Together AI Pricing in 2026: Serverless, Dedicated Endpoints and Fine-Tuning
What Together AI charges per 1M tokens for popular open models, what dedicated endpoints, GPU clusters and fine-tuning cost, and where Together is or isn't the cheapest provider.
Together AI serverless prices for popular open models start at $0.15 per 1M input tokens (gpt-oss-120B, GLM-5.3-Flash and Qwen3.8 Flash), and frontier open models such as Kimi K3 cost $3.00 / $15.00 per 1M input/output tokens. There is no free trial: you need to buy at least $15 of credits to start. The cheapest ways in are the $0.15-input models, or the free Ternary Bonsai 27B. For hardware, H100 GPU clusters cost $3.99 per GPU-hour on demand ($1.99 preemptible), and dedicated inference endpoints start at $5.49 per GPU-hour.
All prices below come from Together AI’s official pricing page and docs, plus our Together AI provider page, as of November 2026.
How does Together AI pricing work?
- Serverless inference is billed per 1M tokens, with separate input, cached-input and output rates. Image models are billed per image and video models per video.
- Prepaid credits. Together is fully prepaid. The first purchase must be at least $15, and later top-ups can be as small as $5. API access stops when your balance reaches zero.
- No free trial. Together’s billing docs say it doesn’t currently offer free trials.
- Batch API. Selected serverless models run at 50% off in batch mode, with a fixed 24-hour window. Other models run at standard rates.
- Dedicated endpoints and GPU clusters are billed per GPU-hour. Fine-tuning is billed per 1M training tokens.
Together AI serverless prices (per 1M tokens)
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT OSS 120B | $0.15 | — | $0.60 |
| GLM-5.3-Flash | $0.15 | $0.03 | $0.50 |
| Qwen3.8 Flash | $0.15 | — | $0.47 |
| DeepSeek V4.1 Flash | $0.30 | $0.006 | $1.20 |
| Qwen3.7 Plus | $0.32 | — | $1.28 |
| Muse Glimmer 30B | $0.35 | $0.04 | $1.50 |
| Qwen3.6 Plus | $0.50 | — | $3.00 |
| DeepSeek V4 Pro | $1.32 | $0.13 | $3.96 |
| GLM-5.2 | $1.40 | $0.26 | $4.40 |
| GLM-5.3 | $1.40 | $0.26 | $4.40 |
| Qwen3.7 Max | $2.50 | $0.50 | $7.50 |
| Kimi K3 | $3.00 | $0.30 | $15.00 |
Together also lists Kimi K3.1 Pro ($2.00 / $10.00), MiniMax M3.1 Flash ($0.15 / $0.60) and Inkling 2 Small ($0.30 / $1.00). Cached input costs between 2% and 20% of the normal input rate (DeepSeek V4.1 Flash cache hits cost $0.006). That matters for agents and chat apps that resend long prompts.
Together’s catalog changes often. Llama 3.3 70B, for example, is now listed only for fine-tuning, not serverless inference. There are currently no serverless embedding or rerank models; those run only on dedicated endpoints.
Images: FLUX.2 [dev] costs $0.0154 per image, FLUX.2 [pro] $0.03, GPT Image 2 $0.053, and Qwen Image $0.0058. Compare these on the cheapest image generation APIs ranking.
Is Together AI the cheapest?
Usually not. For most models, Together charges the model maker’s own list price, and specialist hosts undercut it.
| Model | Together AI | Model maker’s API | Cheaper elsewhere |
|---|---|---|---|
| DeepSeek V4 Pro | $1.32 / $3.96 | DeepSeek: same at peak, $0.66 / $1.98 off-peak | DeepInfra $1.30 / $2.60 |
| DeepSeek V4.1 Flash | $0.30 / $1.20 | DeepSeek: same at peak, $0.15 / $0.60 off-peak | DeepInfra $0.14 / $0.42 (discounted) |
| GLM-5.3 | $1.40 / $4.40 | Z.ai: same | DeepInfra $0.5625 / $2.50 (discounted) |
| GLM-5.2 | $1.40 / $4.40 | Z.ai: same | DeepInfra $0.5625 / $1.80 (discounted) |
| Kimi K3 | $3.00 / $15.00 | — | DeepInfra $2.85 / $14.25 |
| Qwen3.8 Flash | $0.15 / $0.47 | Alibaba: same | DeepInfra $0.113 / $0.382 |
| Qwen3.7 Max | $2.50 / $7.50 | Alibaba: same | No cheaper verified price |
| Qwen3.7 Plus | $0.32 / $1.28 | Alibaba: $0.40 / $1.60 | Together is cheaper than Alibaba |
Two things stand out:
- DeepSeek direct is half price most of the time. DeepSeek’s peak window covers 01:00–04:00 and 06:00–10:00 UTC on weekdays. At every other time, including weekends, DeepSeek charges half of Together’s rate.
- Qwen3.7 Plus is a rare win for Together: 20% below Alibaba’s international price, and Alibaba’s price is itself marked as a limited-time discount.
Together’s case rests on breadth, not price: one account covers LLMs, image and video models, fine-tuning, dedicated endpoints and GPU clusters. Each model page above lists every provider we track. The cheapest LLM APIs ranking covers the rest.
How much do Together AI dedicated endpoints cost?
Dedicated inference gives you single-tenant GPUs with no sharing, custom model support and autoscaling. The pricing page lists on-demand rates per GPU-hour, and Together’s docs say dedicated endpoints bill per minute while they run.
| Hardware | On-demand per GPU-hour |
|---|---|
| NVIDIA HGX H100 | $5.49 |
| NVIDIA HGX H200 | $5.99 |
| NVIDIA HGX B200 | $9.29 |
| NVIDIA HGX B300 | $11.49 |
| NVIDIA GB300 NVL72 | $15.79 |
Reserved pricing, and GB200 NVL72, go through sales. Charges continue until you stop or delete the endpoint.
How much do Together AI GPU clusters cost?
GPU clusters are raw capacity for training or self-managed serving, priced per GPU-hour:
| GPU | Preemptible | On-demand | Reserved 3–30 days | Reserved 31–90 days |
|---|---|---|---|---|
| H100 | $1.99 | $3.99 | $3.69 | $3.45 |
| H200 | $2.99 | $5.99 | $4.99 | $4.15 |
| B200 | $4.65 | $9.29 | $8.69 | $7.99 |
| B300 | $5.75 | $11.49 | $10.59 | $9.89 |
Reservations of 91+ days, plus GB200 and GB300 NVL72 racks, require a capacity request.
How this compares depends on the chip. Together’s $3.99 on-demand H100 is below RunPod’s $4.39 for an H100 SXM in its Secure Cloud. On H200 and B200, though, RunPod is cheaper: $5.29 vs $5.99 and $7.99 vs $9.29. The chip pages linked in the table compare every GPU provider we track; Together’s own rates are on its GPU listing.
How much does Together AI fine-tuning cost?
Fine-tuning is billed per 1M tokens processed: training tokens (dataset size × epochs) plus evaluation tokens. Together offers both LoRA and full fine-tuning, each with separate rates for supervised fine-tuning (SFT) and Direct Preference Optimization (DPO). Every job has a minimum charge.
| Model | SFT per 1M tokens | DPO per 1M tokens | Minimum per job |
|---|---|---|---|
| Qwen3.5 0.8B–9B | from $0.34 | from $0.84 | $4 |
| Qwen3.5 27B | from $1.05 | from $2.62 | $4 |
| Llama 3.3 70B | from $2.03 | from $5.08 | $4 |
| DeepSeek-V3.1 | $7.00 | $17.50 | $20 |
| Kimi K2.6 | $15.00 | $37.50 | $60 |
| GLM-5.2 | $40.00 | $100.00 | $60 |
The pricing page lists two rate tables, and some models appear in both at slightly different prices (Llama 3.3 70B SFT is $2.03 or $2.24). Use the cost estimate in the console, CLI or API before launching a job to see the exact figure. Cancelled jobs are charged only for completed steps, and failed jobs are fully refunded. Serving a fine-tuned model on a dedicated endpoint is billed separately at the GPU rates above.
Is Together AI worth it?
Good fit if you want one provider for serverless open models, fine-tuning and GPU capacity. It also suits workloads that lean on cached input, or Qwen3.7 Plus, where Together beats Alibaba’s own price.
Look elsewhere if price per token is what matters most. DeepInfra undercuts Together on most of the models above, and DeepSeek direct is half price off-peak. A routing library can send traffic to whichever provider is cheapest for each model. For the wider landscape, see our serverless AI inference guide.