Does Replicate charge for failed runs?

Its own documents disagree, and we are not going to resolve it for them. The billing documentation says that for all models, a failed run is not charged. The terms say you will be billed for partial runs and failed attempts depending on the point of failure. Private models and deployments are separately documented as billing for active instance time regardless. SpicyAPI: Failed or timed-out tasks are usually not charged. Some models charge for content refusals as stated on their model page; check those billing terms before submitting.

Can I still set a monthly spend limit on Replicate?

Per its changelog, the account-level monthly spend limit was deprecated and customers who need it are directed to contact support. The remaining control is the prepaid balance, which blocks new work at zero. Ours are per key — day, calendar month and lifetime — all decided inside the transaction that admits the task, and a key can also be restricted to named models and source IP addresses.

Why does per-second billing matter?

Run-time billing depends on active inference; public-model setup and idle time are free. Private models and deployments can also bill setup and idle time. Check which model type you use before comparing it with a SpicyAPI estimate.

Platform comparison

SpicyAPI vs Replicate

Replicate is the widest open-model catalogue in this comparison and the one most people have already used. Two rows are worth reading carefully before you commit to it: what a second of billing buys, and what happened to its monthly spend limit.

Browse models All comparisons

Starting point

What each platform is built around

Replicate

Replicate hosts open-source and community models behind one prediction API and a packaging tool for pushing your own. Most models bill by the time the run takes at a per-second hardware rate; some bill per output. Its terms acknowledge in full capitals that the service can generate objectionable content and that managing that output is the customer's responsibility.

SpicyAPI

We publish a curated catalogue rather than an open one: fewer models, each with a fixed price against its own unit and a published input schema. Nothing bills by execution time, so a slow run costs the same as a fast one.

Row by row

The eight rows worth reading before you pick

Our column is the same on every one of these pages, because our behaviour does not change with who we are being compared against. Only the right-hand column is specific to this platform.

SpicyAPI vs Replicate
What you are comparing	SpicyAPI	Replicate
Content policy	No per-request content review, no safety-checker parameter to negotiate and no content-mode flag on the key. What a model will or will not produce is decided by that model. Unlawful use stays prohibited by the acceptable use policy and the reporting channel is public.	The terms state in capitals that the service can generate objectionable content and that the output is the customer's to manage. The acceptable use policy addresses content only as a general prohibition on obscene or offensive material. A safety checker runs on certain base models and their fine-tunes for web predictions; a disable_safety_checker parameter defaults to false and is documented as API-only.
How you are billed	US dollars, per the unit the task itself is measured in. No credit currency to convert, no subscription, no plan tier, no monthly minimum. A quote call prices the exact body you are about to send and reserves nothing. Balance is topped up by card or crypto (USDT).	Most models bill by run time at a per-second hardware rate, so the same output costs more on slower hardware or a slower run. Some models bill per output image or per thousand output tokens. Private models and deployments also bill for setup and idle time. Credit is prepaid, valid one year and non-refundable.
A failed request	A successful task is settled at what the run used and is not refunded because the output disappointed. Failed or timed-out tasks are usually not charged. Some models charge for content refusals as stated on their model page; check those billing terms before submitting.	The two official documents disagree. The billing documentation states that for all models, a failed run is not charged. The terms state that you will be billed for partial runs and failed attempts depending on the point of failure. Private models and deployments are documented as billing failed and cancelled runs for the time the instance was active.
Request shape	One create call answers 202 with a task ID; results arrive by polling or by signed webhook. Text speaks an OpenAI-compatible surface on the same key. Input is validated against a published schema with additionalProperties closed, and every write accepts an Idempotency-Key for 24 hours.	Asynchronous predictions with polling, plus webhooks that can be filtered by event. A request can also be left open with a wait parameter for up to sixty seconds. Predictions time out after thirty minutes. No OpenAI-compatible chat endpoint is listed in the HTTP API reference.
Spend controls	Set daily, monthly and lifetime spending caps on each key. New keys have no spending cap until you set one. You can also restrict a key to selected models and source IP addresses; unsettled holds count toward its allowance.	The ability to set a monthly spend limit was deprecated, with the changelog directing customers who need it to contact support. What remains is the prepaid balance, which stops new work once it reaches zero, and a reload threshold. No per-key limit is documented.
Official tooling	Native REST plus an OpenAI-compatible surface, our own SDK, CLI, MCP server and Agent Skill, all on the same key.	First-party clients for Node.js, Python, Swift and Go, a hosted MCP server, and a CLI plus a model packaging tool on GitHub. The documentation's client library page lists the MCP server but not the CLI.
Published commitments	A published operational target of 99.9% for our own layer — accepting requests, admitting tasks, delivering callbacks. It is a target, not a service level agreement: no service credits, no compensation, and it excludes the third-party model providers that execute generation, announced maintenance windows and events outside our control. Live status is published at status.spicyapi.ai.	No uptime commitment or service level agreement appears in the public documentation. Documented rate limits are 600 prediction creations per minute and 3000 requests per minute on other endpoints.
Languages	Site, model documentation, console and API error text in ten languages: English, Japanese, Korean, German, French, Spanish, Brazilian Portuguese, Russian, Italian and Polish. Error codes never change with the language, so your branching stays stable.	Documentation observed in English only, with no alternate-language versions advertised.

Swipe horizontally to see all columns.

Content policy

What the platform will refuse, and who decides

Match the exact model version and fine-tune, not just the display name. Check your character reference and prompt on the destination endpoint before moving the whole queue.

How you are billed

What one unit of billing actually is

Separate public-model active time from deployment setup and idle costs. Public-model setup and idle time are free; a cold start can increase waiting without increasing billable inference time.

Request shape

What the HTTP response means

Map Replicate's version/input and prediction ID to the destination schema and task ID. If Prefer: wait returns unfinished work, continue polling instead of submitting another billable job.

Spend controls

What stops a loop that has gone wrong

Give an integration moving from an auto-reloading account a separate capped key. Keep customer-level budgets in your app: a key cap covers everyone sharing that key.

Fit

Which one fits the work you are doing

Pick Replicate when

You want the widest catalogue of open models, including community-published ones, and the ability to package and push your own with their container tooling.
Your work is research or experimentation where breadth matters more than a predictable per-task price.
You already have a Go or Swift integration against their first-party client and no reason to move it.

Pick SpicyAPI when

You need a spend cap you can set yourself. Theirs was deprecated; ours are per key — day, calendar month and lifetime — and enforced at admission.
You want a fixed price per task rather than a per-second rate, so a slow run and a fast run cost the same.
You want one unambiguous answer about failed runs. Their billing page and their terms currently give two.

Models

Models you can call for this

Each link opens that model family's own page: inputs, limits, content policy and the current price per unit.

ByteDanceSeedream 5.0 ProImage editingContent policy: Uncensored AlibabaQwen Image 2.1Text to imageContent policy: Uncensored AlibabaWan 3.0Image to videoContent policy: Uncensored AlibabaQwen Image Edit SpicyImage editingContent policy: Uncensored ByteDanceSeedream 5.0 FlashText to imageContent policy: Uncensored ByteDanceSeedance 2.5 SpicyImage to videoContent policy: Uncensored

FAQ

Questions about Replicate and SpicyAPI

Its terms state in capitals that the service can generate objectionable content and that all output is the customer's to manage, delete or quarantine. Its acceptable use policy addresses content only through a general prohibition on obscene, defamatory or otherwise offensive material, and we found no clause specific to uncensored content. Separately, a safety checker runs on certain base models and their fine-tunes for web predictions, with a disable parameter documented as available through the API. Read both documents yourself.

What we read on Replicate

These are Replicate's official sources. This comparison describes published features, not a benchmark or a test of their service.

Other comparisons

SpicyAPI vs fal.ai SpicyAPI vs Segmind Pricing