Baseten vs Modal vs Replicate 2026: Cheapest Hosting

A separate comparison on this site already prices Replicate against fal and Modal on the flat per-second GPU rate. This one centers Baseten, the platform that comparison skipped, and reframes the question the way a founder actually asks it: to serve a custom model or an open LLM, is it cheaper to rent a dedicated GPU by the minute, rent serverless compute by the second, or pay per token? Those are three different billing units, and the cheapest one depends entirely on how busy your model is. Every rate below is derived transparently from the vendors' own unit prices.

1. Three hosting models, three billing units

Baseten, Modal, and Replicate are often listed together as GPU inference providers, but they are architecturally different products, and that difference drives the bill more than any single rate.

Platform	Product shape	Billing unit	Idle behavior
Baseten	Dedicated model deployments with autoscaling[2]	Per GPU-minute[1]	No charge at zero replicas; warm replicas billed[2]
Modal	Serverless functions on your container[3]	Per GPU-second[3]	Never pay for idle resources[3]
Replicate	Model marketplace plus private deployments[5]	Per GPU-second or per token[4]	Public: idle free. Private: idle billed[5]

Baseten runs your model as a dedicated deployment: you pick the GPU, set a minimum and maximum replica count, and the autoscaler moves between them^[2]. Modal runs your code as a serverless function that grabs a GPU only while executing and releases it the instant the function returns, billing by the second with no idle line at all^[3]. Replicate is a marketplace first: a public model is a shared, always-warm endpoint you call by the second, while a private model or dedicated deployment runs on your own hardware and bills for its full online lifecycle^[4][5]. The billing unit tells you where the cost lives, so read it before the rate.

2. GPU rate card, converted to per-hour

To compare a per-minute platform against two per-second ones, everything below is normalized to a per-hour figure, derived from each vendor's published unit price (per-minute times 60, or per-second times 3,600). Hardware naming and memory configurations differ slightly between vendors, so match the closest equivalent.

GPU	Baseten /hr	Modal /hr	Replicate /hr
H100 80GB	$6.50[1]	$3.95[3]	$5.49[4]
A100 80GB	$4.00[1]	$2.50[3]	$5.04[4]
L40S	not listed[1]	$1.95[3]	$3.51[4]
A10 / A10G	$1.21[1]	$1.10[3]	not listed[4]
L4	$0.85[1]	$0.80[3]	not listed[4]
T4	$0.63[1]	$0.59[3]	$0.81[4]
B200	$9.98[1]	$6.25[3]	not listed[4]

Modal is the cheapest raw GPU rate on every chip all three list, and by a wide margin at the top: its H100 is roughly 40 percent below Baseten's and 28 percent below Replicate's^[1][3][4]. Two caveats keep this from settling the question. Modal applies a region-selection multiplier of 1.5x to 1.75x when you pin your workload to specific regions, which erases much of that lead for latency-sensitive deployments^[3]. And Baseten's per-minute billing rounds finer against long idle gaps than a raw hourly rate implies, because it only meters the minutes a replica actually runs^[1][2]. The rate card ranks the platforms only for a GPU that stays busy. The next section is where the ranking can flip.

3. Cold starts and scale-to-zero

The per-hour rate prices a busy GPU. Most custom-model workloads are not busy; they are spiky, with long idle stretches between bursts. What each platform charges during those stretches, and how badly the wake-up hurts, is the axis that decides the real bill.

• Modal never charges for idle resources: the function holds a GPU only while running and releases it on return, so a model that sees one request an hour pays for seconds of compute, not the wait between^[3]. Cold starts happen on each wake as the container spins up.
• Baseten scales to zero on a minimum replica count of 0, and a deployment scaled to zero incurs no charges^[2]. The cost of that is a cold start on the next request, which can take minutes for large models, so Baseten recommends a minimum of 2 replicas in production to eliminate cold starts, at the price of paying for two warm GPUs continuously^[2].
• Replicate splits by model type. A public model is always warm in a shared pool, so setup and idle are free and you pay only active time, at the cost of occasional cold boots and queue waits when others hit the same model^[5]. A private model or dedicated deployment bills for the full online lifecycle: setup, idle, and active, unless you scale it to zero and accept the cold boots^[5].

The pattern is a triangle. Modal and a scale-to-zero Baseten deployment both pay near zero while idle but eat cold starts on wake. A warm Baseten deployment or a Replicate private model with a minimum replica above zero pays continuously but answers instantly. A Replicate public model outsources the idle cost to Replicate's shared pool, which is the cheapest idle story of all until the shared queue makes you wait. Baseten's edge in this triangle is its cold-start engineering, which is what makes scale-to-zero usable for interactive traffic rather than only batch jobs^[2].

4. Serving an open LLM: GPU-second vs per-token

For an open-weight LLM specifically, there is a second billing unit in play. Instead of renting a GPU and running the model, you can pay per token through a managed endpoint, where the platform absorbs the GPU, the batching, and the idle time.

Approach	Baseten	Modal	Replicate
Rent GPU, run model	Yes, per minute[1]	Yes, per second[3]	Yes, per second[4]
Managed per-token LLM	Model APIs: GPT OSS 120B $0.10/$0.50, DeepSeek V4 $1.74/$3.48 per 1M in/out[1]	None; you run the model[3]	Some models: DeepSeek R1 $3.75/M in, $10/M out[4]

The crossover between the two units is exact and derivable from Baseten's own prices. A dedicated H100 costs $6.50 per hour^[1]. Baseten's Model APIs bill GPT OSS 120B at $0.10 per 1M input and $0.50 per 1M output tokens^[1]. So one H100-hour buys the same dollars as about 13 million output tokens ($6.50 divided by $0.50 per 1M) or 65 million input tokens on the per-token endpoint. Serve fewer than that per GPU-hour and the managed endpoint is cheaper, because it charges nothing between requests; serve more and the dedicated GPU wins. Whether your model clears 13 million output tokens per H100-hour depends on batch size and sequence length, so measure throughput on your workload before deciding rather than assuming it.

Modal is deliberately absent from the per-token column: it sells raw compute, not managed model endpoints, so with Modal you always take the rent-a-GPU path and own the serving stack^[3]. Replicate meets in the middle, exposing per-token pricing on a curated set of language models while still letting you deploy the GPU version yourself^[4]. Baseten is the only one of the three that pairs cheap managed per-token LLM serving with a dedicated-GPU path on the same platform, which is the core of its pitch here^[1].

5. Worked example: an intermittent model

Take one custom model on a single H100, used intermittently: about 90 active GPU-hours per month, in bursts, against an average month of 730 hours. Platform compute cost only:

Configuration	Monthly	Cold starts?
Modal serverless (per second, no idle)	~$356 (90 × $3.95)[3]	Yes, on each wake
Baseten min 0 (scale-to-zero)	~$585 (90 × $6.50)[1][2]	Yes, can take minutes[2]
Replicate private, warm (min 1)	~$4,008 (730 × $5.49)[4][5]	No, but idle is billed[5]
Baseten min 2 (always warm)	~$9,490 (2 × 730 × $6.50)[1][2]	No[2]

For this intermittent shape, Modal's active-only per-second billing is the cheapest by a clear margin, at roughly $356 against Baseten scale-to-zero's $585, because Modal charges nothing for idle and its H100 rate is lower^[1][3]. The two always-warm rows show the cost of buying away cold starts: a Replicate private deployment kept warm bills every idle hour at $5.49, and a Baseten production setup with two warm replicas prices reliability at nearly $9,500 a month^[2][4][5]. Flip the workload to a busy GPU near 730 active hours and the ranking narrows to the raw rate card, where Modal still leads but the gap to Baseten and Replicate is what you would divide across your traffic. Fold whichever path fits into your AI stack cost calculator before committing.

6. Which is cheapest for you?

• Lowest raw GPU rate: Modal, cheapest per hour on every chip all three list, and cheapest for spiky workloads thanks to no idle charge^[3].
• Dedicated deployment with usable scale-to-zero: Baseten, which pairs per-minute billing and no charge at zero replicas with cold-start engineering built for interactive traffic^[1][2].
• Cheap managed per-token LLM serving: Baseten Model APIs, from $0.10/$0.50 per 1M on GPT OSS 120B, when a dedicated GPU would sit under-utilized^[1].
• Simplest deploy, no infra work: Replicate public models, a single API call at the highest per-second rate but zero idle cost^[4][5].
• Free credits to prototype: Modal, $30/mo on Starter and $100/mo on the $250 Team plan^[3].
• Watch the idle line: a Replicate private deployment or a warm Baseten min-replica setup can cost multiples of a scale-to-zero configuration for the same active work^[2][5].

Re-verify each pricing page before you commit: GPU rates move with supply and demand, and per-token model catalogs change monthly. For the hosting layer under a full-stack app rather than the model itself, see Railway vs Render vs Fly pricing, then run the numbers through the AI stack cost calculator.

All rate figures verified against official pricing pages and docs as of 2026-07-17.

Frequently asked questions

Which is cheapest for hosting a model: Baseten, Modal, or Replicate?

On the raw GPU rate, Modal is cheapest across every comparable chip. Its H100 works out to about $3.95/hour ($0.001097/s), against Replicate's $5.49/hour ($0.001525/s) and Baseten's $6.50/hour ($0.10833/min), all verified on the vendors' pricing pages on 2026-07-17. But the rate is not the whole bill. Baseten and Modal both scale to zero and charge nothing while idle, so for an intermittently used model they can undercut a Replicate private deployment that bills idle time. Pick on your utilization pattern, not the sticker rate: Modal for the lowest per-hour compute, Baseten when you need dedicated autoscaling with fast cold starts, Replicate when single-call deploy convenience matters more than the rate.

Does Baseten charge for idle GPU time?

Not when a deployment is scaled to zero: Baseten's autoscaling docs state that a deployment scaled to zero replicas incurs no charges. The tradeoff is that the next request triggers a cold start, which can take minutes for large models. To remove cold starts, Baseten's production guidance is to set the minimum replica count to at least 2, but those warm replicas then bill at the per-minute GPU rate around the clock whether or not they serve traffic. So Baseten does not charge for idle in the scale-to-zero configuration, and does charge for the warm replicas you keep online to avoid cold starts. All figures verified on 2026-07-17.

Should I serve an open LLM on a dedicated GPU or a per-token API?

It comes down to utilization. Renting a dedicated H100 on Baseten costs $6.50/hour no matter how many tokens flow through it. Baseten's own Model APIs bill GPT OSS 120B at $0.10 per 1M input and $0.50 per 1M output tokens, so one H100-hour equals the per-token cost of roughly 13 million output tokens. Serve fewer than that per GPU-hour and the managed per-token endpoint is cheaper because you pay nothing between requests; serve more and the dedicated GPU wins. Whether your model clears that throughput on an H100 depends on batch size and sequence length, so benchmark your workload before committing. Modal has no per-token product, so with Modal you always rent the GPU and run the model yourself.

Do Baseten, Modal, and Replicate offer free credits?

Modal is the clearest: its Starter plan is free and includes $30 in monthly credits, and its Team plan ($250/month) includes $100 in monthly credits, verified on Modal's pricing page on 2026-07-17. Baseten states that new accounts come with credits to experiment with the platform, though it does not publish a recurring monthly figure. Replicate's pricing page does not advertise free credits, so budget for usage from the first run. For a developer who wants to prototype GPU inference without an upfront bill, Modal's recurring monthly credits are the strongest starting position of the three.