Runware
Runware is a developer API that runs image, video, audio, 3D and language models through a single endpoint. It hosts more than 400,000 models on hardware it builds itself and bills strictly per request.
What is Runware?
Runware is a generative AI inference platform sold as an API, operated by the British company Runware Ltd. Its premise is that a product team should not have to integrate a different vendor for every modality: one authentication, one POST /v1 endpoint and one bill cover image, video, audio, 3D, vision and large language models. Every request body is an array of tasks, each carrying a task type and a model identifier, so moving from Flux to a Google or Anthropic model is a change of string rather than a rewrite. The homepage counts 192 image models, 108 video, 38 vision, 37 text, 28 audio, 77 editing operations and seven 3D models, on top of a community library it puts above 300,000 entries, for a total it advertises as 400,000+ models.
What makes the economics work, according to the company, is that it does not rent generic cloud GPUs. The Sonic Inference Engine is a hardware and software stack Runware designs and runs itself, tuned from BIOS and kernel upward, with models kept preloaded across regions and large ones sharded across local GPUs. Its Sonic Inference Pods are modular data centres it says can be deployed in weeks wherever power is cheap. Open-source models are therefore billed on optimised compute time, so a faster model automatically costs less, while closed-source and partner models are billed at a negotiated fixed price per request.
Around that core sit a Playground for testing, model collections curated by task, LoRA style training, a Model Upload path for private weights, and a beta serverless tier split between raw GPU compute billed by the second and an API Gateway that puts a customer's own model behind a managed endpoint. Tooling covers TypeScript and Python SDKs, a CLI, an MCP server, ComfyUI nodes and OpenAI-compatible calls, with documentation written to be read by language models. The company began in 2023 as PicFinder, a real-time image generator, and took the Runware name in 2024.
What it does
- Generate images from a text prompt or from a source image
- Generate and edit video, including image-to-video and video-to-video
- Call large language models with streaming and tool calling
- Produce speech, music and sound effects, and transcribe audio
- Edit images: inpainting, outpainting, object erase, virtual try-on, background removal, upscaling
- Train style LoRAs and upload your own checkpoints, safetensors and LyCORIS files
- Rent raw GPU compute by the second or publish your own model behind a dedicated API
When to use Runware / When not to
A quick filter to help you decide if Runware is the right fit.
When to use Runware
- Product and engineering teams adding image, video or audio generation to an existing application without provisioning GPUs
- High-volume platforms whose margin depends on the cost of each generation, such as Quora, OpenArt, NightCafe, Freepik or Higgsfield AI
- Developers who want to swap between OpenAI, Google, Anthropic, Flux or Kling models by changing one string rather than rewriting an integration
- Teams training their own style LoRAs or bringing their own checkpoints, safetensors and LyCORIS weights and keeping them private
- Enterprises that need per-team access controls, spend caps, usage monitoring, SSO with SAML and a negotiated SLA
When not to use Runware
- Non-technical users looking for a ready-made generation interface: everything here starts with an API key and code
- Anyone wanting to evaluate output quality before committing, since the Playground and Studio both sit behind a login
- People expecting a mobile or desktop app, or a browser extension, none of which exist
- Teams that need a permanent free tier: the only free element is a one-off 2 USD credit granted at sign-up
- European buyers whose procurement requires a published data processing agreement and a named list of infrastructure subprocessors
How to use Runware
A typical end-to-end flow, from setup to results.
- Create an account on the sign-up page; a 2 USD credit is granted with a business email address and no card is required
- Generate an API key from the dashboard
- Send a POST request to the single endpoint with an Authorization Bearer header carrying that key
- Build the body as an array of tasks, each with a task type such as imageInference, videoInference, audioInference, 3dInference or textInference, a task identifier and a model identifier
- Change modality or provider by editing the model field, since the request shape stays the same
- Add the includeCost parameter to read back the exact amount deducted for each request
- Choose the transport: REST for stateless work, WebSockets for persistent low-latency sessions, server-sent events for token streaming
- Collect asynchronous results through a per-task webhook URL or by polling
- Prototype in the Playground before wiring anything, or let an agent do the integration through the MCP server, the CLI or the ComfyUI nodes
- Top up the prepaid balance, set auto-reload and low-balance alerts, and monitor usage from the dashboard
Pros & Cons
Pros
- One endpoint, one key and one invoice replace a separate integration per model provider
- Unit prices are published and openly benchmarked against fal, Replicate and WaveSpeed, with the cheapest image models quoted at 0.0006 USD per image
- Open-source models are billed on optimised compute time, so the price falls automatically as Runware speeds the model up
- Failed and rate-limited requests are not charged, and there is no billing for cold starts or idle GPU time
- No subscription and no commitment: the prepaid balance does not expire
- A contractual commitment not to train on customer inputs, outputs, Training Data or Trained Models, with ownership of outputs left to the customer
- Broad developer tooling and documentation designed to be consumed by coding agents through llms.txt and per-model JSON schemas
Cons
- The model-by-model price grid is rendered in JavaScript, so the pricing page shows no figures without it and the exact cost of a request means opening the Playground
- Prices vary across thousands of parameters such as resolution, duration, steps and batch size, which makes budgeting a feature awkward
- No permanent free plan, only a one-off 2 USD credit at sign-up, and purchases are non-refundable except as discretionary credits
- No data processing agreement is published and no subprocessor list covers the inference infrastructure
- SOC 2 and ISO 27001 are asserted only through a footer badge and an FAQ answer, with no trust page, certificate or compliance portal behind them
- Neither the terms nor the privacy notice carries a visible effective date, so the legal commitments cannot be dated
- The Playground and Studio require an account, so nothing can be evaluated before signing up, and the serverless tier is still in beta
Pricing & Plans
There is no permanent free plan. Runware operates on a strict pay-as-you-go basis, with no subscription, no commitment and no minimum spend; the account runs on a prepaid balance that does not expire. New users receive a one-off credit of 2 USD when they sign up with a business email address, with no payment card required. Beyond that, the lowest published entry point is 0.0006 USD per image, applying to the cheapest image models such as FLUX.2 [klein] 4B, FLUX.1 [schnell] and Real-ESRGAN; the vendor states that image generation generally runs between 0.0006 and 0.24 USD per image depending on the model, resolution and quality settings. Raw GPU compute is billed separately by the second, from 1.99 USD per GPU-hour at launch. All amounts are denominated in US dollars.
- managed inference across 400
- 000+ models
- billed per request for fixed-price models or on optimised compute time for open-source models
- raw GPU and CPU capacity billed by the second
- from 0.0000044 USD per second for a vCPU node and 1.99 USD per GPU-hour at launch
- with reserved capacity discounted up to 50 percent
- a customer model published behind a dedicated public or private endpoint
- billed per token
- image
- video
- audio or asset
- dedicated capacity
- custom SLAs
- priority routing
- volume-based pricing
- 24/7 on-call engineering
- a shared support channel and granular access controls
- quoted through the sales team
- 2 USD granted once at sign-up with a business email address
Data, GDPR & hosting
A consolidated view of how Runware handles your data.
GDPR overview
The GDPR position is documented but incomplete. The controller is named as Runware Ltd at its Harrow address, and the privacy notice cites both the GDPR and the UK GDPR, listing the legal bases relied on: consent, contract, legal obligation, vital interests and legitimate interests. Rights of access, copy, rectification, erasure, restriction, portability, objection and non-submission to automated decision-making are opened to people in the EEA, the UK, Switzerland and Canada, and are exercised through the contact form or by email. A GDPR Compliant badge appears in the footer of every page. What is missing is just as clear: no representative is designated under Article 27, no data protection officer is named, no data processing agreement is published, and neither the terms nor the privacy notice display an effective date.
Who owns the data?
The terms are explicit that the customer keeps ownership. As between the two parties, you own your Outputs, subject to third-party rights in your inputs and to the licence of the model used, and you own the Trained Models you build from your own Training Data, over which Runware claims no ownership. In return you grant Runware a worldwide, non-exclusive, royalty-free, perpetual and transferable licence over your Contributions and Generations, used to operate the service; that licence expressly does not extend to Training Data or Trained Models. Signing up also lets Runware display your company name and logo in its marketing, revocable in writing within five working days.
Reuse rights
Outputs can be reused without asking Runware for permission. Official and partner models carry commercial usage rights through Runware's own agreements, so no separate licence fee is owed to the model provider, and generations can go straight into production. Community models are different: commercial use is governed by the licence the model creator published, and Runware links to that source so the terms can be checked before shipping. Two limits remain regardless of the model: third-party rights in the inputs or the outputs, such as copyright, trademark or a person's image, and the prohibited-use rules in the terms. Runware states it never uses inputs, outputs, Training Data or Trained Models to train its own models or anyone else's, and that generated content is purged from its servers unless storage is explicitly enabled.
Data retention & training
Hosting summary
The terms state plainly that the services are hosted in the United States, Germany and Romania, and that using them from elsewhere means consenting to a transfer into those three jurisdictions. The footer of every page advertises data residency in the EU and the US, so the two statements overlap without matching exactly. Requests are routed automatically between regions to a pod that already has the model loaded, and the company presents its Sonic Inference Pods as modular data centres it can place near users and align with local regulatory frameworks; it also states that its data centres run on renewable energy, mainly hydro and solar. The website itself sits behind Cloudflare on an anycast address. The governing entity is Runware Ltd in the United Kingdom, and disputes fall under English law. No list of infrastructure subprocessors is published, so the identity of the parties operating those facilities cannot be checked from the site.
Things to keep in mind
Risks and trade-offs to weigh before adopting Runware.
- SOC 2 and ISO 27001 are claimed only in a footer badge and an FAQ answer; there is no trust page, certificate or third-party compliance portal to back them, so do not treat them as verified
- No data processing agreement is published and no subprocessor list covers the inference infrastructure, which will block a procurement review that requires both
- No Article 27 representative and no data protection officer are designated, even though the operator is a post-Brexit British company processing data from European users
- Signing up grants Runware a perpetual, transferable licence over your Contributions and Generations, and the right to use your company name and logo in its marketing until you object in writing
- Purchases are non-refundable and any refund is at the vendor's discretion and in service credits only, so an over-provisioned balance is money you cannot get back
- Because generation is cheap and instant, it is easy to lose track of spend and of what has been produced; the spend caps, low-balance alerts and per-team controls exist for a reason and are worth setting up on day one
- Runware provides no consumer-facing age verification and pushes that duty onto you, so any product you build on top inherits the responsibility for who sees what it generates
Setup & Integrations
Technical difficulty
Moderate, and strictly for developers. There is no no-code path to the API itself: you need an API key and the ability to send an authenticated POST request whose body is an array of tasks. A first generation is genuinely quick, and the site promises going live in hours. Scaling is where the work is, since webhooks, polling, streaming, rate limits and error handling all have to be understood. TypeScript and Python SDKs, a CLI, ComfyUI nodes, an MCP server and OpenAI-compatible calls reduce the effort, and connectors to n8n, Make and Zapier are advertised.
Deployment
Integrations
Behind Runware
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Runware.
Frequently asked questions
What can I generate through the Runware API?
How many models are available?
How does billing work?
Is there a free plan or a free trial?
What is the cheapest price for an image?
Is my data used to train models?
Who owns the images and models I generate?
Where is the data hosted?
Can I try it before writing any code?
Who is behind Runware?
Should you pick Runware?
Runware is a serious piece of infrastructure aimed squarely at people who write code. Its argument is narrow and well made: one endpoint, one key and one invoice for every modality, on hardware the company designs and operates itself, which is what allows it to publish unit prices well below the platforms it names on its own comparison pages. For a product team already spending real money on generation, that argument is easy to test, because the prices are public, failed requests are not billed and nothing has to be committed in advance. The contractual position is unusually clear too: outputs and trained models belong to the customer, and the terms rule out training on customer data rather than leaving it to a setting.
The reservations are mostly about what cannot be checked. The model-by-model price grid only renders with JavaScript, and the real cost of a feature depends on resolution, duration, steps and batch size, so budgeting takes work. SOC 2 and ISO 27001 appear as a footer badge and an FAQ line with no certificate, trust page or compliance portal behind them. No data processing agreement is published, no subprocessor list covers the inference infrastructure, and neither the terms nor the privacy notice shows an effective date. There is also no way to judge output quality before creating an account.
So: a strong fit for engineering and product teams shipping generative features at volume, and for enterprises willing to negotiate the missing paperwork directly. A poor fit for anyone who wanted a generation interface rather than an API, and a cautious one for a procurement team that needs its compliance evidence in writing before it signs.
- Choosing a selection results in a full page refresh.
- Opens in a new window.