Together AI Pricing Explained: Plans, Costs and What You Get

Together AI pricing explained: serverless per-token rates, dedicated endpoints, GPU clusters and fine-tuning, with rates verified on the official pricing page.

By Galactic News Editorial · Published 2026-11-10T07:24:18+00:00 · Updated 2026-11-10T07:24:18+00:00

Together AI charges by use, not by subscription. Serverless inference bills per million tokens. Dedicated endpoints and GPU clusters bill per GPU hour. Fine-tuning bills per million training tokens. There is no per-seat fee for serverless use. This guide explains each pricing tier with rates taken from Together AI's own pricing page, so you can estimate a real monthly cost.

The short version: small models cost cents per million tokens. Large frontier models cost dollars. A dedicated H100 endpoint costs $5.49 per hour. A GPU cluster H100 costs $3.99 per hour on demand. The right tier depends on volume, latency needs, and how much control you want.

Together AI pricing at a glance

Together AI groups its pricing into serverless inference, provisioned throughput, dedicated inference, GPU clusters, fine-tuning, and sandboxes. Serverless is the default starting point. You send requests to an API. Together AI runs the model. You pay per token. No capacity planning is needed. No idle GPU costs apply.

Provisioned throughput reserves capacity in throughput units. It suits steady, high-volume traffic. Dedicated inference reserves a whole GPU for one customer's model. GPU clusters rent raw NVIDIA hardware by the hour. Fine-tuning trains an open model on your data. Each tier is priced separately. Most teams start serverless and move up the ladder as volume grows.

Serverless inference: what a million tokens costs

Serverless rates vary widely by model. Small models are cheap. Large reasoning models cost far more. Output tokens cost more than input tokens on most models. Cached input costs less than fresh input. The pricing page lists a cached rate for many models.

At the low end, Ternary Bonsai 27B is listed at $0.00 for input and output. The gpt-oss-120B model costs $0.15 per million input tokens and $0.60 per million output tokens. DeepSeek V4.1 Flash costs $0.30 input and $1.20 output, with cached input at $0.006. Qwen3.7-Plus costs $0.32 input and $1.28 output.

Mid-range models cost more. DeepSeek V4 Pro costs $1.32 per million input tokens and $3.96 output. GLM-5.3 costs $1.40 input and $4.40 output. Qwen3.7-Max costs $2.50 input and $7.50 output. At the top of the list, Kimi K3 costs $3.00 per million input tokens and $15.00 output, with cached input at $0.30.

Image, audio, and video models bill differently. Image models bill per image. FLUX.2 [pro] costs $0.03 per image. GPT Image 2 costs $0.053 per image. SD XL costs $0.0019 per image. Text-to-speech bills per million characters. Cartesia Sonic-3 costs $65.00 per million characters. Kokoro-82M TTS costs $4.00 per million characters. Transcription bills per audio minute. Whisper Large v3 costs $0.0015 per minute.

Serverless requests are best-effort. Under high demand they can be throttled. The Batch API suits work that does not need an instant response. It bills below real-time rates for eligible models. Check the pricing page for the batch rate on the model you use.

Dedicated inference and GPU clusters

Dedicated inference puts a model on GPUs reserved for one customer. It suits production workloads that need steady latency. The H100 dedicated endpoint costs $5.49 per GPU hour on demand. The B200 dedicated endpoint costs $8.99 per GPU hour. Reserved rates are quoted by sales.

GPU clusters rent raw hardware. The H100 costs $3.99 per GPU hour on demand. The H200 costs $5.99. The B200 costs $8.19. The B300 costs $9.99. Preemptible nodes cost less. The H100 preemptible rate is $1.99 per GPU hour. The H200 preemptible rate is $2.99. The B200 preemptible rate is $4.09.

Reserved clusters cost less with longer commitments. A 91-to-180-day reservation drops the H100 to $3.19 per GPU hour. The same term drops the B200 to $6.79. Shorter terms sit between on-demand and the longest term. Enterprise capacity is quoted directly.

Fine-tuning and sandbox pricing

Fine-tuning bills per million tokens processed in training. The price depends on the model and the method. Supervised fine-tuning on Llama 3.1 8B costs $0.34 per million tokens. Llama 3.3 70B costs $2.03. Llama 4 Maverick costs $5.60. DeepSeek-V3.1 costs $7.00. Each job carries a minimum charge shown on the pricing page. For Llama 3.1 8B the minimum is $4.00.

Direct preference optimization costs more than supervised fine-tuning. For Llama 3.1 8B it costs $0.84 per million tokens. For Llama 3.3 70B it costs $5.08. For DeepSeek-V3.1 it costs $17.50. The price is based on the sum of tokens in the training dataset times the number of epochs, plus any evaluation tokens.

Code sandboxes bill by compute. Each vCPU costs $0.0446 per hour. Each GiB of RAM costs $0.0149 per hour. A code interpreter session of 60 minutes costs $0.03. Shared filesystem storage costs $0.16 per GiB per month.

How to estimate your monthly cost

Start with the model you plan to use. Find its input and output rates on the pricing page. Estimate your monthly tokens. A typical chat app might send 10 million input tokens and 2 million output tokens per month. On gpt-oss-120B that costs $1.50 for input and $1.20 for output. The total is $2.70 per month. On Kimi K3 the same traffic costs $30.00 for input and $30.00 for output. The total is $60.00 per month.

Model choice dominates the bill. A small model can cost twenty times less than a large one for the same traffic. Test the small model first. Move up only when quality demands it. Cached input also cuts cost. Repeated system prompts and conversation history can hit the cache rate.

Watch for model removals. Together AI removes serverless models on short notice. A model you depend on can leave the list. Keep a fallback model ready. Dedicated endpoints keep a model running after its serverless listing ends.

Set a spend alert if your plan offers one. Track usage weekly while a project ramps up. A bug that loops API calls can burn a budget in hours. A rate limit or a usage cap catches it early.

Bottom line

Together AI pricing is pure pay-per-use. Serverless chat starts near free on small models and reaches $3.00 input and $15.00 output per million tokens on Kimi K3. Dedicated H100 endpoints cost $5.49 per hour. GPU cluster H100s cost $3.99 per hour on demand. Fine-tuning starts at $0.34 per million tokens on small models. Pick the smallest model that meets your quality bar. Check the pricing page before you commit, because rates change often.