Last updated: July 17, 2026

Baseten vs Modal vs Replicate for one warm inference endpoint

Source note: this is a mixed, mostly source-backed HostFleet comparison built from the providers’ current pricing and deployment docs plus HostFleet’s local provider notes. The sourced layer covers pricing units, scaling defaults, cold-start behavior, queueing, billing semantics, rate limits, and retention caveats. The estimate layer is narrow and explicit: 30-day warm-endpoint cost examples derived from the published GPU rates.

If you are comparing Baseten vs Modal vs Replicate for a production endpoint, the real question is not which vendor can run inference. All three can. The real question is what happens when the endpoint must stay ready instead of sleeping at zero.

This piece stays narrow on purpose:

• about one or a few always-available inference endpoints
• not a benchmark shootout
• not a model-token pricing roundup
• not a generic serverless GPU comparison

If you need the broader market map first, start with Every serverless GPU host compared. If your main question is still endpoint ownership rather than warm-cost math, pair this with RunPod vs Modal vs Replicate for shipping a small inference API. For the single-provider context, keep Baseten for AI inference APIs and jobs, Modal for AI inference APIs and jobs, and Replicate for AI inference APIs and jobs beside this. If your workload may not need a dedicated GPU endpoint at all, Cloudflare Workers AI vs self-hosted GPU is the right contrast.

The short answer

My practical verdict is simple: Modal wins the warm-floor math, Baseten wins if managed dedicated inference is the point, and Replicate only looks cheap while you stay in its shared or prediction-first product shape.

Product shape matters before the GPU line item

A shallow comparison misses the product shape difference:

• Modal is a Python application platform with GPUs attached. You can keep containers warm with min_containers, stretch idle life with scaledown_window, and ship a custom API without buying a more opinionated inference platform.
• Baseten is a managed dedicated inference platform. The docs are built around deployments, replica policy, and production posture rather than app-first developer ergonomics.
• Replicate is a model and prediction product first. Public models feel cheap because setup and idle time are free, but deployments flip into normal dedicated-infrastructure billing.

That difference is why one headline number never answers the buying question.

Estimate: what one warm endpoint roughly costs for 30 days

These are estimate examples, not vendor quotes. Assumptions:

• 30-day month
• one continuously warm unit
• GPU price only
• no storage, egress, or extra platform charges
• Baseten from published per-minute rates
• Modal and Replicate from published per-second rates
• Replicate public-model pricing is excluded because it is a different product shape

Those estimates come from the current published rates checked on 2026-07-17:

• Modal T4 $0.000164/s and A100 80 GB $0.000694/s
• Baseten T4 $0.01052/min and A100 80 GB $0.06667/min
• Replicate T4 $0.000225/s and A100 80 GB $0.001400/s

The table is not saying the products are identical. It is saying that once you force each platform into a warm dedicated-serving shape, the bill differences become visible very quickly.

What the docs say about sleeping and waking up

Modal

Modal’s docs say containers idle for up to 60 seconds by default before shutdown. The scaling guide exposes the core warm-capacity knobs directly:

• min_containers
• max_containers
• scaledown_window

The same guide also states per-function limits of 2,000 pending inputs and 25,000 total inputs, with higher pending limits for spawned async jobs. In practice, Modal is friendly when you want to buy a little warmth without turning the whole service into a permanently on box.

Baseten

Baseten’s docs are explicit that the default autoscaling posture is conservative:

• min_replica defaults to 0
• max_replica defaults to 1
• the docs say that default max_replica of 1 effectively disables autoscaling
• the docs also recommend at least 2 minimum replicas for production if you want to avoid cold starts and add redundancy

Baseten’s billing docs are also unusually clear:

• usage is metered by the minute
• a replica is billed for every minute it is observed as up
• cold starts and model load are billed
• idle warm replicas are billed
• scaled-to-zero idle time is not billed

That makes Baseten coherent, but not magically cheap. The platform gets expensive the moment production readiness means keeping replicas awake.

Replicate

Replicate’s billing docs split the product in two:

• for public models, you only pay for active processing time; setup and idle time are free
• for private models and deployments, you pay for setup, idle, and active time

That is the biggest reason Replicate feels cheap in demos and expensive when converted into a dedicated warm endpoint. The product was not originally optimized around your team owning a permanently warm custom-serving surface.

Replicate also carries two operational caveats that matter for production apps:

• default API rate limits are 600 prediction-create requests per minute and 3000 requests per minute for other endpoints
• API-created prediction inputs, outputs, files, and logs are deleted after 1 hour by default

Those are not deal-breakers, but they are reminders that Replicate remains prediction-shaped even when you buy more control.

Where each one actually wins

Choose Modal first if:

• your team wants a custom Python API
• you care about warm-cost efficiency more than platform polish theater
• you want fine control over prewarm behavior without buying a whole dedicated inference vendor posture

Choose Baseten first if:

• the product really needs a managed dedicated inference platform
• rollout flow, replica policy, and production inference posture are part of the value
• paying more than Modal is acceptable because the control plane is the point

Choose Replicate first if:

• public or official models already fit the use case
• prediction-style APIs and webhooks are acceptable
• you want to avoid paying for warm dedicated capacity for as long as possible

Skip Replicate deployments as the default answer if the whole goal is one cheap always-ready endpoint. The docs do not support that story.

FAQ

Which one is cheapest for a warm endpoint?

On current published GPU prices, Modal. That does not make it universally best. It just means the warm-floor math is friendlier.

Which one is best for a serious managed inference service?

Usually Baseten. It is the most clearly dedicated-serving platform of the three.

Why does Replicate often feel cheaper than this comparison suggests?

Because public-model billing is a different product shape. Once you move into deployments and hold capacity online, Replicate starts billing like dedicated infrastructure.

Final verdict

If I had to compress the decision into one sentence, it would be this: Modal is the best default when you need one warm custom endpoint at the lowest published floor, Baseten is the premium choice when managed dedicated inference is the feature you are buying, and Replicate is strongest when you can avoid a warm dedicated endpoint altogether.

That is the honest July 17, 2026 answer to Baseten vs Modal vs Replicate for one warm inference endpoint.

Sources

• Baseten pricing - https://www.baseten.co/pricing/ - checked 2026-07-17
• Baseten resources - https://docs.baseten.co/deployment/resources - checked 2026-07-17
• Baseten autoscaling overview - https://docs.baseten.co/deployment/autoscaling/overview - checked 2026-07-17
• Baseten request lifecycle - https://docs.baseten.co/deployment/autoscaling/request-lifecycle - checked 2026-07-17
• Baseten billing and usage - https://docs.baseten.co/organization/billing - checked 2026-07-17
• Modal pricing - https://modal.com/pricing - checked 2026-07-17
• Modal scaling - https://modal.com/docs/guide/scale - checked 2026-07-17
• Modal cold starts - https://modal.com/docs/guide/cold-start - checked 2026-07-17
• Modal timeouts - https://modal.com/docs/guide/timeouts - checked 2026-07-17
• Replicate pricing - https://replicate.com/pricing - checked 2026-07-17
• Replicate billing - https://replicate.com/docs/topics/billing - checked 2026-07-17
• Replicate rate limits - https://replicate.com/docs/topics/predictions/rate-limits - checked 2026-07-17
• Replicate data retention - https://replicate.com/docs/topics/predictions/data-retention - checked 2026-07-17
• Replicate create a deployment - https://replicate.com/docs/topics/deployments/create-a-deployment - checked 2026-07-17
• Replicate monitor a deployment - https://replicate.com/docs/topics/deployments/monitor-a-deployment - checked 2026-07-17
• HostFleet Baseten note - /opt/hostbot/data/ai-hosting/notes/2026-07-16-baseten-pricing-limits.md
• HostFleet Modal note - /opt/hostbot/data/ai-hosting/notes/2026-06-11-modal-pricing-limits.md
• HostFleet Replicate note - /opt/hostbot/data/ai-hosting/notes/2026-06-22-replicate-pricing-limits.md
• HostFleet provider matrix - /opt/hostbot/data/ai-hosting/providers.csv