Modal vs Replicate Serverless GPU Pricing 2026

Quick Answer

Modal vs Replicate serverless GPU pricing 2026: Modal H100 $3.95/hr ($0.001097/sec), A100 80GB $2.50/hr; Replicate H100 $5.49/hr ($0.001525/sec), A100 80GB $5.04/hr. Modal cold-starts ~5s with GPU memory snapshots; Replicate ~60s+ on private deployments but unbilled cold starts on public models.

Last verified: Sep 16, 2026.

At a glance

• Modal H100: $3.95/GPU-hr ($0.001097/sec) | ~5s cold start
• Replicate H100: $5.49/GPU-hr ($0.001525/sec) | ~60s+ private, free on public
• Modal A100 80GB: $2.50/GPU-hr ($0.000694/sec)
• Replicate A100 80GB: $5.04/GPU-hr ($0.001400/sec)
• Modal B200: $6.25/GPU-hr ($0.001736/sec)
• Replicate 2x B200: $9.95/GPU-hr
• Modal H200: $4.54/GPU-hr
• Modal Starter: Free + $30/mo credits
• Modal Team: $250/mo + $100 credits
• Replicate: No free credits; pay-as-you-go

Why Modal vs Replicate matters for inference deployments in 2026

Modal and Replicate are the two serverless GPU clouds most often compared for inference workloads — they solve similar problems with different economics.

Modal is Python-native (decorator-style deployment, broader compute primitives) and is competitive on raw GPU rates. Replicate is inference-first with a Cog container format and a model marketplace at a price premium on raw GPU time. Both bill per-second; both scale to zero; both charge nothing for idle resources. The decision is about workload shape: Modal for custom code with frequent invocations; Replicate for spiky traffic on known open models (morphllm.com, June 2026).

The price gap is meaningful. Modal H100 at $3.95/GPU-hr is 28% cheaper than Replicate H100 at $5.49/GPU-hr. Modal A100 80GB at $2.50/GPU-hr is half the price of Replicate A100 80GB at $5.04/GPU-hr. For sustained production inference, Modal is materially cheaper. For spiky low-volume traffic on known open models, Replicate unbilled cold starts on public models deliver better dollar cost.

Modal per-GPU rate card

Modal publishes per-second GPU rates across 9 SKUs with no idle billing and explicit CPU/memory line items.

Verified August 2026 (modal.com/pricing):

• T4: $0.000164/sec = $0.59/GPU-hr
• L4: $0.000222/sec = $0.80/GPU-hr
• A10: $0.000306/sec = $1.10/GPU-hr
• L40S: $0.000542/sec = $1.95/GPU-hr
• A100 40GB: $0.000583/sec = $2.10/GPU-hr
• A100 80GB: $0.000694/sec = $2.50/GPU-hr
• H100: $0.001097/sec = $3.95/GPU-hr
• H200: $0.001261/sec = $4.54/GPU-hr
• B200: $0.001736/sec = $6.25/GPU-hr
• B300: $0.001972/sec = $7.10/GPU-hr

Modal also bills CPU at $0.0000131 per physical core per second and memory at $0.00000222 per GiB per second. There is no charge for idle resources. The minimum CPU is 0.25 cores.

Modal Starter is free with $30/month in credits — enough for hobby projects. Modal Team is $250/month plus compute with $100/month in credits and 1,000 containers + 50 GPU concurrency. Enterprise plans are custom.

Replicate per-GPU rate card

Replicate publishes per-second GPU rates with a split billing model: public models bill per-second of active runtime only; private models bill for setup, idle, and active time.

Verified August 2026 (replicate.com/pricing):

• CPU (small): $0.09/hr
• CPU: $0.36/hr
• T4: $0.81/GPU-hr ($0.000225/sec)
• L40S: $3.51/GPU-hr ($0.000975/sec)
• A100 80GB: $5.04/GPU-hr ($0.001400/sec)
• H100: $5.49/GPU-hr ($0.001525/sec)
• 2x A100: $10.08/GPU-hr
• 2x H100: $10.98/GPU-hr
• 2x B200: $9.95/GPU-hr

For language models, Replicate also bills per token (DeepSeek R1 at $3.75/M input, $10/M output). For image and video models, Replicate bills per output (per-image or per-second-of-video). These per-output pricing models can be cheaper than raw GPU time for spiky low-volume workloads where the GPU would otherwise sit idle.

Replicate has no free credit tier. It is pay-as-you-go from the first GPU-second. The unbilled cold start on public models is the closest equivalent to a free tier — for known open models with sporadic traffic, cold-start costs do not appear on the invoice.

Cold start time: Modal's key advantage

Modal cold-starts 10x faster than Replicate private deployments — a meaningful advantage for latency-sensitive scale-to-zero.

Modal containers boot in about 1 second, with GPU memory snapshots cutting a vLLM cold start from 45 seconds to roughly 5 seconds. Replicate private deployments commonly see 60+ second cold boots. For latency-sensitive workloads (chat applications, real-time inference), the cold-start gap is the deciding factor between the two platforms (morphllm.com, June 2026).

For spiky workloads where cold starts are tolerable (image generation, batch inference, scheduled jobs), the per-GPU rate difference dominates the cost calculation — Modal wins on price. For latency-sensitive workloads (chat, real-time APIs), Modal wins on both price and cold-start time. Replicate's only advantage is unbilled cold starts on public models, which matters for known open models with sporadic traffic.

Side-by-side Modal vs Replicate pricing

Break-even utilization math

For sustained traffic, Modal is cheaper than Replicate by 30-40%; for spiky traffic on known models, Replicate unbilled cold starts can win.

The break-even calculation for sustained inference traffic on H100:

• Modal H100 at $3.95/GPU-hr x 730 hrs = $2,884/mo for 24/7 allocation
• Replicate H100 at $5.49/GPU-hr x 730 hrs = $4,008/mo for 24/7 allocation
• Modal savings at full utilization: $1,124/mo

For spiky traffic (less than 10% utilization), both platforms bill only while serving requests. Modal's $0.001097/sec is cheaper than Replicate's $0.001525/sec for the same workload. Replicate's only advantage is unbilled cold starts on public models — a real cost saving if your workload invokes known open models with sporadic traffic.

For an 8-GPU H100 cluster running 30 days at 18 hours/day (4,320 GPU-hours): Modal at $3.95/GPU-hr = $17,064; Replicate at $5.49/GPU-hr = $23,717. Modal is $6,653 cheaper before storage, egress, and MLOps overhead (gpucloudcost.com, August 2026).

How to choose between Modal and Replicate in 2026

The decision hinges on traffic pattern, deployment surface preference, and budget.

1. Custom Python workloads with frequent invocations: Modal. Decorator-style deployment, lower per-second rates, faster cold starts, $30/mo free credits.
2. Latency-sensitive scale-to-zero chat applications: Modal. ~5s cold starts with memory snapshots make scale-to-zero usable instead of a billing trap.
3. Image and video generation with one-API-call deployment: Replicate. Catalog of pre-hosted models, per-image or per-second-of-video pricing is competitive with raw GPU rates.
4. Spiky low-volume calls on known open models: Replicate. Unbilled cold starts on public models deliver better dollar cost than Modal.
5. Agent code execution with sandboxing: Modal. Isolated sandboxes for agent code, which Replicate does not offer.
6. Cost-sensitive production inference: Modal. Lower per-second GPU rates (H100 $3.95/hr, A100 80GB $2.50/hr) plus pay-by-the-cycle billing make sustained traffic cheaper.

What enterprise buyers should do next

Three actions for organizations evaluating Modal vs Replicate in 2026.

1. Run a 30-day pilot on Modal Starter tier (free $30/mo credits). The free tier is enough to evaluate Modal's Python-native deployment surface, cold-start performance, and per-second billing model. If your workload exceeds $30/mo in compute, evaluate Modal Team ($250/mo + $100 credits).
2. Replicate the same workload on Replicate for comparison. For known open models, Replicate's unbilled cold starts and per-token pricing on language models deliver competitive unit economics. For custom code, Modal wins on every axis.
3. Model cold start time into your latency budget. Modal's ~5s cold start is acceptable for most chat and inference workloads; Replicate's ~60s+ private cold start is a barrier for real-time APIs. If your workload is latency-sensitive, Modal is the only viable choice.

What to watch next

Three near-term datapoints. First, Modal's B300-class pricing — current rates top out at B300 $7.10/GPU-hr; expect GB300 listings in late 2026 as Vera Rubin production ramps. Second, Replicate's acquisition by Cloudflare (announced June 2026, reported in devtoollab.com) — the integration could push Replicate rates lower as Cloudflare's network and capacity come into play. Third, Modal vs Replicate on Blackwell Ultra (B300) cold-start time — current Replicate cold starts on private B200 deployments can run 90s+; if Modal maintains its ~5s cold start advantage on B300, the gap widens materially (devtoollab.com, August 2026).