RunServerless Early access

Run a container on a GPU.
Skip the cluster.

Serverless runs your container as a job that finishes or an endpoint that serves requests. The GPU is provisioned for it and metered per second, and compute billing stops when the job ends or the endpoint stops.

Request access Read the API preview

Being built for the first cohort. Join by request.

Preview API · subject to change

import fantasti

fx = fantasti.Client()        # reads FANTASTI_API_KEY

job = fx.jobs.create(
    image="ghcr.io/acme/train:latest",
    command=["python", "train.py", "--epochs", "3"],
    gpu="H200:8",
    timeout="24h",
    capacity=fantasti.Spot(max_price="follow"),
    mlflow="support-sft",     # experiment name
)
for line in job.logs(follow=True):
    print(line)

Output · illustrative

job_4n7w  placing   H200:8 · spot · max_price follow
job_4n7w  running   mlflow experiment support-sft
epoch 1/3 · checkpoint saved
epoch 2/3 · checkpoint saved
epoch 3/3 · checkpoint saved
job_4n7w  exited 0  container disk removed · metering stopped

$ fantasti job submit --image ghcr.io/acme/train:latest \
    --gpu H200:8 --timeout 24h --spot --max-price follow \
    --mlflow support-sft -- python train.py --epochs 3
$ fantasti job logs job_4n7w --follow

Output · illustrative

job_4n7w  placing   H200:8 · spot · max_price follow
job_4n7w  running   mlflow experiment support-sft
epoch 1/3 · checkpoint saved
epoch 2/3 · checkpoint saved
epoch 3/3 · checkpoint saved
job_4n7w  exited 0  container disk removed · metering stopped

{
  "mcpServers": {
    "fantasti": { "type": "http", "url": "https://mcp.fantasti.ai/mcp" }
  }
}

Job tools · early access

jobs_create   confirms a launch above the key's per-call budget
jobs_logs

curl -X POST https://api.fantasti.ai/v1/jobs \
  -H "Authorization: Bearer $FANTASTI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
        "image": "ghcr.io/acme/train:latest",
        "command": ["python", "train.py", "--epochs", "3"],
        "gpu": "H200:8",
        "timeout": "24h",
        "capacity": { "mode": "spot", "max_price": "follow" },
        "mlflow": "support-sft"
      }'
# { "id": "job_4n7w…", "status": "placing" }

Shapes: job · endpoint
Capacity: on-demand · spot with max_price
Metering: per second, while it runs
Tracking: managed MLflow

§01 Two shapes Job · Endpoint

Two shapes.

For engineers who want a GPU for a task, not a machine.

Job Runs to completion Training, fine-tuning, preprocessing, evaluation. Restarts on failure. fx.jobs.create(image=…, gpu="H200:8")
Endpoint Serves requests at a URL A model or app behind HTTPS, with a token if you want one. Stop it and compute billing stops. fx.endpoints.create(image=…, port=8000)
Dev Interactive work Use a Workspace. Same image, same bill. /workspaces

§02 Jobs Illustrative

Jobs: start, run, finish.

Give Fantasti an image, a command and a GPU. The job runs until it exits or reaches its timeout, restarts if it fails, and its container disk is removed when it ends. Jobs can run on spot with a max price.

Timeoutsset a limit per job, up to 7 days
Spotfollow the market or set a max price
Trackedlog to managed MLflow

Your agent can do this. jobs_create and jobs_logs are MCP tools.

MCP server Early access

HGX H200 8-GPU baseboard, line drawing — One job, from request to exit. Illustrative.

7 days is the longest timeout a job can set.

Timeout · set per job

EQ 01 Source · Serverless API preview

timeout="24h"The job on this page

A job ends when its command exits or at its timeout, whichever comes sooner. Its container disk is removed, and metering stops with it.

§03 Endpoints Preview API

Endpoints: a model at a URL.

Serve a container at a managed HTTPS address, protected by a token. Stop it when you are not using it and compute billing stops with it. During early access each endpoint runs as one replica.

Preview API · subject to change

ep = fx.endpoints.create(
    image="ghcr.io/acme/serve:latest",
    args=["--model", "s3://acme-models/support-sft"],
    gpu="H100:1",
    port=8000,
    auth="token",
)
print(ep.url)    # https://ep-…  (managed HTTPS)
ep.stop()        # compute billing stops; start it again later

Output · illustrative

ep_9c3t  placing   H100:1 · on-demand
ep_9c3t  running   https://ep-… · token required · 1 replica
ep_9c3t  stopped   compute billing stopped

$ fantasti endpoint create support-sft \
    --image ghcr.io/acme/serve:latest \
    --gpu H100:1 --port 8000 --auth token \
    -- --model s3://acme-models/support-sft
$ fantasti endpoint stop support-sft    # start it again later

Output · illustrative

ep_9c3t  placing   H100:1 · on-demand
ep_9c3t  running   https://ep-… · token required · 1 replica
ep_9c3t  stopped   compute billing stopped

curl -X POST https://api.fantasti.ai/v1/endpoints \
  -H "Authorization: Bearer $FANTASTI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
        "image": "ghcr.io/acme/serve:latest",
        "args": ["--model", "s3://acme-models/support-sft"],
        "gpu": "H100:1",
        "port": 8000,
        "auth": "token"
      }'
# { "id": "ep_9c3t…", "status": "placing" }

Addressmanaged HTTPS, read from ep.url
Tokenauth="token" protects the address
Stoppeda stopped endpoint costs nothing

API reference

A request to an endpoint. Schematic · not to scale.

A client sends a request and the token over HTTPS to the endpoint's managed address. The address checks the token and passes the request to your container, which listens on port 8000 on one H100:1 replica and runs until you stop it. Connections: Your client to Managed address: request; Managed address to Your container: 1 replica.

Your client
Any HTTPclient
sends
request + token
Managed addressHTTPS
Checksthe token
url
https://ep-…
auth
"token"
Your containerH100:1
Runs untilyou stop it
port
8000
replicas
1

Your client
sends
request
+ token
Any HTTP client
Managed address
url
https://ep-…
auth
"token"
Checks the token
Your container (highlighted)
gpu
H100:1
port
8000
replicas
1
Runs untilyou stop it

Your client
Any HTTP client
sends
request + token
Managed addressHTTPS
Checksthe token
url
https://ep-…
auth
"token"
Your containerH100:1
Runs untilyou stop it
port
8000
replicas
1

Over time, the endpoint ep_9c3t is placed and then runs, metered per second. When you stop it, nothing is metered. When you start it again it runs and is metered per second, until you stop it.

§04 Spot and limits Illustrative

Spot for jobs, with your max price.

Every GPU type has a spot price that moves with supply and demand and can change as often as every 15 minutes. Pick one of two modes for each request.

Mode 01

Follow the market
fantasti.Spot(max_price="follow")
No ceiling. You pay the spot price for each interval, and Fantasti never stops your work because the price rose. Spot capacity can still be reclaimed.
Mode 02Drawn in the plot

Set a max price
fantasti.Spot(max_price=2.40)
Choose the most you will pay per GPU-hour. Your nodes run while the spot price is at or below it, and you pay the spot price, not your max. If the spot price rises above your max, those nodes stop.

How spot works

PLT 01Spot price and a max price over 72 hours Illustrative series · not Fantasti market data

Job: H100:1
Max price: $2.40 /GPU·hr
Ran: 51 h 15 m of 72 h
Stopped: 4×
Resumed: 4×

Illustrative series, not Fantasti market data. A step line shows a spot price for one GPU type over 72 hours in 15-minute steps, between $1.91 and $2.72 per GPU-hour. With the max price at $2.40, the nodes run 51 h 15 m of 72 hours, stop 4 times when the price rises above the max, and resume 4 times when it falls back. You pay the spot price for each interval, not the max.

Limits in early access

Job timeout: Up to 7 days; Set per job. A job ends at its exit or at its timeout.
Endpoint replicas: One per endpoint; One replica serves every request during early access.
Max price: Per request, per GPU-hour; It limits what you pay. It does not reserve capacity.

§05 Rates Prices reviewed 2026-10-10

Metered per second, at the rate of the GPU.

A job is metered from its start to its exit, an endpoint while it runs. Both appear as lines on your Fantasti statement, at the rate of the GPU they run on.

TAB 01Rates for the job and the endpoint on this page Source · Fantasti price list Reviewed 2026-10-10

Rates for the job and the endpoint on this page: what each runs on, its capacity mode, the price per GPU-hour and per GPU-second, and when it is metered
Shape	Runs on	Capacity	Per GPU-hour	Per GPU-second	Metered
Job job_4n7w	8 × H200 SXM	Spot `max_price="follow"`	From $0.87	From $0.000242	From its start to its exit
Endpoint ep_9c3t	1 × H100 SXM5	On-demand	$5.40	$0.001500	While it runs. Nothing while it is stopped.

Metered by the second, billed hourly. "From" is the lowest spot price for that GPU. The spot price moves with supply and demand. The spot price always stays below the on-demand rate for the same GPU. All prices

§06 Which product 4 of 7 products

Serverless, Batch Inference or Compute?

TAB 02Serverless, Batch Inference or Compute Source · Fantasti

Serverless, Batch Inference or Compute
Question	Serverless Early access	Batch Inference	GPU Instances	GPU Clusters
You bring	A container	A model and a dataset	An image, or your own stack	A scheduler, or nothing but SSH
Fantasti runs	One job or one endpoint	A sharded inference run	One machine with 1 or 8 GPUs	Many 8-GPU nodes on one fabric
Ends	When the job exits or you stop the endpoint	When every shard is done	When you stop it or its term ends	When you stop it or its term ends
Keeps	Nothing on the container disk	Outputs and checkpoints	Its volumes	Its volumes and shared filesystem
Capacity	On-demand or spot	Spot or on-demand	On-demand, spot, reserved	On-demand, spot workers, reserved
Use it when	The job is one container	The job is one model over one dataset	You need a GPU and root	The job spans nodes

§07 Q&A 4 questions

Sheet: 01 / 01
Title: Serverless: rates, fit and answers
Metering: Per second
Reviewed: 2026-10-10

§08 Request Early access

Bring a container.
We bring the GPU.

Serverless is in early access. Companies can request a place in the first cohort. Tell us what you want to run.

Request access Read the API preview

Run a container on a GPU.
Skip the cluster.

Two shapes.

Jobs: start, run, finish.

Endpoints: a model at a URL.

Spot for jobs, with your max price.

Follow the market

Set a max price

Limits in early access

Metered per second, at the rate of the GPU.

Serverless, Batch Inference or Compute?

Questions about Serverless.

Do I manage any VMs?

How is it billed?

Can a job run on spot?

Do endpoints autoscale?

Bring a container.
We bring the GPU.

Jobs: start, run, finish.

Endpoints: a model at a URL.

Follow the market

Set a max price

Limits in early access

Do I manage any VMs?

How is it billed?

Can a job run on spot?

Do endpoints autoscale?

Bring a container. We bring the GPU.

Bring a container.
We bring the GPU.