Is Together AI a media generation platform?

Not primarily, by their own framing. They describe themselves as the AI native cloud and their documentation leads with running, training and serving open-source models. Image and video are one category alongside chat, code, vision, GPU clusters, fine-tuning and evaluation. If media is your main workload that focus is worth weighing; if open-weight text is, it is a reason to pick them.

Do they charge for failed requests?

Their documentation answers this for batch work — failed requests in the error file are not billed — and we could not find a statement covering synchronous or real-time requests. 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.

What about free credits?

Their documentation states there is no free trial and that access requires a minimum five dollar credit purchase. We also have no free tier. The difference is what the balance is: theirs is credits used as a measurement unit, ours is US dollars, and there is no conversion step between the price you read and the amount that settles.

Platform comparison

SpicyAPI vs Together AI

Together AI is an open-model inference cloud first and a media platform second, which is the honest starting point for this comparison. If your workload is open-weight text at volume, that focus is a reason to pick them.

Browse models All comparisons

Starting point

What each platform is built around

Together AI

Together AI runs open-source models across serverless inference, dedicated endpoints, GPU clusters and fine-tuning. Text is priced per million tokens, images per image, videos per video, GPUs per hour. The platform is fully prepaid with a five dollar minimum purchase and no free trial, and the OpenAI-compatible surface covers chat, embeddings, images, speech and transcription.

SpicyAPI

Our centre of gravity is the opposite: generative media first, with OpenAI-compatible text on the same key. We do not sell GPU capacity, clusters or fine-tuning, and if that is what you need, they are the better fit.

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 Together AI
What you are comparing	SpicyAPI	Together AI
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.	No separate acceptable use policy document; the restriction lives in the terms of service, which prohibit using the service to communicate material that is obscene, alongside a right to suspend access for breach. Image generation exposes a disable_safety_checker parameter, documented as running by default and returning HTTP 422 when triggered.
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).	Prepaid credits used as a measurement unit, with a five dollar minimum purchase and no free trial. Text per million tokens, images per image, video per video, transcription per minute, speech per million characters, GPUs per hour. Fine-tuning carries a per-model minimum charge. Fees are documented as 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.	Documented for batch work only: failed requests in the error file are not billed, and cancelling does not refund responses already generated. We found no statement covering synchronous or real-time requests.
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.	An OpenAI-compatible surface covering chat, completions, embeddings, image generation, speech and transcription. Images are synchronous. Video is asynchronous create-and-poll and is documented as Together-native, not exposed through the OpenAI SDK. We found no webhook or callback for media generation; a separate batch API handles bulk work.
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.	A maximum monthly spending limit exists at account level — the 402 error code names it — but we found no documentation for setting or adjusting it, and none for a per-key budget. Documented usage limits are dynamic rate limits rather than spend.
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 Python and TypeScript SDKs, a CLI, a documentation MCP server that exposes docs retrieval rather than inference, and a published set of agent skills.
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. Provisioned throughput is sold on a one-month minimum term, and enterprise contracts are billed under their own terms.
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; the documentation sitemap carries no language prefixes or alternates.

Swipe horizontally to see all columns.

Content policy

What the platform will refuse, and who decides

For unfiltered image workflows, read the model's safety setting alongside Together's terms. A documented disable_safety_checker parameter is not a promise that every requested scene is supported.

How you are billed

What one unit of billing actually is

Separate serverless token or image rates from dedicated-GPU hours and fine-tuning charges. Include your existing minimum commitments when deciding whether migrating a small workflow actually saves money.

Request shape

What the HTTP response means

Keep supported OpenAI chat methods, updating the key, base URL and model ID. Together video uses its native job API; replace its polling and outputs.video_url mapping separately.

Spend controls

What stops a loop that has gone wrong

Rate limits manage request throughput, not your application's dollar budget. Set a capped key for the migrated workflow and retain per-user quotas and retry limits in your app.

Fit

Which one fits the work you are doing

Pick Together AI when

Your workload is open-weight text at volume. That is what the platform is built around and the OpenAI-compatible surface is broader than ours.
You need fine-tuning, dedicated endpoints, GPU clusters or reserved throughput. We sell none of those.
You want one vendor covering inference and training rather than a media API plus something else.

Pick SpicyAPI when

Media is the main workload rather than a side category, and you want webhooks on media jobs rather than polling.
Your requests are being refused under a terms clause prohibiting obscene material, and you need the model's own behaviour to be the limit.
You want spend caps you can set yourself, per key and per day, rather than an account limit whose configuration is not documented.

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.

AlibabaWan 3.0Text to videoContent policy: Uncensored ByteDanceSeedream 5.0 ProText to imageContent policy: Uncensored ByteDanceSeedance 2.5 SpicyImage to videoContent policy: Uncensored ByteDanceSeedance 2.5Text to videoContent policy: Uncensored AlibabaQwen Image 2.1Text to imageContent policy: Uncensored AlibabaWan 3.0 SpicyImage to videoContent policy: Uncensored

FAQ

Questions about Together AI and SpicyAPI

Their terms of service prohibit using the service to communicate material that is obscene. There is no separate acceptable use policy document — we checked the usual paths and they return not found — so that clause in the contract is the operative rule. Image generation additionally runs a safety checker by default, with a documented parameter to disable it and HTTP 422 returned when the checker fires.

What we read on Together AI

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

Other comparisons

SpicyAPI vs OpenRouter SpicyAPI vs Replicate Pricing