Baseten vs Replicate

Both run AI models on GPUs behind an API, but they lean different ways. Baseten is built to serve and train your own models in production, in the US, the EU or your own cloud, while Replicate starts from a catalog of thousands of ready models you call with one request.

The companies and products

Baseten

Baseten is an inference cloud: it runs AI models on GPUs for production apps, and sells hosted open models and training.

You package a model with Truss, its open-source format, or bring a Docker image, and Baseten runs it as replicas on Kubernetes that scale up with traffic and down to zero. It also sells Model APIs (open models priced per token) and training jobs on the same GPUs. Sandboxes, isolated environments where an agent runs commands and works with files, are in private preview; this page compares bochka with both the CPU instances and the sandbox preview.

Kind of product: Inference cloud: dedicated model deployments, Model APIs, training; Sandboxes in private preview
Isolation: Containers on Kubernetes, a dedicated namespace per customer, network isolation by Calico and Cilium; GPUs never shared across users
GPUs: Nvidia T4, L4, A10G, A100, H100, H100 MIG, H200, B200 and RTX PRO 6000, up to 8 per instance, plus multi-node
CPU instances: 1 vCPU / 2 GiB up to 16 vCPU / 64 GiB, on shared CPU instance types that cannot pick a region
Regions: United States and European Union for GPU deployments; self-hosted in your own AWS, GCP or OCI account on Enterprise
Billing: Per minute while a workload runs, including image builds and cold starts; no plan fee on Basic, starting credits for new workspaces
Compliance: Says it is SOC 2 Type II certified and HIPAA compliant

Replicate

Replicate runs AI models with an API: thousands of community and official models, plus your own models packaged with Cog.

You call a public model with one API request, fine-tune one on your data, or package your own model with Cog, Replicate's open-source tool that builds a Docker image with an HTTP prediction server. Replicate runs it on CPU or NVIDIA GPU hardware as deployments that scale up with traffic and down to zero. It is built for teams that serve models, not for agents that need a machine of their own, so this page compares the places where the two overlap.

Company: Founded 2019, Y Combinator Winter 2020, San Francisco; joined Cloudflare in November 2025 and continues as its own brand with the same API
Product: Model inference API: public models, fine-tunes, and custom Cog models run as autoscaling deployments
Isolation: Each model runs as a Docker container built by Cog; the docs do not name the isolation layer
Hardware: CPU (1 or 4 vCPU), T4, L40S, A100 80GB, H100 and H200, up to 8 GPUs per instance; most multi-GPU and all H200 need a committed spend contract
Run time: Each prediction times out after 30 minutes unless support raises it
Hosting: Infrastructure subprocessors (AWS, CoreWeave, GCP) listed in the United States; no region choice documented
Open source: Cog is open source and its images run anywhere Docker runs; the hosted platform is not

Which use cases each one fits

For each use case, Battleships prices every sandbox provider on the same workload and sorts them from the cheapest monthly bill up. “24th cheapest of 98” means 23 of the other 97 providers run that workload, with every feature it needs, for less. “With limits” means the provider runs it but misses something, so it is listed after every full match. Open a use case for its whole ranking and what each provider misses.

Use case and its workload	Baseten: price rank	Replicate: price rank	bochka: price rank
Software factory 100 at once · 4 vCPU / 8 GB · 2,200 sessions of 8 h a month A coding agent needs root, Docker and a disk that is still there tomorrow; neither Baseten nor Replicate documents Docker inside, and neither keeps a disk.	216th of 367, with limits $2,619 a month	268th of 367, with limits $4,340 a month	3rd cheapest of 367 $634 a month
Always-on AI employees 50 machines 24/7 · 2 vCPU / 4 GB An assistant that lives for months loses its files on both: Baseten preview sandboxes are deleted after 7 days by default and Replicate keeps no disk.	242nd of 366, with limits $3,258 a month	335th of 366, with limits $11,300 a month	5th cheapest of 366 $657 a month
Code interpreter bursts 100 at once · 1 vCPU / 2 GB · 200,000 sessions of 1 min a month Short bursts suit Replicate's per-second billing at $295 a month, while Baseten bills whole minutes plus cold starts and 15 idle minutes by default, for $1,856.	319th of 345, with limits $1,856 a month	258th of 345, with limits $295 a month	186th of 345, with limits $2,000 a month
Sandboxes for your users 300 at once · 2 vCPU / 4 GB · 30,000 sessions of 23 min a month Neither is built to hand your users a machine: Baseten's preview sandboxes cap at 10 at once by default, and Replicate gives an endpoint, not a VM.	277th of 361, with limits $1,868 a month	317th of 361, with limits $3,705 a month	Cheapest of 361 $209 a month
RL rollouts 1,000 at once · 2 vCPU / 4 GB · 1,000,000 sessions of 3 min a month Replicate's per-second billing makes RL rollouts cheaper there than on Baseten, at $17,900 against $31,140 a month, but each prediction stops at 30 minutes.	322nd of 360, with limits $31,140 a month	174th of 360, with limits $17,900 a month	34th of 360, with limits $2,000 a month
Agent evals 250 at once · 4 vCPU / 8 GB · 15,000 sessions of 20 min a month Replicate fits agent evals at $1,440 a month and Baseten does not, but bochka runs them on full VMs for $200.	297th of 364, with limits $1,609 a month	67th cheapest of 364 $1,440 a month	10th cheapest of 364 $200 a month
CI runners 40 at once · 8 vCPU / 16 GB · 6,000 sessions of 12 min a month	320th of 377, with limits $1,119 a month	311th of 377, with limits $864 a month	4th cheapest of 377 $86 a month
24/7 dev boxes 1 machine 24/7 · 4 vCPU / 8 GB	170th of 368, with limits $130 a month	306th of 368, with limits $226 a month	18th cheapest of 368 $26 a month
Computer-use agents 50 at once · 4 vCPU / 8 GB · 20,000 sessions of 15 min a month Neither Baseten nor Replicate has a desktop or computer-use product; bochka streams a 60fps desktop with Chromium from every sandbox.	294th of 369, with limits $1,799 a month	263rd of 369, with limits $1,296 a month	6th cheapest of 369 $200 a month
Browser-use agents 50 at once · 2 vCPU / 4 GB · 30,000 sessions of 10 min a month	318th of 367, with limits $1,194 a month	322nd of 367, with limits $1,296 a month	7th cheapest of 367 $200 a month
Scraping & bots (startup scale) 50 at once · 1 vCPU / 2 GB · 100,000 sessions of 3 min a month	220th of 346, with limits $183 a month	272nd of 346, with limits $439 a month	6th cheapest of 346 $90 a month
Mass scraping & bots (undetectable) 1,000 at once · 1 vCPU / 2 GB · 3,000,000 sessions of 2 min a month	126th of 338, with limits $3,654 a month	221st of 338, with limits $8,775 a month	71st of 338, with limits $1,800 a month
Parallel red teaming 200 at once · 2 vCPU / 4 GB · 5,000 sessions of 30 min a month	254th of 360, with limits $381 a month	306th of 360, with limits $855 a month	2nd cheapest of 360 $200 a month
GPU eval jobs 10 at once · 8 vCPU / 32 GB · 500 sessions of 30 min a month Both fit GPU eval jobs: Replicate comes to $1,338 a month and Baseten to $1,625, and bochka has no GPUs at all.	24th cheapest of 98 $1,625 a month	22nd cheapest of 98 $1,338 a month	Can't run it
macOS agents 5 at once · 4 vCPU / 8 GB · 1,000 sessions of 20 min a month	Can't run it	Can't run it	Can't run it

Key features

Feature	Baseten	Replicate	bochka
Kind of product	Inference cloud for serving and training your models, plus Model APIs; agent sandboxes in private preview	Model inference API: thousands of public models, fine-tunes, and your own Cog models	Sandboxes for AI agents: full VMs with coding agents, a desktop and hosting ready
Ready-made models	Model APIs: hosted open models priced per token	Thousands of community and official models for images, video, speech and text, plus fine-tuning	None; your agent calls the model API you choose
GPUs	Nvidia T4, L4, A10G, A100, H100, H200, B200 and RTX PRO 6000, up to 8 per instance, plus multi-node	T4, L40S, A100 80GB, H100 and H200, up to 8 per instance; most multi-GPU and all H200 need a committed spend contract	None, CPU only
Scale to zero and wake on request	Deployments scale to zero and start on the next request; standby sandboxes resume with files kept	Deployments scale to zero and start again on the next prediction, between min and max instances you set	No; you start a stopped sandbox through the API
Run time	No documented maximum for a deployment; it runs until it scales down or you remove it	Instances can stay on, but each prediction times out after 30 minutes unless support raises it	No session cap on paid plans; you set the time to live
Shell access	SSH once enabled for your organization	No SSH; you call the model through the HTTP API or client libraries	SSH, root and systemd, like a normal Ubuntu server
Root, Docker and /dev/kvm	Model containers run as the non-root app user; Docker inside and /dev/kvm are not documented	The model is the container; Docker inside and /dev/kvm are not documented	Docker, root, systemd and /dev/kvm on every machine
Files after a stop	Replicas start fresh from the image; preview sandboxes are deleted with their files after 7 days by default	No persistent disk; API inputs, outputs and logs are deleted after an hour by default	Disk snapshotted every minute and on stop; a stopped sandbox costs nothing and keeps it
Fork	No fork or clone; new replicas cold-start from the image	No snapshots or fork; you add replicas of the same image	Fork a running sandbox with its disk
Agent sandboxes	Private preview, access on request, 10 per organization at once by default	None	On every plan, self-serve, 100 to 2,000 at once by plan
Desktop and computer use	None; sandboxes run commands and file operations	No desktop or browser product	60fps desktop with Chromium, computer use and browser control
Keeping keys out of the machine	Secrets are passed into the model as arguments or files under /secrets	Secrets are sent to the model as inputs and redacted from logs afterwards	AI credentials proxy for model providers and any HTTPS API you register
Regions	US or EU for GPU deployments; shared CPU instances cannot pick a region	Infrastructure subprocessors listed in the United States; no region choice documented	EU only: Germany, Finland and France
Running it yourself	Self-hosted or hybrid in your own AWS, GCP or OCI account on Enterprise	Cog is open source and its images run anywhere Docker runs; the hosted platform is not	No

Every feature Battleships tracks (43)

Feature	Baseten	Replicate	bochka
Full VM (own kernel)	No	No	Yes
Docker inside	No	No	Yes
Nested virtualization	No	No	Yes
Root	Not published	Not published	Yes
Systemd	Not published	Not published	Yes
Desktop GUI	No	No	Yes
Browser	No	No	Yes
Computer-use API	No	No	Yes
SSH	Yes	No	Yes
Persistent disk	No	No	Yes
Pause/resume	No	No	Yes
Automatic snapshots	No	No	Yes
Fork/clone	No	No	Yes
Live fork (no pause)	No	No	Yes
Memory fork	No	No	No
Custom image (Docker or snapshot)	Yes	Yes	Yes
≥24 h sessions	Yes	Yes	Yes
Idle auto-stop	Yes	Yes	No
Wake on request	Yes	Yes	No
HTTPS preview URLs	Yes	Yes	Yes
Raw TCP inbound	Not published	Not published	Yes
Public IPv4	No	No	No
Egress allowlist	Yes	No	No
Secret proxy	No	No	Yes
Secret proxy for any API	No	No	Yes
Preinstalled agents	No	No	Yes
Agent harness API	No	No	Yes
MCP server	Yes	Yes	No
GPU	Yes	Yes	No
Windows	No	No	No
Arm64	No	No	No
Self-hosting	Yes	Yes	No
BYOC	Yes	No	No
Open source	No	Yes	No
Scoped API keys	Not published	Not published	Yes
Spend limits	No	Yes	Yes
Usage per sandbox (API)	Yes	No	Yes
Audit logs	Not published	Yes	No
SSO	Yes	No	No
SOC 2	Yes	Not published	No
HIPAA	Yes	Not published	No
EU data residency	Yes	No	Yes
Zero data retention	Not published	Not published	Yes

Performance

BasetenNot measured

ReplicateNot measured

bochka26.69 runs/s

There is no comparable CPU measurement for either: neither Baseten nor Replicate was part of the public StarSling benchmark run, so we cannot say which is faster at CPU work. bochka did 26.69 Node.js build-and-test runs per second at 4 vCPU / 8 GB in that run, on dedicated AMD Ryzen 9 9950X hosts. Both are built for GPU inference, which this test does not cover.

Node.js build-and-test loop at 4 vCPU / 8 GB, higher is better. Measured with the open hpc-sandbox-benchmarks suite. Baseten and Replicate have no published run in this suite yet, so there is no like-for-like number to compare.

Pricing in five scenarios

Each bill is priced from the providers' public price lists by the open Battleships engine, with the cheapest plan that runs the scenario.

Scenario	Baseten, a month	Replicate, a month	bochka, a month
One coding agent, workdays 4 vCPU / 8 GB, 50 GB disk, 8 hours a day, 22 days a month, files kept between days	$34 · Basic · with limits	$57 · Usage · with limits	$20 · $20/mo
One agent that never stops 4 vCPU / 8 GB, 50 GB disk, running 24/7 all month	$136 · Basic · with limits	$237 · Usage · with limits	$26 · $20/mo
50 agents in parallel 50 machines of 4 vCPU / 8 GB at once, 8-hour sessions, 1,100 sessions a month, Docker inside	$1,701 · Basic · with limits	$2,851 · Usage · with limits	$317 · $20/mo × 3 seats
Sandboxes for your users 2 vCPU / 4 GB, 30,000 sessions of 10 minutes a month, up to 300 at once	$1,246 · Basic · with limits	$1,620 · Usage	$200 · $100/mo × 2 seats
1,000 short jobs at once 2 vCPU / 4 GB, 1,000 at once, 3 minutes each, 200,000 jobs a month	$6,228 · Basic · with limits	$3,580 · Usage	$2,000 · $2000/mo · with limits

Baseten: Baseten bills per minute, with partial minutes rounded up, from image build and cold start until the replica stops; there is no plan fee on Basic. By default a replica stays billed for 15 idle minutes before it scales down.

Replicate: Replicate bills per second by hardware from prepaid credit that expires after a year. Public models bill only while they run; private models and deployments also bill setup and idle time.

Replicate: Plans start at $20 a month, and the whole fee is usage credit spent per second on running sandboxes. A stopped sandbox costs nothing and keeps its disk.

Limits in those scenarios

Neither Baseten nor Replicate has a persistent disk, so both miss the workday agent, the always-on agent and the 50-agent fleet: files are gone when an instance stops, and the fleet also needs Docker inside, which neither documents. Replicate fits the user sandboxes and the 1,000-job burst, while Baseten's CPU instances are an app platform rather than a sandbox, and its preview sandboxes run 10 at once by default. Each Replicate prediction stops after 30 minutes unless support raises it. On bochka the 1,000-job burst is queued, because it needs more starts per day than the $2,000 plan allows.

Scenario	Baseten	Replicate	bochka
One coding agent, workdays	adjacent product, not a like-for-like sandbox missing persistent disk different kind of product: app platform no persistent disk: files lost when it stops or sleeps	missing persistent disk no persistent disk: files lost when it stops or sleeps disk beyond 0 GiB: price unknown egress price not published	None that bite
One agent that never stops	adjacent product, not a like-for-like sandbox missing persistent disk different kind of product: app platform no persistent disk: files lost when it stops or sleeps	missing persistent disk no persistent disk: files lost when it stops or sleeps disk beyond 0 GiB: price unknown egress price not published	None that bite
50 agents in parallel	adjacent product, not a like-for-like sandbox missing persistent disk, Docker inside different kind of product: app platform no persistent disk: files lost when it stops or sleeps	missing persistent disk, Docker inside no persistent disk: files lost when it stops or sleeps disk beyond 0 GiB: price unknown egress price not published	None that bite
Sandboxes for your users	adjacent product, not a like-for-like sandbox different kind of product: app platform disk beyond 0 GiB: price unknown egress price not published	disk beyond 0 GiB: price unknown egress price not published	None that bite
1,000 short jobs at once	adjacent product, not a like-for-like sandbox different kind of product: app platform disk beyond 0 GiB: price unknown egress price not published	disk beyond 0 GiB: price unknown egress price not published	max 2,400 starts/day (needs ~6,667) max 600 starts/hour (needs ~1,000) bursts above 90 starts/min are queued (a burst of 333 takes ~4 min to start)

Which one should you use?

Use Baseten when

you serve or train your own models in production, need B200s or multi-node GPUs, or must run in the US, the EU or your own cloud account.

Bigger and newer GPUs. For large models, Baseten offers B200 and RTX PRO 6000 next to H100 and H200, up to 8 per instance and across nodes.
Pick the US or EU, or your own cloud. Teams with data rules can run GPU deployments in the US or EU, or self-host in their own AWS, GCP or OCI account.
Built for your own models in production. Package a model with Truss or a Docker image and Baseten scales replicas with traffic, with training on the same GPUs.

Use Replicate when

you want to call ready-made models from a large catalog, fine-tune one, or ship a Cog model that scales to zero and bills by the second.

Thousands of models, one call away. If you want an image, video, speech or text model without hosting it, Replicate's catalog starts with one API request.
Per-second billing for public models. Public models bill only for the seconds they run, which suits light or uneven traffic.
Cog runs anywhere. Cog, Replicate's open-source packaging tool, builds standard Docker images you can also run on your own servers.

Where bochka fits

Baseten and Replicate are for serving models on GPUs, and for that they are the right choice. If what you are really running is agents that write code, build containers or use a browser, they need a machine of their own, which neither gives you. Keep the models on Baseten or Replicate and let the agents work on bochka.

A full VM with root and Docker

Every bochka sandbox has its own kernel with root, systemd, Docker and /dev/kvm. Baseten containers run as a non-root user and Replicate's model is the container, and neither documents Docker inside, so an agent's docker build has nowhere to run.

A disk that stays, and forks

bochka snapshots the disk every minute and on stop, and you can fork a running sandbox. Baseten replicas start fresh and preview sandboxes are deleted after 7 days; Replicate keeps no disk, so each run sets up again and pays for that time.

Long sessions with agents and a desktop

bochka has no session cap on paid plans, ships Claude Code and Codex inside, and streams a 60fps desktop. A Replicate prediction stops at 30 minutes, and neither has a desktop, so you would build the agent loop and browser yourself.

Use bochka when your agents need a machine of their own: a full VM with root, Docker and /dev/kvm, a disk snapshotted every minute, forks, coding agents and a 60fps desktop, alongside whichever model API you call. Compare it one to one: bochka vs Baseten and bochka vs Replicate.

Questions people ask

Is Baseten cheaper than Replicate?

On the CPU scenarios we priced, Baseten usually comes out lower, such as $136 against $237 a month for one agent that never stops, though neither fits that job. Replicate is lower for 1,000 short jobs at once and for GPU eval jobs, at $1,338 against $1,625 a month.

Which is faster, Baseten or Replicate?

We have no comparable measurement for either: neither ran the public StarSling CPU benchmark. Both are built for GPU inference, so speed depends on the GPU you pick. bochka did 26.69 runs per second on that CPU test.

Which has more GPU choices?

Baseten lists T4, L4, A10G, A100, H100, H200, B200 and RTX PRO 6000, up to 8 per instance and across nodes. Replicate offers T4, L40S, A100 80GB, H100 and H200, and most multi-GPU and all H200 need a committed spend contract. bochka has no GPUs.

Can Baseten or Replicate run Docker inside?

Neither documents it. Baseten runs your model as a container under a non-root user, and Replicate's Cog builds your model into a Docker image without Docker inside it. Every bochka sandbox has root, Docker and /dev/kvm.

Where does my data run?

Baseten lets GPU deployments pick the US or the EU, and Enterprise customers can self-host in their own cloud. Replicate lists its infrastructure in the United States with no region choice documented. bochka runs only in the EU, in Germany, Finland and France.

Do Baseten or Replicate have sandboxes for AI agents?

Baseten has agent sandboxes in private preview, 10 per organization at once by default, deleted after 7 days by default. Replicate has none; it runs models behind an API. bochka gives every plan full VM sandboxes with coding agents and a desktop.

Try bochka

Install the CLI, then run bochka new.

Product preview