Together AI

Together AI is an AI cloud platform built around model inference, customization, and compute. Its strongest focus is open and open-weight AI, giving developers a way to run models through APIs without building the underlying GPU infrastructure themselves. The platform also goes beyond simple API calls with batch inference, dedicated deployments, GPU clusters, and fine-tuning.

That makes Together AI different from an AI chatbot and different from a pure model marketplace. It is closer to development infrastructure for teams that want to build applications around open models and control more of the serving and customization workflow.

The current Together AI model library lists 200+ models across chat, image, vision, video, audio, transcription, code, embeddings, reranking, and moderation. Its catalog includes models from organizations such as Qwen, Meta, Mistral AI, DeepSeek, MiniMax AI, ZAI, Moonshot AI, and Black Forest Labs, with selected third-party media models also appearing in the catalog.

Together AI Is More Than a Model API

At first glance, Together AI can look like another service where you obtain an API key and call an LLM. That description misses most of the platform.

The service is split into several layers. Serverless Inference is the simplest entry point: choose a supported model, send requests through the API, and pay according to usage. Dedicated Inference moves the workload onto single-tenant GPU instances. Provisioned Throughput is designed around reserved capacity and predictable token throughput. GPU Clusters target larger training and inference jobs, while Fine-Tuning is aimed at adapting open models to specific tasks. Batch Inference handles large asynchronous workloads.

That structure matters because teams often start with a small API integration and later discover that they need more control. A prototype may be perfectly comfortable on serverless inference. A production workload with predictable traffic might benefit from reserved capacity. A custom model may need dedicated hardware. A specialized application may justify fine-tuning.

Together AI attempts to cover those stages without forcing the team to move to a separate infrastructure provider.

Its OpenAI-compatible API also reduces migration work. Together’s documentation shows the standard OpenAI client pointed at Together’s endpoint, allowing developers already familiar with that interface to make relatively small integration changes.

How the Inference Side Works

The simplest Together AI workflow is serverless inference.

A developer creates an API key, chooses a model from the catalog, sends a request, and receives the result. There is no GPU cluster to configure for this path. The platform handles serving and scaling behind the API.

The model library is useful here because it gives developers a central place to inspect available models and their supported modalities. The current catalog includes text and code models alongside image, video, audio, transcription, embeddings, reranking, and moderation options.

Together also emphasizes performance on serverless workloads. Its current serverless page advertises up to 2.75× faster inference and describes optimizations such as adaptive speculative decoding. Those are provider claims rather than independent benchmark results, so they should be read as performance positioning rather than a universal guarantee for every model and workload.

Another useful option is Batch Inference. Together says batch jobs can process up to 30 billion enqueued tokens per model per user, with most serverless batch workloads priced at up to 50% below the real-time API. The service is designed for asynchronous workloads where immediate responses are not necessary.

That could be a practical fit for large-scale summarization, classification, dataset processing, evaluation runs, or other jobs where millions of requests need to be processed efficiently.

Fine-Tuning Is a Major Part of the Platform

Fine-tuning is one of the areas where Together AI becomes more interesting than a standard inference API.

The platform supports model customization through techniques including supervised fine-tuning with LoRA and Direct Preference Optimization with LoRA. Pricing is based on the number of training and evaluation tokens processed, and minimum charges vary by model.

Together’s 2026 fine-tuning updates also expanded support for tool-calling, reasoning, and vision-language training. A March 2026 release stated that the service could train models with up to 1 trillion parameters, while a September update added support for newer open-weight models, live experiment tracking, finer training and data-processing controls, and lower pricing on selected models.

For developers, this changes the purpose of the platform. Instead of choosing a model only because it already performs well on a general benchmark, a team can start with an open model and adapt it to a specific domain or workflow.

There is also a useful ownership angle. Together says customers retain ownership of their model weights and intellectual property and can move those weights to another provider or self-host them later. That portability is important for companies that want a managed training and inference service without making the provider itself the permanent home of the model.

Dedicated GPUs Give Teams More Control

Serverless inference is convenient, but it is not the answer for every workload.

Together AI’s Dedicated Inference product provides single-tenant GPU instances with guaranteed performance, support for custom models, autoscaling, and traffic-spike handling. Current pricing lists NVIDIA HGX H100 at $5.49 per GPU-hour and HGX B200 at $8.99 per GPU-hour for on-demand dedicated inference on the pricing page.

Together also offers GPU clusters for larger-scale workloads. Current hardware options include H200 and B200 configurations, with different pricing depending on whether the capacity is on-demand or reserved. Together’s H200 page, for example, currently lists $5.99 per GPU-hour for on-demand use and lower rates for longer reservation windows.

This part of Together AI makes the service relevant to teams doing more than API inference. Researchers and ML engineers can move into dedicated hardware when they need custom serving, training, or greater control over compute.

The trade-off is straightforward: GPU infrastructure introduces more complexity and a much larger cost surface than simply paying per token. A small application may never need it.

Pricing Requires Looking at the Workload

Together AI does not use a conventional “$20 per month” subscription model for its main developer offering.

The current support documentation says Together uses prepaid credits, with a minimum $5 credit purchase. It also states that there is no current free trial and no negative balance or postpaid spending limit for standard credit-based usage.

For serverless inference, pricing is model-specific and generally calculated per million input and output tokens. Current examples include MiniMax M3 at $0.30 per million input tokens and $1.20 per million output tokens, DeepSeek V4 Flash 0731 at $0.14 input and $0.28 output, and Llama 3.3 70B at $1.04 input and $1.04 output.

Other workloads use different billing units. Images can be priced per image or megapixel, speech services can be priced per character or audio minute, video is priced per video, dedicated infrastructure uses GPU-hours, and fine-tuning is token-based.

This is one of Together AI’s biggest learning curves. There is no single answer to “How much does it cost?” because the answer depends on which product, model, hardware type, and workload you choose.

For small-scale API usage, the pay-as-you-go model is easy enough to understand. For larger deployments, teams should estimate traffic, token volume, concurrency, caching, batch workloads, and GPU requirements before deciding what the monthly bill might look like.

Privacy and Data Controls Matter More for Production

Together AI’s privacy policy provides a useful distinction between normal service operation and Zero Data Retention.

The company states that data collected from customers is not used to train its models without explicit opt-in and consent. It also provides a Zero Data Retention setting that can prevent prompts, submitted content, and outputs from being stored or used for training, product improvement, or secondary purposes beyond what is needed to provide the service.

That is a positive control for production teams, but privacy should still be treated as an implementation decision rather than a blanket assumption.

The exact model, region, deployment type, retention setting, and contract can matter. Teams handling sensitive information should review Together’s current privacy and contractual documentation before sending regulated or confidential data.

Who Fits Together AI Best?

Together AI is particularly relevant to developers who want open or open-weight models without managing every GPU and serving component themselves.

It fits applications that need API-based inference, model experimentation, large batch processing, fine-tuning, custom model deployment, or access to dedicated GPUs. It can also make sense for teams that want ownership of fine-tuned model weights and the option to move those weights elsewhere later.

It is less compelling when the requirement is simply “give me every major closed frontier model behind one API.” That is closer to the value proposition of a multi-provider gateway such as OpenRouter. Independent 2026 comparisons consistently place OpenRouter in the multi-provider routing category, while Together AI is described as an open-model inference, fine-tuning, and compute platform.

The same distinction applies to self-hosting. Platforms such as RunPod can provide more direct control over raw GPU infrastructure, while Hugging Face offers a broader model ecosystem. Fireworks AI overlaps more directly on hosted open-model inference.

Together AI’s appeal comes from putting several pieces together: model access, inference, fine-tuning, batching, and compute.

For teams building around open models, that makes the platform worth considering as a broader AI development environment rather than simply another place to call an LLM.