What it is
Together AI is a cloud built for open-weight AI models. It covers the whole range: a shared API for models like Kimi K3, DeepSeek V4, GLM-5.3, Qwen and gpt-oss, a batch API, fine-tuning jobs, dedicated endpoints that run a model on GPUs reserved for you, and raw GPU clusters for training.
Prices on the shared API are per million tokens; for example gpt-oss-120B is $0.15 input and $0.60 output, DeepSeek V4 Flash 0731 is $0.14 and $0.28, and the newer V4.1 Flash is $0.30 and $1.20. Batch jobs run at up to 50% less. Dedicated H100 endpoints are $5.49 an hour. Rented H100 clusters cost $3.99 an hour on demand, $1.99 if you accept that the job can be interrupted, and $3.19 with a 91 to 180 day reservation.
The platform is fully prepaid. There is no free trial, the first credit purchase must be at least $15, and API access stops when the balance reaches zero.
Our review
Together's advantage is range, and the pricing rewards staying with it as a workload grows. A team can prototype on the shared API, fine-tune gpt-oss-120B for $2.50 per million training tokens, then serve it on a dedicated endpoint, all on one prepaid balance. Few rivals cover training clusters as well as inference.
Shared API prices are in line with the market, and the catalogue moves fast enough to need watching. Its October 2026 pricing page lists gpt-oss-120B at $0.15 input and $0.60 output per million tokens, but DeepSeek V4.1 Flash costs more than twice the older V4 Flash 0731 on input ($0.30 against $0.14). A team pinned to a model name will see its bill change when it upgrades. Cached input is heavily discounted on the larger models, and the batch API is billed at up to 50% less.
Dedicated hardware is where Together clearly undercuts. A dedicated H100 endpoint is $5.49 an hour against $8.00 at Fireworks AI, and a B200 is $8.99 against $13.00. Training clusters are cheaper still at $3.99 an hour for an H100 on demand and $1.99 preemptible, though H200 endpoints and the GB200 and GB300 systems are sales-only.
Operationally it has become simpler. The old Build Tier system and its spend-based rate limits are gone; the changelog says most serverless users will no longer hit limits, though 429 errors still occur under heavy demand. The trade is a prepaid model with no trial and a hard stop at zero balance, so production accounts need auto-recharge on a card. Prompt storage is a privacy setting rather than off by default, and Zero Data Retention has to be switched on, after which inputs and outputs are not stored or used for training.
Where it earns its place
A two-engineer start-up building a document chatbot
Serve an open model behind an OpenAI-style API without running GPUs
Point an OpenAI SDK at Together and call gpt-oss-120B. 750 million input and 250 million output tokens a month cost about $262 at list prices, less with cached input or batch.
An ML engineer fine-tuning a support model on 2 billion tokens
Train a LoRA on Llama 3.1 8B and serve it
Supervised fine-tuning on Llama 3.1 8B is $0.34 per million tokens, so about $680 for the run. Serving it on one dedicated H100 around the clock is about $4,000 a month.
A research lab finance lead
Budget a three-month pre-training run on 64 H100s
At the 91 to 180 day rate of $3.19 per GPU-hour, 64 H100s for 90 days come to about $441,000, against about $552,000 on demand. Preemptible capacity at $1.99 is cheaper but can be interrupted.
Before you commit
- Setup
- Create an account, buy at least $15 of credit, generate an API key and call it with the OpenAI-compatible SDK; no free trial.
- Integrations
- OpenAI-compatible chat, embeddings and batch endpoints; Together SDKs; fine-tuning and GPU clusters in the same console.
- Data and privacy
- Not used for training without opt-in; Zero Data Retention is a setting you must enable; SOC 2 Type 2 and ISO 27001:2022, with reports via trust.together.ai.
- Cost at scale
- One dedicated H100 endpoint running all month is about $4,000 ($5.49 x 730 hours); an 8-GPU H100 cluster is about $23,300 a month on demand or $18,600 on a 91 to 180 day reservation.
- Limits
- Serverless has no fixed per-tier rate limits since October 2026 but can return 429s under load; batch allows up to 30 billion enqueued tokens per model.
What's good
- Inference, fine-tuning and GPU clusters on one prepaid account
- Dedicated H100 endpoints at $5.49/hour, below Fireworks' $8.00
- Batch API at up to 50% off
- SOC 2 Type 2 and ISO 27001:2022
What's not
- No free trial; $15 minimum first purchase
- API access stops at zero balance unless auto-recharge is on
- Zero Data Retention must be enabled manually
- H200, GB200 and GB300 capacity needs a sales call
Best for
Teams building on open models who want to move from API calls to fine-tuning and their own GPUs with one provider.
Alternatives
Together AI: common questions
- Is Together AI worth it?
- A single bill for open models at every stage, from a $15 API top-up to a reserved H100 cluster, at prices that undercut the specialist hosts on dedicated hardware. We rate it 4.5 out of 5.
- How much does Together AI cost?
- Prepaid, $15 minimum first purchase, no free trial. Per 1M tokens, e.g. gpt-oss-120B $0.15 in / $0.60 out, Kimi K3 $3 / $15; batch up to 50% off. Dedicated endpoints: H100 $5.49/hour, B200 $8.99/hour. GPU clusters: H100 $3.99/hour on demand, $1.99 preemptible. Fine-tuning from $0.34 per 1M tokens.
- What are the downsides of Together AI?
- No free trial; $15 minimum first purchase; aPI access stops at zero balance unless auto-recharge is on; zero Data Retention must be enabled manually; h200, GB200 and GB300 capacity needs a sales call.
- Who is Together AI best for?
- Teams building on open models who want to move from API calls to fine-tuning and their own GPUs with one provider.
- How much does Together AI cost for a team or in production?
- One dedicated H100 endpoint running all month is about $4,000 ($5.49 x 730 hours); an 8-GPU H100 cluster is about $23,300 a month on demand or $18,600 on a 91 to 180 day reservation.
- Is Together AI safe for company data?
- Not used for training without opt-in; Zero Data Retention is a setting you must enable; SOC 2 Type 2 and ISO 27001:2022, with reports via trust.together.ai.
- What are the best Together AI alternatives?
- The closest alternatives we have reviewed are Fireworks AI, Groq, Modal and OpenRouter.
Answers checked Oct 22, 2026.
Make Together AI? Show this rating on your site
