Together AI Review
Serverless and dedicated GPU inference for open-source models, plus fine-tuning
Is this your product? Claim this page · Request a change
Looking for a Together AI alternative? See our ranked comparison.→What is Together AI?
Together AI is an inference and training platform for open-source models. You call serverless endpoints for chat, vision, image, video, and audio models by the token or by the unit of output, or you rent dedicated GPUs (H100, H200, B200) by the hour to host your own model or run a fine-tune. It also runs a fine-tuning pipeline (SFT, LoRA, DPO) billed by training tokens, with a separate hourly bill once the resulting model is deployed.
The pitch is breadth and price: hundreds of open models on one API, at rates generally below the closed-model labs, with the option to drop down to raw GPU rental when you need more control than a managed endpoint gives you.
Together AI screenshots
Who it's for
- ✓ Teams already committed to open-weight models (Llama, DeepSeek, Qwen, GLM, Kimi) who want one API instead of standing up their own serving stack
- ✓ Teams that want to fine-tune an open model and then serve it without moving to a second vendor
Who should look elsewhere
- ✗ Teams that want a published, fixed rate limit before they build against it
- ✗ Teams evaluating on a strict $0 budget, since there's no free trial and no free tier
Pros
- + Every price is on the public pricing page, model by model, with no tiers or sales calls required
- + Covers serverless inference, dedicated GPUs, fine-tuning, and image/video/audio generation under one account and one bill
- + Reserved GPU cluster pricing drops noticeably below on-demand (H100 cluster on-demand $3.99/hr vs. reserved down to $3.29/hr for 91-180 days)
Cons
- – No free trial. The docs require a minimum $5 credit purchase before the API will respond to any call
- – Serverless rate limits are dynamic and unpublished; the docs tell you to read response headers instead of a fixed number, which makes capacity planning harder
- – A dedicated endpoint or fine-tuned model keeps billing by the GPU-hour even while idle, so low-traffic deployments can cost far more per month than the fine-tuning job itself
Together AI pricing
What you pay for
Together AI has no subscription plans. You pay per million tokens for serverless model calls, per GPU-hour if you rent dedicated capacity or a cluster, and per token plus hosting time if you fine-tune a model. Every rate is published on the pricing page down to the individual model, so nothing here is quote-only. The catch is there's no free trial: you need at least $5 in purchased credit on the card before the API will answer a request.
You pay for what you consume rather than a per-seat fee, so cost scales with usage.
| Plan | Price | Highlights |
|---|---|---|
| Pricing is quote-only. Contact the vendor for a quote. | ||
No plan tiers. You pay per million tokens on serverless inference (from $0.03/$0.12 input/output on small models up to $1.74/$3.48 on frontier models like DeepSeek V4 Pro, with some models running higher still, e.g. $4.50 output on Kimi K2.6), per GPU-hour on dedicated single-tenant endpoints ($5.49/hr on-demand for H100, $8.99/hr for B200, H200 is quote-only, reserved rates are quote-only too), per GPU-hour on rentable GPU clusters ($3.99/hr H100, $5.99/hr H200, $8.19/hr B200 on-demand, dropping to $3.09-$7.99/hr depending on GPU and commitment length from 7 to 181+ days), per image or per video on generation models, per minute on transcription, and per million tokens plus hourly hosting on fine-tuning. There is no free trial: the docs require a minimum $5 credit purchase before you can call the API, and your balance must stay positive or access is suspended. The $0.03 startingPrice above is the cheapest serverless model's input rate, not a plan price.
Pricing verified July 7, 2026 · source
How Together AI's pricing compares
Together AI next to its closest alternatives on entry price, billing, and whether pricing is public.
| Tool | Starting price | Billing | Free option | Pricing disclosed |
|---|---|---|---|---|
| Together AI | Usage-based | usage-based | No | Public |
| Baseten | Custom / quote | usage-based | No | Not disclosed |
| Modal | $0.0000131/core-second | usage-based | Yes | Partly public |
| Replicate | $0.000025/second | usage-based | No | Public |
| Fireworks AI | Usage-based | usage-based | Trial ($1 in free credits for new accounts to test serverless inference) | Partly public |
| RunPod | Usage-based | usage-based | No | Public |
Is Together AI still actively developed?
Last significant update: July 2026. Announced an $800M Series C, with investors including NVIDIA, Aramco Ventures, and Vista Equity, plus a commitment of over 500 MW of compute capacity for future GPU buildout.
Top Together AI alternatives
Together AI FAQ
Does Together AI have a free tier or free trial?+
No. The docs state Together AI does not currently offer free trials, and platform access requires a minimum $5 credit purchase before you can make API calls.
How is Together AI priced?+
Pay-as-you-go with no subscription tiers. Serverless model calls are billed per million tokens (rates vary by model, from about $0.03 to over $4 per million), GPU clusters are billed per hour on-demand from $3.99/hr for an H100 up to $8.19/hr for a B200 (dedicated single-tenant endpoints run higher, $5.49-$8.99/hr on-demand), and fine-tuning is billed per million training tokens plus a separate hourly hosting cost once deployed.
What are Together AI's rate limits?+
There are no fixed, published per-model limits. Limits are dynamic and scale with the model's live capacity and your recent usage, so you're expected to check the x-ratelimit-reset response header rather than plan against a fixed number.