Loading tool details…
Complete pricing guide for Together AI. Compare all plans, analyze costs, and find the perfect tier for your needs.
Not sure if free is enough? See our Free vs Paid comparison →
Still deciding? Read our full verdict on whether Together AI is worth it →
Pricing sourced from Together AI · Confirm current pricing and terms with the vendor.
A check mark means listed in that plan’s record. A dash means not listed, rather than confirmed unavailable. Verify current vendor terms.
Detailed feature comparison coming soon. Visit Together AI's website for complete plan details.
View Full Features →Groq specializes in ultra-low-latency inference on custom LPU hardware — often 3-10x faster tokens/sec but with a smaller model catalog. Fireworks AI is Together's closest peer in scope, with a similar catalog and strong function-calling. Together tends to win on catalog breadth (Llama 3.3/4, DeepSeek, Qwen 3, Kimi K2, Mistral) and on the three-tier deployment story (serverless, dedicated, cluster). Groq wins on latency; Fireworks and Together trade blows on price per model. Teams often A/B test the same model across all three.
Llama 3.3 70B Instruct is roughly $0.88/M tokens blended (input+output) on serverless. DeepSeek-V3 is around $1.25/M. Together Turbo variants offer quantized inference at a discount for latency-tolerant workloads. All numbers move as new SKUs land, so check the pricing page before signing long-term contracts. Compared to Claude Sonnet or GPT-4-class closed models, Together is meaningfully cheaper — often 3-10x — with the tradeoff that top-end reasoning quality still slightly favors the frontier closed models.
Yes. Together supports both LoRA and full-parameter fine-tuning on most catalog models, priced per training token, with your dataset (JSONL). After training, the tuned model is available on a dedicated endpoint or as a serverless deployment depending on plan. This is one of the platform's biggest differentiators versus most closed-model providers, which offer fine-tuning only on select models with less flexibility.
Together offers HIPAA-eligible environments and is SOC 2 Type II certified for enterprise customers. BAAs are available on enterprise plans. If you're handling PHI, that setup lives inside a dedicated environment rather than the shared serverless tier, and pricing is quote-based.
Serverless is best for variable or bursty traffic — you pay per token with no infra to manage, but latency and throughput are shared. Dedicated endpoints reserve GPUs for you, giving consistent latency and higher throughput at an hourly per-GPU price. Rule of thumb: if you're running steady inference above a few million tokens per day for a single model, dedicated usually costs less than serverless and performs better. Below that, serverless is simpler and cheaper.
AI builders and operators use Together AI to streamline their workflow.
Try Together AI Now →Ultra-fast LLM inference cloud powered by custom LPU (Language Processing Unit) chips — the fastest hosted API for Llama, Kimi K2, Qwen, and many open-source models.
Compare Pricing →Serverless cloud for running Python code, AI models, and GPU workloads at scale with per-second billing.
Compare Pricing →