Published Updated
In one or two sentences, before anything else.
Atlas Cloud is a New York API service, run by ATLAS CLOUD AI INC., that sells one key and one prepaid balance for 400+ models: Claude, GPT, Gemini and Grok, open models it hosts, and media models such as Seedance, Kling and Veo. It charges Anthropic's list prices for Claude and takes Alipay, WeChat Pay and crypto.
What it can do, one sentence each, and whether we actually tested it.
One key and one prepaid USD balance for Claude, GPT, Gemini and Grok, self-hosted DeepSeek, GLM, Kimi, Qwen and MiniMax, and media models such as Seedance 2.5, Kling V3.0, Wan 3.0, Veo, Seedream 5.0 Pro and Suno.
Atlas Cloud claims 400+ models. Its public LLM list had 126 entries on October 21, 2026, and 85 of them, including every Claude, GPT, Gemini and Grok entry, were marked is_ready: false, which it does not explain.
https://api.atlascloud.ai accepts OpenAI Chat Completions, Completions, Responses and Images, Anthropic Messages and Google Gemini generateContent with one key.
Per Atlas Cloud's protocol docs: Anthropic server tools such as web search are unavailable, and cache_control is ignored when a model is served through a translated protocol.
atlascloud-mcp gives MCP clients 13 tools: docs search, model lists, image, video and audio generation, LLM chat, media upload, job status, balance, usage and costs.
Runs locally through npx. Atlas says it checks parameters against each model's schema first, so invalid requests fail without spending credits.
AtlasCloudAI/atlas-cloud-skills adds an atlas-cloud core skill (SKILL.md with API references), a Seedance 2.5 skill and a video-prompt skill, plus 25 recipes, for Claude Code, Codex CLI and Gemini CLI.
Installed with npx skills add; the skills need an Atlas Cloud API key.
Monthly plans from 10 to 500 USD with weekly point caps for coding agents such as Claude Code, Codex, Cursor and OpenCode, which Atlas says cost 45% to 65% of the underlying model price.
Paid through Stripe only and not usable for image, video or audio. Atlas's setup page uses GLM-5.1 and does not say whether Claude or GPT are included.
Failed LLM requests and failed, timed-out or moderated media jobs are not billed.
Per Atlas Cloud's billing docs. One Trustpilot reviewer says support corrected an overcharge.
Request headers can set media files and request records to expire after 0 to 336 hours.
Upstream model providers keep their own retention, which these headers do not change.
Top-ups by card through Stripe, Alipay, WeChat Pay, cryptocurrency or wire transfer, with invoices on request and credit lines for enterprises.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
About 4,000 tool-calling requests replayed against Atlas Cloud's Kimi K2-0905 and K2 Thinking endpoints and compared with Moonshot's own API.
What happened
On K2-0905 Atlas passed the trigger bar, but only 925 of 1,277 tool calls (72.44%) had valid JSON schemas, second-lowest of 16 entries; on K2 Thinking it passed with 95.74%. The test is a year old and covers one model family.
Case 2
Input
Add atlascloud-mcp to Claude Code with npx -y atlascloud-mcp and ATLASCLOUD_API_KEY, ask for a Seedance image-to-video clip from a local picture, and collect it with atlas_get_prediction.
What happened
The server uploads the picture for a temporary URL, checks the parameters before submitting, and bills the job to the same balance. Jobs are asynchronous, so the agent polls for the result.
Case 3
Input
Claude Code with ANTHROPIC_BASE_URL set to https://api.atlascloud.ai (no /v1), an Atlas key and the model zai-org/glm-5.1.
What happened
Claude Code speaks Anthropic Messages, which Atlas accepts. Atlas's protocol docs say Anthropic server tools such as web search are not available and cache_control is ignored when a model is served through a translated protocol.
Case 4
Input
The same Wan 3.0 video request run on Atlas Cloud and on Runware.
What happened
The reviewer said Atlas Cloud's Wan 3.0 output looked like a weaker model than Runware's. One user's claim with no clips or numbers shared.
What it does badly, with the evidence, and what we could not settle.
Weak
Poor tool-call schemas on Kimi K2-0905.
Evidence: Moonshot's K2 Vendor Verifier (November 2025) found 72.44% valid schemas, second-lowest of 16 hosts; Moonshot, DeepInfra, Fireworks and Novita AI scored 100%.
The Gemini 4 Argon discount is overstated.
Evidence: Atlas labels Gemini 4 Argon at 1.60 / 8.00 USD per million tokens as 60% off an official 4 / 20 USD, but Google charges 2 / 10 USD until November 30, 2026, so the discount on October 21 was 20%.
Non-refundable credits that expire.
Evidence: Top-ups start at 25 USD and are final, non-refundable and non-transferable; credits expire 365 days after purchase and any balance is forfeited on account closure.
No published rate limits, status page or pay-as-you-go SLA.
Evidence: Atlas's docs give no rate-limit numbers and no rate-limit headers. They claim 99.9% API uptime, but the SLA covers only instance services under a signed order, and no status page was found.
Contradictory compliance claims.
Evidence: Atlas's policies page says it holds no SOC 1, SOC 2 or ISO 27001 certification and is not built for HIPAA, while its ask.atlascloud.ai page claims SOC 2 and HIPAA compliance.
Its terms bar publishing benchmarks without consent.
Evidence: The Acceptable Use Policy forbids publishing benchmark results without Atlas Cloud's written consent. The terms name New York law; the Acceptable Use Policy names Delaware.
Not for buyers who need compliance or a clear source for each route.
Evidence: Anyone who needs certified compliance, refunds, published rate limits or a documented source for every Claude, GPT and Gemini route should use the makers' own APIs or a provider that states these terms.
Uncertain
What the -ccmax model ids are.
Why unsure: The model list includes ids ending in -ccmax, named Claude Sonnet 5.5 CC Max and Claude Haiku 5.5 CC Max. Atlas Cloud does not explain what CC Max means or how those routes are supplied.
What is_ready: false means.
Why unsure: 85 of 126 LLM entries, including every closed model, carry is_ready: false, and Atlas's docs do not define the flag.
How the -developer Gemini variants are priced below Google.
Why unsure: Gemini 4 Argon -developer sells at 1.20 / 6.00 USD, 40% under Google's current list, with no explanation.
Whether ByteDance confirms the reseller claim.
Why unsure: Atlas calls itself an authorized U.S. reseller of Seedance 2.5 and Seedream 5.0 Pro; no statement from ByteDance or BytePlus confirming it was found.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Operator | ATLAS CLOUD AI INC., 1540 Broadway, New York, NY; founder and CEO Jerry Tang | Funding not disclosed; PitchBook lists an early-stage deal with VCV Digital Group. Source: atlascloud.aichecked |
| Claude on Atlas Cloud, per million tokens | Opus 5.5 4 / 20 USD, Sonnet 5.5 2 / 10 USD, Haiku 5.5 0.10 / 0.50 USD | The same as Anthropic's list prices. Source: atlascloud.aichecked |
| Gemini 4 Argon on Atlas Cloud, per million tokens | 1.60 USD input, 8.00 USD output, labelled 60% off | Google's price is 2 / 10 USD until November 30, 2026 and 4 / 20 USD from December 1, so the discount on October 21 was 20%. Source: atlascloud.aichecked |
| Platform fee | None stated; failed LLM requests and failed, timed-out or moderated media jobs are not billed | Source: atlascloud.aichecked |
| Minimum top up and free credit | 25 USD per top-up; credits expire after 365 days; no standing free tier | Atlas's credits page lists referral bonuses and an open-source sponsorship program instead. Source: atlascloud.aichecked |
| Payment methods | Cards through Stripe, Alipay, WeChat Pay, cryptocurrency; wire transfer and invoices for large amounts | Source: atlascloud.aichecked |
| Models | 400+ claimed; 126 LLM entries in the public list, 85 marked is_ready: false | Source: api.atlascloud.aichecked |
| API formats | OpenAI Chat Completions, Completions, Responses and Images; Anthropic Messages; Gemini generateContent; tool calling and structured outputs | No Anthropic server tools such as web search; cache_control ignored on translated protocols; no rate-limit headers. Source: atlascloud.aichecked |
| MCP | Official atlascloud-mcp server with 13 tools for docs, models, media generation, chat, balance and usage | Source: atlascloud.aichecked |
| Agent Skills | AtlasCloudAI/atlas-cloud-skills: atlas-cloud, seedance-2-5-skill and universal-video-prompt-skill | Source: atlascloud.aichecked |
| Coding agents | Claude Code, Codex, Cursor, OpenCode, OpenClaw, Hermes and Trae setups; MCP setups for many more clients | Source: atlascloud.aichecked |
| Privacy and data retention | No training on content without written permission; content kept only briefly; per-request expiry from 0 to 336 hours; hosted in the US | The policies page disclaims SOC 2, ISO 27001 and HIPAA; DPA and zero-retention addendum for enterprise. Source: atlascloud.aichecked |
| Status page and uptime | No status page found; OpenRouter one-day uptime 99.11% to 100% on four open models | The docs claim 99.9% API uptime, but the SLA covers only instance services under a signed order. Source: openrouter.aichecked |
| Refunds | Top-ups are final, non-refundable and non-transferable; balance forfeited on account closure | Source: atlascloud.aichecked |
| Website visits, September 2026 (atlascloud.ai) | 981,000 (−30.4% on August); 9th of 10 on AICPB's AI model routers list | Source: mp.weixin.qq.comchecked |
Answer first, then why, an example, the limits and what to do next.
Atlas Cloud is a real New York company with public terms and paying users, but several of its claims do not line up. Its policies page disclaims SOC 2 and HIPAA while another Atlas page claims both, its Gemini discount is measured against a future Google price, and it does not explain its -ccmax Claude model ids.
Limits
Its Acceptable Use Policy forbids publishing benchmarks without its written consent, so little independent testing exists.
Atlas Cloud charges the makers' list prices for Claude and most closed models, so it is not cheaper there. Claude Sonnet 5.5 costs 2 / 10 USD per million tokens on both; Gemini 4 Argon is 20% under Google's current price, and some open models were cheaper through its OpenRouter endpoints.
Why
The Coding Plan, sold separately, claims 45% to 65% of the underlying price for coding models.
Atlas Cloud publishes an official MCP server, atlascloud-mcp, with 13 tools. Coding agents such as Claude Code, Codex, Cursor and Gemini CLI can use it to list models, generate images, video and audio, chat with LLMs and check balance and usage; it also publishes Agent Skills for Claude Code, Codex and Gemini CLI.
Claude Code works with Atlas Cloud through its Anthropic Messages endpoint at https://api.atlascloud.ai. Atlas's setup page uses GLM-5.1 as the model; Anthropic server tools such as web search are unavailable, and prompt caching is ignored when a model is served through a translated protocol.
Atlas Cloud top-ups are final and non-refundable. Failed LLM requests and failed, timed-out or moderated media jobs are not billed, but unused credits expire 365 days after purchase and any balance is forfeited if you close the account.
The main alternatives are fal, Replicate, WaveSpeedAI and Kie.ai for media models, and OpenRouter for language models. fal and Replicate price Nano Banana Pro at 0.15 USD per 1K image against Google's 0.134 USD; Kie.ai is cheaper but does not say how it gets access.
This page is built from Atlas Cloud's site, docs, policies and model list, the makers' price pages, OpenRouter's endpoint data, public tests by Moonshot AI and the SGLang team, Trustpilot reviews and Hacker News comments, all checked on October 21, 2026. Atlas Cloud's claims are labelled as its own.
According to AICPB's traffic rankings, atlascloud.ai had 981,000 visits in September 2026, down 30.4% on August: 9th among AI model routers.
Atlas Cloud is run by ATLAS CLOUD AI INC. from New York, with Jerry Tang as founder and CEO; PitchBook dates it to 2024. It began as a GPU cloud and SGLang-based inference host and now sells one API key and one prepaid US dollar balance for what its home page calls 400+ models. Its public LLM list had 126 entries on October 21, 2026: 32 from Anthropic, 23 from OpenAI, 15 from Google, 12 Qwen, 11 ByteDance, 9 DeepSeek and smaller groups from Z.ai, xAI, Moonshot and MiniMax. 85 of them, including every Claude, GPT, Gemini and Grok entry, are marked is_ready: false, which Atlas does not explain.
The home page leads with media. Video models include Seedance 2.0 and 2.5, Kling V3.0, Wan 2.7 and 3.0, Veo and Vidu; image models include Seedream 5.0 Pro, GPT Image 2 and Nano Banana Pro; audio includes Suno and ElevenLabs. Atlas calls itself an authorized U.S. reseller of ByteDance's Seedance 2.5 and Seedream 5.0 Pro. Atlas is also a provider inside OpenRouter, with 30 models.
Some model ids carry suffixes that name a route: -aws for Claude through AWS, -az for GPT through Azure, and cheaper -developer variants of Gemini. The list also includes ids ending in -ccmax, named "Claude Sonnet 5.5 CC Max" and "Claude Haiku 5.5 CC Max". Atlas Cloud does not explain what CC Max means or how those routes are supplied.
| Model, per million input / output tokens | Atlas Cloud | Maker's list price |
|---|---|---|
| Claude Opus 5.5 | 4 / 20 USD | 4 / 20 USD (Anthropic) |
| Claude Sonnet 5.5 | 2 / 10 USD | 2 / 10 USD (Anthropic) |
| Gemini 3.8 Flash | 0.75 / 3.75 USD | 0.75 / 3.75 USD (Google) |
| Gemini 4 Argon | 1.60 / 8.00 USD, labelled 60% off | 2 / 10 USD until November 30, 2026; 4 / 20 USD from December 1 (Google) |
| DeepSeek V4 Pro 0813 | 1.32 / 3.96 USD | 1.32 / 3.96 USD peak, 0.66 / 1.98 USD off-peak (DeepSeek) |
Atlas Cloud's prices come from its pricing page. Claude and most closed models cost the same as the makers' lists. The Gemini 4 Argon discount is measured against Google's price from December 1, 2026; against the price in force on October 21 it is 20%, or 40% for the -developer variant at 1.20 / 6.00 USD. On OpenRouter some Atlas endpoints were cheaper than its own list, such as DeepSeek V4.1 Flash at 0.135 / 0.54 USD against 0.30 / 1.20 USD.
For media, Atlas listed Wan 2.7 image-to-video at 0.10 USD a second at 720p and 0.15 USD at 1080p in July 2026, and Nano Banana Pro edits at 0.14 USD per image up to 2K, against Google's 0.134 USD. A separate Coding Plan, 10 to 500 USD a month with weekly point caps, claims 45% to 65% of the underlying model price; its credits cannot be used for media.
No Artificial Analysis provider page or other independent latency benchmark of Atlas Cloud's API has been published. Buyers should know that its Acceptable Use Policy forbids publishing benchmark results without Atlas Cloud's written consent.
Trustpilot shows 3.6 out of 5 from 6 reviews. Reviewers praise the prices and video range; complaints cover poor video output, argumentative support, and one user's claim (October 1, 2026) that Wan 3.0 seems routed to a weaker model than on Runware. On Hacker News, one OpenRouter user put AtlasCloud on an ignore list of providers that "always suck" (September 2026), while others use it for GLM 5.3 Flash or praise its coding plan's limits. One commenter said in October 2025 that proxies including AtlasCloud did not enforce OpenAI's ID verification for restricted models. Several promotional HN comments in February 2026 carried tracking links to Atlas.
atlascloud-mcp (GitHub: AtlasCloudAI/mcp-server), run locally with npx -y atlascloud-mcp. Its 13 tools search the docs, list models, generate images, video and audio, chat with LLMs, upload media, poll jobs and read balance, usage and costs. Setups cover Claude Code, Codex CLI, Gemini CLI, Goose, Cursor, Claude Desktop, Windsurf, Kilo Code, OpenCode and others. Passing MCP tools inside an API request is not documented.atlas-cloud (a SKILL.md entry point with API references), seedance-2-5-skill and universal-video-prompt-skill, plus 25 recipes, installed with npx skills add AtlasCloudAI/atlas-cloud-skills for Claude Code, Codex CLI and Gemini CLI.https://api.atlascloud.ai; Codex, Cursor, OpenCode, OpenClaw, Hermes and Trae use the OpenAI-compatible endpoint. The examples use GLM-5.1, and the page does not say whether the Coding Plan covers Claude or GPT.generateContent. Tool calling works in Chat Completions, Responses, Messages and Gemini; structured outputs in Chat Completions and Responses. Anthropic server tools such as web search are unavailable, cache_control is ignored on translated protocols, and Responses ignores previous_response_id, so you resend the whole conversation.Atlas says request content is not used to train or benchmark hosted models without written permission and is kept "only briefly"; media files and request records can be set to expire per request, from 0 to 336 hours, while upstream providers keep their own retention. Services are hosted in the US, and enterprises can ask for a DPA, a zero-retention addendum and a subprocessor list. Its policies page says it holds no SOC 1, SOC 2 or ISO 27001 certification and is not built for HIPAA, while its ask.atlascloud.ai page claims SOC 2 and HIPAA compliance.
Top-ups start at 25 USD and are final, non-refundable and non-transferable; credits expire after 365 days and a balance is forfeited on account closure. Rate limits are not published. The terms name New York law; the Acceptable Use Policy names Delaware. The docs claim 99.9% API uptime, but the SLA covers only instance services under a signed order, and no status page was found.
Skip Atlas Cloud if you need certified compliance, refunds, published rate limits or a documented source for every closed-model route. It fits developers who want Seedance, Kling, Wan and Veo next to Claude and GPT on one balance, especially those paying with Alipay, WeChat Pay or crypto.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
WaveSpeedAI is an API platform, run by a Singapore company and a Hong Kong company named in its terms, that sells 1,000+ image, video, audio and 3D models plus an LLM service on one balance. It matches fal on Kling 3.0, charges above Google's list for Veo 3.1 Fast and accepts Alipay and WeChat Pay.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
Replicate is an API for running AI models, owned by Cloudflare since a $57.4 million acquisition closed on December 1, 2025. It sells 100+ official models such as Veo 3.1, Nano Banana Pro and Kling 3.0 per output, and 50,000+ community models by the second of GPU time.
fal is a San Francisco API platform, run by Features & Labels, Inc., that sells per-output access to more than 1,000 image, video, audio and 3D models on one prepaid balance. Its open models are cheap, but Google partner models such as Veo 3.1 Fast and Nano Banana Pro cost more on fal than on Google's own API.
Eden AI is a Lyon-based API gateway, run by Eden AI SAS, that puts about 500 language models and 250 'expert' models for OCR, speech, translation and image work behind one key and one prepaid balance. It passes provider prices through and, since January 2026, adds a 5.5% platform fee on self-serve credits.