About Together AI
Together AI is an AI-native cloud platform for building and operating applications with open and custom models. Developers can access language, vision, image, audio, video, transcription, embedding, reranking, and moderation models through serverless APIs. Teams can begin with pay-as-you-go inference and move to batch processing, dedicated endpoints, custom containers, or GPU clusters as workloads grow.
The platform also provides fine-tuning, model evaluations, secure development sandboxes, managed storage, and research-optimized infrastructure. Its model library includes options from DeepSeek, Meta, Google, Qwen, Mistral, OpenAI, MiniMax, and other providers, with pricing that varies by model and workload. Readers comparing infrastructure and developer platforms can browse more AI development tools, automation tools, and data and analytics tools.
Key Features
Serverless inference : Runs supported open models through managed APIs without requiring customers to provision or maintain inference infrastructure.
Batch inference : Processes large asynchronous workloads with serverless or private deployments for data preparation, evaluation, and offline generation.
Dedicated inference : Deploys models on reserved infrastructure for teams that need predictable capacity, control, latency, and production economics.
Dedicated containers : Hosts custom models and generative media workloads on accelerated GPU infrastructure.
Model fine-tuning : Supports supervised fine-tuning, LoRA, full fine-tuning, and direct preference optimization across selected open models.
GPU clusters : Provides on-demand and reserved NVIDIA H100, H200, B200, GB200, and GB300 capacity for training and inference workloads.
Developer sandbox : Creates secure, scalable code environments for AI applications and agents, including snapshots and rapid resume workflows.
Evaluations and model shaping : Measures model quality and adapts behavior to domain data before production deployment.
Developer resources : Includes documentation, playgrounds, cookbooks, demos, model selection tools, voice-agent guidance, and Together Chat.
Pros
✔ Broad model library covers text, vision, image, audio, video, transcription, embeddings, reranking, and moderation.
✔ Serverless APIs let teams prototype without managing GPU infrastructure.
✔ Batch, dedicated, and container options support a path from experiments to high-volume production.
✔ Fine-tuning and evaluations are integrated with deployment infrastructure.
✔ On-demand and reserved GPU clusters support larger training and custom inference workloads.
✔ Research-driven optimizations target throughput, latency, and infrastructure efficiency.
Cons
✖ Costs vary by model, token direction, media type, hardware, and deployment option, making estimates more involved.
✖ Production teams must monitor token consumption and GPU hours carefully to control spend.
✖ Dedicated infrastructure and AI Factory arrangements require sales consultation or custom commitments.
✖ Model availability, names, rates, and performance can change as the catalog evolves.
✖ Fine-tuning requires suitable datasets, evaluation methodology, and ML engineering expertise.
✖ Organizations should review privacy, retention, security, and model-license terms for each workload.
Plans & Pricing
| Service | Type | Price | Usage Limit | Inclusions |
|---|---|---|---|---|
| Serverless Inference | Pay as you go | Varies by model and modality | Usage-based | Managed chat, vision, image, audio, video, transcription, embedding, reranking, and moderation APIs. |
| Example: MiniMax M3 | Token pricing | $0.30 input / $1.20 output per 1M tokens | Usage-based | Cached input is listed at $0.06 per million tokens on the official pricing page. |
| Example: gpt-oss-120B | Token pricing | $0.15 input per 1M tokens | Usage-based | Representative open-model serverless inference; verify its current output and cached-input rates online. |
| Standard Fine-Tuning | Training usage | From $0.48 per 1M processed tokens | $4 minimum job charge | Supervised LoRA for models up to 16B starts at $0.48; full fine-tuning and DPO cost more, with rates increasing by model size. |
| GPU Clusters | Hourly compute | H100 $3.99, H200 $5.99, B200 $8.19 per GPU/hour | On-demand or reserved capacity | Accelerated compute for training and inference; longer reserved terms may receive lower rates. |
| Dedicated / Enterprise | Custom | Contact sales | Custom capacity | Dedicated model or container inference, custom infrastructure, AI Factory, storage, support, and enterprise requirements. |
Source: Together AI pricing. Verify current model rates, cached and batch discounts, fine-tuning charges, GPU availability, minimums, and enterprise terms before deployment.
FAQs
Q1: What is Together AI used for?
Together AI is used to run, batch-process, fine-tune, evaluate, and deploy open or custom AI models on serverless APIs, dedicated infrastructure, and GPU clusters.
Q2: Which model types does Together AI support?
The platform supports chat, vision, image, audio, video, transcription, embeddings, reranking, moderation, and custom model workloads.
Q3: How does Together AI pricing work?
Serverless language models are generally billed per million input and output tokens, media models use modality-specific units, fine-tuning is billed by processed training tokens, and GPU clusters are billed per GPU-hour.
Q4: Can Together AI fine-tune models?
Yes. Together AI supports supervised fine-tuning, LoRA, full fine-tuning, and direct preference optimization for selected models and sizes.
Q5: Does Together AI offer dedicated inference?
Yes. Teams can use dedicated model inference or dedicated containers when they need reserved capacity, custom models, stronger control, and predictable production performance.
Q6: Which GPUs are available?
Together AI lists NVIDIA H100, H200, B200, GB200, and GB300 infrastructure, although availability and public hourly pricing differ by hardware.
Published on: July 5, 2026
