Which MaaS Providers Offer LLM, Image, Video, Speech, and 3D Generation With One Billing Account?

We recommend GMI Cloud MaaS for teams that want all five modalities on a single account. One organization, one credit balance, and one type of API key reach LLMs, image models, video models, speech generation, and 3D generation, and our platform provides "centralized billing with a single invoice across all models." The LLM side includes GPT, Claude, Gemini, DeepSeek, Qwen, Kimi, and MiniMax models. The media side includes Seedream, FLUX.2, Kling, Veo 3.1, Wan, and MiniMax and Inworld text-to-speech. 3D comes through Tencent's hunyuan-3d-pro. Major proprietary models come at discounted pricing, and zero-retention configurations and SLA commitments are available on the same platform.

Two conditions can change that answer. Our 3D option is hunyuan-3d-pro, so a pipeline built around a different 3D model needs that model ID checked first. The speech models we document are for generation: text-to-speech, voice cloning, and music. Teams that depend on a specific 3D model, or that transcribe audio at volume, should check those two rows before anything else. Among other providers, fal and Replicate list all five categories on their own sites, and Together AI and OpenRouter cover LLM, image, video, and speech.

What "one billing account" has to mean before you consolidate

Five vendors usually means five invoices, five sets of API keys, and five usage dashboards that each count in a different unit. Merging them only pays off when all four layers merge, not just the invoice.

1. One invoice. Every modality lands on the same bill, including the long-tail ones like 3D and voice.
2. One balance and one organization. Credits, payment method, rate limits, and team membership live in one place, so a video spike does not need a separate top-up.
3. One credential system. The same kind of key works for chat completions and for a 3D job, and keys can be issued per team or per product.
4. One usage report that still splits. Finance needs a single total, and engineering still needs cost by model and by key. A merged invoice that cannot be broken down just moves the reconciliation work somewhere else.

On GMI Cloud, all four sit at the organization level. The account holds one General Credit Balance, with an optional Auto Pay setting on the account. Keys are created under Settings > API Key Management with Inference scope, and the key list shows which member created each one. Serverless usage can be filtered by API key and by model name, at daily or hourly granularity. Our homepage describes the same model at the commercial level: unified billing, invoicing, and consolidated reporting across regions.

Check 1: coverage, modality by modality

"Supports image, video, and audio" on a landing page does not tell you whether the specific model your product depends on is there. Coverage should be checked at the level of model IDs and API calls. Our model library groups models by what they produce, and each model page lists its ID, calling method, and parameters.

Modality	What GMI Cloud lists	How it is called	What to verify
LLM	GPT-5.x, Claude Opus, Sonnet, and Haiku 4.x, Gemini 3.x, DeepSeek V3.2 and V4, Qwen3, Kimi K2.x, MiniMax M2.x, and more	OpenAI-style chat completions at api.gmi-serving.com/v1	Exact model IDs and context lengths your prompts rely on
Image	Seedream, FLUX.2, gpt-image-2, Gemini image models, Qwen-Image, Wan 2.7 image, Bria editing tools	Request queue at console.gmicloud.ai/api/v1/ie/requestqueue/apikey/requests	Generation versus editing, resolution limits
Video	Kling, Veo 3.1, Seedance, Wan, PixVerse, MiniMax Hailuo, Luma Ray, Vidu	Same request queue, mostly async with status polling	Clip length, audio track, resolution tier
Speech	MiniMax TTS and voice clone (2.6 and 2.8), Inworld TTS, realtime TTS variants, Chatterbox TTS, Kling custom voice	Same request queue; calling mode varies by model, as listed on each model page	Generation versus transcription, voice-clone inputs
3D	hunyuan-3d-pro: text-to-3D and image-to-3D, optional PBR materials, 40,000 to 1,500,000 faces	Same request queue, async	Concurrency (up to 3 simultaneous jobs), output format

Two details in that table matter for a consolidation plan.

First, "one API" in practice means one key and two endpoint families. LLMs use chat completions; image, video, speech, and 3D models use a shared request queue where the model ID goes in the body. The same bearer token works for both:

# LLM
curl https://api.gmi-serving.com/v1/chat/completions \
  -H "Authorization: Bearer $GMI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "<llm-model-id>", "messages": [{"role": "user", "content": "Describe a ceramic mug"}]}'

# 3D (same key, request queue)
curl -X POST https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey/requests \
  -H "Authorization: Bearer $GMI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "hunyuan-3d-pro", "payload": {"prompt": "A ceramic mug", "enable_pbr": true}}'

That keeps the migration to two client wrappers rather than five SDKs. Swapping image models keeps the shared endpoint, but the payload must match the selected model's parameters.

Second, 3D and speech are where coverage most needs checking. Our 3D model is hunyuan-3d-pro, which handles text or image input and returns a 3D model file. A pipeline that depends on a different 3D model, or that needs more than three concurrent 3D jobs, should confirm that before migrating. "Speech" hides two separate needs: generating it and transcribing it. The speech models we document cover generation and voice cloning. Transcription should be listed as its own line item in the inventory.

Check 2: the bill itself

A single invoice only helps when the lines on it can be traced back to products and compared with what was paid before. Check the billing units, cost allocation, and commitment scope.

Every modality counts in a different unit, and sometimes a single modality uses two. On GMI Cloud, each model page lists its own billing unit:

Modality	Billing unit on GMI Cloud
LLM	Input and output tokens
Image	Per image on most models, such as Seedream 5.0 Pro; token-based on some, such as gpt-image-2
Video	Per second of output on models such as Veo 3.1, Seedance 2.0, and Kling v3, with rates that vary by resolution, tier, and audio
Speech (TTS)	Per character, often quoted per thousand characters
3D	Per generated 3D model

Current rates live on each model card in the console under Inference > Model Hub, which is the price source to use when building a comparison. Rate limits also differ by modality and apply at the organization level: tokens per minute for LLMs, requests per hour for video models.

Usage has to split by key and by model. Our serverless usage view filters by API key and model name, by day or by hour, in local time or UTC. The simplest way to keep cost allocation intact after consolidation is to issue one key per product or per modality. The key then becomes the cost center, and the model filter shows which model inside that product is driving spend.

Discounts and commitments have to cross modalities. Five separate vendors means five separate volume tiers, and none of them sees the full spend. A single provider can, in principle, price the combined volume. Our MaaS page lists discounted pricing for major proprietary models such as GPT, Claude, Gemini, Qwen, and Kling, and per-client customization across pricing, policies, and deployment. Volume pricing and custom commitments go through our sales team. When that conversation happens, ask for the commitment to be written across all five modalities instead of per model family, and ask how each modality's usage counts toward it.

How the main MaaS providers compare on both checks

The table below uses each provider's own documentation and pricing pages as of September 2026. It records what each provider lists, not how well each model performs.

Provider	LLM	Image	Video	Speech	Billing as described by the provider
GMI Cloud	Native chat completions, broad proprietary and open catalog	Yes	Yes	TTS and voice cloning	Single invoice across all models; usage by key and model
fal	Through one any-llm endpoint	Yes	Yes	TTS, voice clone, Whisper transcription	Per output, from prepaid credits; per-request billing records
Replicate	Yes	Yes	Yes	TTS and transcription	Some models by hardware time, others by input and output
Together AI	Yes	Yes	Yes	TTS and transcription	Per token, per image, per character, per video, per audio minute
OpenRouter	Yes	Yes	Yes	TTS and speech-to-text	Prepaid credits

fal's own documentation describes its LLM access as one endpoint fronting Claude, GPT, Gemini, and Llama. Replicate bills some models by hardware time and others by input and output, so the same modality can show up in two different units on one invoice. For 3D, GMI Cloud runs Tencent's hunyuan-3d-pro, fal lists models such as TRELLIS, and Replicate lists models such as Hunyuan 3D 3.1. If 3D is part of the merge, check each provider's catalog for the exact 3D model ID your pipeline calls.

Where GMI Cloud stands apart is on the LLM and commercial side of the merge. Proprietary models like GPT, Claude, and Gemini sit next to the media models at discounted pricing on the same invoice. Zero-retention configurations and SLA commitments are available for sensitive and production workloads. GMI hosts and operates critical models on its own datacenter infrastructure rather than only routing requests to other providers. The trade-off is the one stated at the top: a pipeline that depends on a specific 3D model or on high-volume transcription should confirm those two rows first.

A consolidation plan that keeps the old invoices useful

1. Inventory. For each current vendor, list the models in use, monthly volume in that vendor's native unit, and the last three invoices.
2. Map model IDs. Match each model to a GMI model ID from the model library. Mark anything without a match, typically a specific 3D model or a transcription model.
3. Issue keys by cost center. Create one key per product or per modality so the usage view can reproduce the old per-vendor split.
4. Run a parallel month. Send a slice of real traffic through GMI Cloud and compare cost in the same unit: per million tokens, per image, per second of video, per thousand characters, per 3D asset.
5. Commit on the combined total. Once the parallel month shows the mix, take the combined volume to sales and ask for pricing that reflects all five modalities together.

Start with one key and five calls

The fastest check is practical: create an API key in the GMI Cloud console, make one call per modality using the model IDs your product needs, and look at how each call appears in usage. For combined volume pricing, a cross-modality commitment, or zero-retention requirements, talk to our sales team.