Atlas Cloud

What is Atlas Cloud?

Atlas Cloud is an AI API aggregation platform that puts more than 400 models for chat, image, video and audio behind one API key, one endpoint and one billing account. It is operated by ATLAS CLOUD AI INC., a company registered in New York, and its legal terms are governed by New York law.

The product is infrastructure, not an application. There is no canvas or project workspace — you write code, pass a model identifier and receive a completion or a generated file. What it sells is the removal of per-vendor integration work: instead of separate keys, SDKs and invoices for a text, image and video provider, you hold one of each.

The catalogue reports its own totals and filters by modality. It currently lists 432 models: 217 video, 119 image, 69 LLM and 27 audio or 3D. Video is the largest part by count, which is unusual for a platform that also presents itself as an LLM gateway, and it shapes both the pricing model and the integration pattern below.

Core features

• OpenAI-compatible chat: the chat endpoint accepts the OpenAI SDK or any HTTP client. Point the client at the platform base URL, swap the model string, and streaming, tool use and function calling keep working unchanged — there is no separate SDK to learn for text.
• Two base paths, one key: chat completions run against the /v1 base URL while image, video and audio generation run against /api/v1. The same bearer key authenticates both, but the paths are not interchangeable.
• Synchronous text, asynchronous media: chat returns inline. Image and video run as predictions — post to a submit endpoint, receive a prediction id, then poll until the status reaches succeeded. Polling roughly every two seconds is the documented pattern, and webhooks are optional.
• Per-task endpoints: a model family is not one call with a mode flag. Text-to-video, image-to-video and reference-to-video are separate catalogue entries with separate prices, so switching input type means switching endpoint and re-checking the rate.
• Exact model identifiers: the id copied from a catalogue card is the literal string passed in the request. A model that is not available to your account returns 404 rather than a distinct availability error.
• Media upload endpoint: a dedicated call accepts a file and returns a temporary URL for a later generation task — the documented route for inputs too large to inline as Base64.
• MCP server: one JSON block naming the npx package and the key environment variable wires the catalogue into MCP-compatible clients, so an agent can call models from plain language without provider glue code.
• Per-request retention headers: two optional headers set how long generated media and the request record survive, independently of each other, on a call-by-call basis.

Where the abstraction stops

The unified key is a billing and authentication abstraction, not a semantic one. Each model keeps its own parameters, resolution tiers and price ladder, and the platform does not normalise those away. Two models that both accept an image and a prompt can differ by an order of magnitude in cost per second and expose entirely different aspect-ratio and duration options.

Guide

1. Create an account and generate a key in the console. The key is shown once at creation; if it is lost it cannot be recovered and has to be replaced with a new one, so it belongs in a secret store or environment variable from the first minute.
2. Decide which base path the work needs. Text goes to the /v1 OpenAI-compatible route; images, video and audio go to /api/v1. A typo in the path returns 401, not 404, because the gateway authenticates before it routes — so a working key failing on authentication usually means the URL is wrong.
3. Copy the exact model id from its catalogue card and read the billing block on the same card, which states whether the model bills per second, per image or per token and whether optional parameters add separate line items.
4. For media, submit asynchronously and poll. If a request larger than 50 MB is needed, upload the file first or pass a URL — a body over that ceiling is rejected with 413.
5. Set retention headers on the submit call if the workload has a data-handling requirement, since these cannot be applied retroactively once the job has run.
6. Implement client-side back-off before production, because the endpoints do not report how much quota is left.

Use cases and examples

The platform fits pipelines that span modalities. A product that writes ad copy with an LLM, renders a key visual, then animates it into a clip touches three vendors in a conventional stack and one here, with a single balance and logs traceable by request id. Draft-heavy pipelines are a second fit: because rates rise sharply with resolution, a workflow that drafts at a low tier and finishes selected outputs at a high tier fits inside one catalogue, not two accounts.

Agent-driven work is a third case. With the MCP server configured, a coding assistant can generate or edit an image inside a larger task without an integration being written first — useful exploratory ground, though the same billing and retention rules apply to every agent call.

Who is it for

• Development teams shipping multi-modal features benefit most, because the saving is proportional to the number of vendors replaced. A team using one modality gains little from aggregation.
• Teams already on the OpenAI SDK have the shortest migration path for text, since the chat surface is a drop-in change of base URL and model string.
• Cost-sensitive batch producers get a real lever in the per-resolution price ladders, which make draft-then-finish workflows cheaper than rendering everything at final quality.
• Agent and tooling builders can reach the catalogue through the MCP server without writing provider integrations.

It is a poor fit for non-technical creators, since there is no editor and every output arrives through code. It is also unsuitable for work under HIPAA, FISMA or similar regulation: the terms state the services are not tailored to those regimes and that people whose interactions fall under them may not use the services at all. Teams that must publish latency or throughput comparisons should read the usage restrictions first.

Platforms

Access is programmatic. There is no desktop or mobile application; the surfaces are the HTTP API, the web console used for keys and usage, and the MCP server for agent clients.

Because the chat route is OpenAI-compatible, the documentation lists third-party clients that work by configuration alone: desktop chat applications and self-hosted web front ends can be pointed at the platform host, with the /v1 suffix required in the base URL. Editors and coding agents connect through the MCP server rather than a native plugin.

Two constraints bound availability. The services are hosted in the United States, and use from another region is treated as consent to transferring and processing personal data there. A request can also be refused on regional grounds with HTTP 451, which the documentation classifies as not retryable — region is an access condition to verify during evaluation, not an error to code around.

Pricing

Billing is prepaid and usage-based rather than subscription-based, and the shape of the charge follows the modality.

Resolution, not model choice alone, is the dominant cost variable for video. One hosted family is listed at 0.099 USD per second at 720P, 0.144 at 1080P and 0.279 at 4K on the same card. A ten-second clip can differ several times in cost on tier alone.

Two further mechanics change the arithmetic. The catalogue shows the platform rate beside the provider's official rate and labels the gap as a percentage discount — but that discount differs by model, so it is a per-model property, not a platform-wide rate. And some models use composite pricing, where an optional request parameter adds its own charge: one listed model adds 0.117 USD per second when prompt expansion is enabled on top of a 0.4 USD per second base. The headline number on a card can be the base component only.

Credits must be bought in advance and expire 365 days from purchase. All purchases are final: credits are non-refundable, non-transferable and cannot be exchanged for currency. Any remaining balance is forfeited immediately on termination or account deletion, and an export-control violation triggers termination plus irrevocable forfeiture. Published rates may change without notice, effective as posted, so a cost model built on today's catalogue should be re-checked rather than assumed stable.

Alternatives

A 2026 comparison article on a site whose footer reads "Powered by AtlasCloud" sets the service against fal.ai and Replicate, crediting fal.ai with the larger published endpoint count and the lowest baseline per-inference price, and positioning this platform for production teams needing high-concurrency infrastructure and compliance tooling. Treat it as vendor-side material. A directory review names Replicate, Together.ai and OpenRouter as the comparison set and frames the distinction as coverage breadth rather than raw model count.

That review also describes the pricing as usage-based with no published public tiers, expecting buyers to talk to sales. That conflicts with the current catalogue, which publishes per-model rates beside each provider's official price. The discrepancy may reflect a change since the review was written; either way, price against the live catalogue rather than a secondary summary.

Limitations

• No quota visibility in responses. Rate limits apply per account and per model, but the LLM and media endpoints return no rate-limit or retry-after headers. Remaining quota cannot be read from a response, so back-off has to be implemented client-side by inference rather than by instruction.
• A blind retry can bill twice. A generation submission that timed out may already have been accepted, so retrying it can create — and bill for — a second job. The documented mitigation is to submit asynchronously and poll, which is a design constraint on the client, not a setting.
• Streaming success is not proof of completion. Once a stream has opened, failures arrive as an event inside the stream rather than as an HTTP error, so a 200 status on a streaming call does not guarantee a complete response.
• Hard 50 MB request body ceiling. Bodies above it are rejected with 413, which pushes large inputs through the upload endpoint or a URL rather than inline data.
• Generated media cannot be kept beyond 14 days. Storage defaults to 14 days and the retention header maximum equals that default, so per-request settings can only shorten it, never extend it. After expiry the output URLs return 404, making downloads to your own storage a requirement rather than an optimisation for anything with a longer life.
• Resale is restricted. Partners may not expose the platform APIs directly to their end users and may not build a thin wrapper for raw resale; substantial added value is required. This constrains a common business model built on aggregated APIs.
• Benchmark publication requires written consent. Publishing benchmark results, latency or throughput metrics, or comparative evaluations to any third party needs prior written permission. Internal testing is permitted only if the results are not disclosed externally — a restriction that matters for teams whose procurement or engineering culture involves publishing comparisons.
• The compliance signals conflict. The site footer displays a SOC 2 certified and HIPAA compliant badge, while the data retention policy states that absent a separate executed written agreement, the vendor does not represent that it currently maintains SOC 1, ISO 27001 or any specific regulatory certification. The legal terms separately exclude HIPAA-regulated use. These vendor-published statements do not agree; settle any compliance requirement in writing before relying on either.
• Output carries no originality or non-infringement warranty. Generated content may not be unique or accurate, and no representation is made that it will not infringe third-party rights.
• The governing-law clauses do not match. The acceptable use policy names Delaware law and the exclusive jurisdiction of Delaware courts, while the legal terms name New York law and New York courts. The two documents are not reconciled anywhere, so which forum applies to a dispute is unsettled on the published text alone.

FAQ

Q1. Do I need a new SDK to use it?

Not for text. The chat endpoint is OpenAI-compatible, so pointing the OpenAI SDK at the platform base URL and swapping the model string is enough, with streaming, tool use and function calling working unchanged. Image, video and audio use a different base path and an asynchronous submit-and-poll pattern, so those do need their own integration.

Q2. Why did my request return 401 when the key works elsewhere?

Because the gateway authenticates before it routes, a typo in the endpoint path also produces 401 rather than 404. When a known-good key fails, check the URL — including whether the call should go to the /v1 chat path or the /api/v1 media path — before rotating the key.

Q3. How long are my generated files kept?

Fourteen days by default. A per-request header can shorten that to as little as one hour, but because the header maximum equals the default it cannot extend storage past 14 days. Once media expires the output URLs return 404, so anything you need to keep has to be copied to your own storage first.

Q4. Is my prompt data used to train models?

Content submitted through platform-controlled services is not used to train, fine-tune or evaluate the platform's own hosted models without written authorisation. That commitment does not extend to third-party providers, whose own terms apply where they process the content, and the vendor does not claim all providers maintain identical practices.

Q5. Can I get a refund on unused credits?

No. Credits are purchased in advance, expire 365 days from purchase, and are non-refundable and non-transferable. A remaining balance is also forfeited on termination or account deletion, so buying credits well ahead of consumption carries real risk.

Q6. Can I resell access through my own product?

Only with substantial added value. Exposing the platform APIs directly to your end users is prohibited, and building a thin wrapper for raw resale is explicitly not permitted, so a pure pass-through business is outside the terms.

Q7. Can I publish latency benchmarks comparing it with other providers?

Not without prior written consent. The usage terms restrict publishing benchmark results, performance metrics including latency and throughput, and comparative evaluations to third parties. Internal testing is allowed provided the results stay internal.

Q8. Is it certified for regulated workloads?

The signals conflict and should be settled in writing. A SOC 2 and HIPAA badge appears in the site footer, the retention policy declines to represent that any specific certification is currently maintained absent a separate executed agreement, and the terms state the services are not tailored to HIPAA or FISMA and may not be used where those laws apply.