Together AI

Together AI, styled as the "AI Native Cloud," is a full-stack platform for running, fine-tuning, and training open-source AI models at production scale. Rather than building its own closed frontier models, Together AI focuses on making the best open-source models, such as DeepSeek, Llama, Qwen, GLM, and Kimi, fast, affordable, and easy to deploy through serverless inference APIs, dedicated infrastructure, and its own vertically integrated GPU cloud. The platform's OpenAI-compatible API lets teams switch from closed-model providers with minimal code changes while accessing over 200 models across text, vision, image, audio, and video.

Founded in June 2022 by Vipul Ved Prakash, Ce Zhang, Chris Ré, and Percy Liang, Together AI is grounded in systems research from its team, including foundational work like FlashAttention, and channels that research directly into product performance gains such as 2x faster inference and up to 90% faster pre-training. Beyond serverless inference, the platform offers batch processing for large asynchronous workloads, provisioned throughput with SLA-backed capacity, dedicated single-tenant GPU deployments, and on-demand or reserved GPU clusters spanning NVIDIA H100 through GB300 hardware. For teams shaping their own models, Together AI provides LoRA and full fine-tuning, direct preference optimization, and evaluation tooling, all priced per token based on model size.

Together AI serves close to 500,000 developers and enterprises, with customers like Cursor, Decagon, Zoom, Quora, ElevenLabs, and Salesforce building production AI products on its infrastructure; Cursor uses the platform for real-time, low-latency inference at scale, while Decagon achieved 6x cost reduction per conversational turn versus GPT-5 mini. The company also contributes actively to the open-source ecosystem through datasets like RedPajama-V2 and ongoing research published at venues including NeurIPS, ICML, and ICLR, and offers a startup accelerator program to help early-stage companies build on open models.

Together AI uses pay-as-you-go pricing across its products. Serverless inference is billed per model, priced per 1M tokens for text (e.g., MiniMax M3 at $0.30 input/$1.20 output, GLM-5.2 at $1.40 input/$4.40 output, gpt-oss-120B at $0.15 input/$0.60 output), per image for image generation (e.g., FLUX.1 [schnell] at $0.0027/image), per video for video models (e.g., ByteDance Seedance 2.5 at $0.115/video, Google Veo 3.0 at $1.60/video), and per audio minute or character for speech models. Dedicated Inference runs on single-tenant GPUs starting at $5.49/GPU/hour on-demand for NVIDIA HGX H100 and $8.99/hour for HGX B200, with reserved options available via sales. GPU Clusters offer on-demand rates from $3.99/hour (H100) to $8.19/hour (B200), with reserved pricing dropping as low as $3.19/hour for 181+ day H100 commitments. Sandbox compute costs $0.0446/vCPU/hour and $0.0149/GiB RAM/hour, with Code Interpreter sessions at $0.03 per 60-minute session. Fine-tuning is priced per 1M tokens processed, ranging from $0.48 (LoRA, up to 16B parameters) to $8.00 (full fine-tuning, 70-100B parameters) for standard models, with specialized model pricing (e.g., DeepSeek-R1, GLM-5) ranging $5-$40 per 1M tokens plus a minimum job charge. Managed Storage costs $0.16/GiB/month.