allowance is the standout number in this whole comparison: $30 renewed every month, on a plan that costs nothing. Fireworks gives $1 once. Together publishes nothing. What $30 a month buys, at August 2026 rates:

The Team plan is worth reading carefully. It costs $250 and includes $100 of credits, so the true platform fee is $150 per month. What you buy for that is concurrency: 50 simultaneous GPUs against 10, and 5,000 containers against 100. If your workload is a large fan-out (embedding a corpus, running an eval sweep) the concurrency limit binds long before the credit does.

Worked example: a fine-tuning run

A LoRA fine-tune on a single H100, four hours of wall clock, with 8 CPU cores and 64 GiB of memory attached.

Runtime: 4 hours = 14,400 seconds

GPU    14,400 s  x  $0.001097            = $15.80
CPU    14,400 s  x  8 cores x $0.0000131 =  $1.51
Memory 14,400 s  x  64 GiB  x $0.00000222=  $2.05
                                    ------------
                                     TOTAL  $19.35

Two comparisons make that number meaningful. Fireworks managed training would bill the same job per million training tokens instead: $0.50 per million for LoRA SFT on a model up to 16B, so a 30-million-token run is $15.00 with no infrastructure to think about. And a reserved H100 from Together at $3.19 per hour is $12.76 for four hours, but only if you already have the reservation and have amortized the other 726 hours in the month.

Worked example: an embedding batch

Embedding a corpus is the workload Modal is best at, because it is bursty, parallel, and finished when it is finished.

Shape: 20 containers x 540 s each = 10,800 GPU-seconds on L4
Each container: 4 CPU cores, 16 GiB

GPU    10,800 s x $0.000222              = $2.40
CPU    10,800 s x 4 cores x $0.0000131   = $0.57
Memory 10,800 s x 16 GiB  x $0.00000222  = $0.38
                                    -----------
                                     TOTAL $3.35

Wall clock: 9 minutes (the containers ran in parallel)
Covered by the free $30 monthly credit, about 9 times over.

The same job on a single L4 you rented would take three hours of wall clock and cost about the same, because per-second billing means parallelism is free. That is the actual pitch: you are buying the scheduler, not the discount.

Modal against renting raw GPUs

The comparison people want is Modal versus just renting a GPU. Here it is with real numbers.

Reserved H100 (Together, 181+ days): $3.19/hr x 730 hr = $2,328.70/month
Modal H100:                          $3.95/hr of ACTUAL runtime

Break-even runtime = $2,328.70 / $3.95 = ~590 hours/month
                   = ~81% of a 730-hour month

Below ~80% utilization, Modal is cheaper.
Above it, buy the reservation.

Questions people ask

Keep reading

Continue through Inference providers with Cheapest LLM API: the honest list, by capability class.