Together AI Pricing 2026: Serverless Tokens, Dedicated Endpoints and GPU Hours

Short answer: Together AI is prepaid and usage-based, with no subscription. Serverless models bill per million tokens, from $0.15 input and $0.60 output on gpt-oss-120B up to $3.00 and $15.00 on Kimi K3. A dedicated H100 endpoint costs $5.49 per GPU-hour, and raw H100s in a GPU cluster cost $3.99 an hour on demand. You need a $15 first credit purchase to start. Checked on Together's pricing page, 30 October 2026.

The pricing question most teams actually face is serverless vs dedicated: below roughly a billion tokens a month, per-token billing almost always wins, and above that the answer depends on which model you run. The scenario tables below show where the line falls.

Together AI pricing at a glance

Prices from Together AI pricing, dedicated inference pricing docs and batch docs, read on 30 October 2026.

Together has no free trial. The credits docs say the first purchase must be at least $15 and later top-ups at least $5, credits do not expire, and API access stops when the balance hits zero. That prepaid model is a real safety feature for a side project: you cannot wake up to a surprise bill.

How much do serverless models cost on Together AI?

Output prices vary about 100-fold between models, so the model you pick is the budget. The page even lists one model, Ternary Bonsai 27B, at $0.00. A selection from the pricing page:

If tokens are new to you, what a token is in AI explains the unit. Output tokens cost three to six times as much as input on most large models, so a chatbot that writes long answers costs far more than a classifier that returns one word. Cached input is the other lever: on DeepSeek V4.1 Flash a cached token costs $0.006 per million against $0.30 uncached, so long system prompts that repeat on every call get much cheaper.

One change that catches people out: Llama 3.3 70B is no longer on Together's serverless API. Its model page says so, and the pricing page lists no 70B Llama for serverless. You can still run it, but only on a dedicated endpoint billed by the GPU-hour.

What does a dedicated endpoint cost per month?

Dedicated model inference bills for hardware, not tokens. An H100 80GB costs $5.49 an hour and a B200 180GB costs $8.99 an hour, both on demand; H200, B300 and GB300 are quote-only. Larger models need more GPUs, and Together's docs say a four-GPU configuration costs four times the single-GPU rate.

All figures are our arithmetic: hourly rate × hours, per replica, before any reserved discount.

Billing runs by the minute, and only while a replica is ready to serve. Provisioning and cold starts are not billed, and a deployment scaled to zero costs nothing. That makes the second row realistic for internal tools: set an inactivity timeout and the endpoint stops itself overnight. Reserved capacity, at a lower effective rate for a committed term, is quote-only.

Dupple data: the same H100 costs 38% more per hour as a managed dedicated-inference replica ($5.49) than as an on-demand cluster GPU ($3.99), our arithmetic from Together's pricing page on 30 October 2026. The premium buys autoscaling, traffic routing and a ready inference API; a cluster gives you bare GPUs and you run the serving stack yourself.

Serverless or dedicated: which is cheaper at 50M and 500M tokens?

Here is the monthly serverless bill at two volumes, assuming three input tokens for every output token, a common shape for chat and retrieval apps. Compare each figure with $3,952.80 for one always-on H100 replica.

Our arithmetic: 75% of tokens at the input price, 25% at the output price, no caching, no batch discount.

The answer is lopsided. At 50M tokens a month, serverless is cheaper than a dedicated H100 for every model on the list, by a wide margin. At 500M, only Kimi K3 comes close to the cost of a replica, and only because its output price is high.

Dedicated still makes sense in three cases. The model you need is not on serverless, as with Llama 3.3 70B. You need guaranteed latency and no shared rate limits. Or your traffic is so steady that a replica stays busy all day, which Together's own docs name as the point where dedicated usually wins. Before you commit, check whether one H100 can actually carry the model: a model that needs two or four GPUs doubles or quadruples the replica cost.

How much does batch inference save?

Batch jobs run from one uploaded JSONL file and return results within a 24-hour window, at up to 50% off serverless rates, with a separate rate-limit pool. On the GLM-5.3 example above, that would take a 500M-token month from $1,075 to as little as $537.50 if the model qualifies for the full discount (our arithmetic).

The trade-off is latency. Batch suits classification, evaluations, synthetic data and overnight summaries, not anything a user is waiting on. Together keeps batch input and output files for 7 days, so download results before they expire.

What do GPU clusters cost?

GPU clusters rent raw GPUs by the hour, with longer commitments bringing the rate down:

Prices per GPU per hour from Together AI pricing, same date. Terms of 181 days or more, and GB200 and GB300 NVL72 systems, are quote-only.

Preemptible GPUs cost half the on-demand rate but can be taken back, so they fit fault-tolerant training with checkpoints, not serving. Shared filesystem storage alongside a cluster costs $0.16 per GiB a month. For how these rates compare with other GPU clouds, Devshot keeps a cheapest GPU cloud comparison.

What does fine-tuning cost?

Fine-tuning bills per million tokens processed: training set size × epochs, plus any evaluation tokens. LoRA supervised fine-tuning starts at $0.34 per 1M tokens on small models such as Llama 3.1 8B and Qwen3.5 9B, and Llama 3.3 70B costs $2.03 per 1M. Each job carries a minimum charge, $4 on most small models and up to $60 on the largest.

Worked example: a 20M-token dataset trained for 3 epochs on Llama 3.3 70B is 60M tokens, or $121.80 with LoRA SFT (our arithmetic). Serving the result then needs a dedicated endpoint, since the base model is not on serverless.

Together also sells post-training compute by the GPU-hour. B200 time is on a promotion at $9.00 an hour instead of $11.00 until 14 December 2026, so budget for the higher figure on any run after that date.

The costs that surprise new users

• No free tier. There is no trial; the $15 first purchase is the entry fee.
• Idle replicas bill. A dedicated endpoint with a minimum replica count above zero bills around the clock, traffic or not.
• Multi-GPU models multiply. A model that needs four H100s costs four times the single-GPU rate per replica.
• Models leave serverless. Llama 3.3 70B moved off the serverless API, so a model you build on today may need a dedicated endpoint later.
• Displayed image and video prices are minimums. The pricing page says they refer to the lowest resolution and duration settings.

If you are comparing providers rather than models, a router can send each request to the cheapest host for that model: see our list of LLM gateways. Running small models on your own machine costs nothing per token, and the best free local LLM tools cover that route. For the apps that sit on top, the best RAG tools and the best AI agent frameworks are the next stop, and OpenRouter pricing breaks down the main multi-provider alternative.

How we checked these prices

We read Together AI's pricing page, the Llama 3.3 70B model page, and the docs on dedicated inference pricing, batch inference and credits on 30 October 2026. Scenario volumes, the 3:1 input-to-output split and every monthly total are our own arithmetic from list prices. Together's model list changes often, so check the model you plan to use on the pricing page before you budget. A summary of the same plans sits on Together AI's Toolradar pricing profile.

FAQ

Bottom line

Start on serverless, pick the cheapest model that does the job, and use batch for anything that can wait; how to reduce LLM API costs lists the other levers. Move to a dedicated endpoint only when the model you need is not on serverless or your traffic keeps a replica busy most of the day, and run the per-month arithmetic on the exact GPU count first.

Cite this: Dupple, "Together AI Pricing 2026: Serverless Tokens, Dedicated Endpoints and GPU Hours", November 2026.

Pricing out the AI stack for your team? Start the Dupple X trial.