Together AI

What Together AI actually meters

Together meters three distinct things and never mixes them into one bill line. On serverless it's the token, split into input, cached input and output, and cached input carries its own published rate on most models rather than a blanket multiplier. MiniMax M3 charges $0.30 input against $0.06 cached. GLM-5.2 charges $1.40 against $0.26, an 81% discount, while Qwen3.5-397B-A17B charges $0.60 against $0.35 and saves only 42%. The discount isn't uniform, so caching pays back very differently model to model.

On dedicated model inference the meter switches to hardware. Together bills per GPU-hour, measured by the minute, per replica, and only while a replica is ready to serve. Provisioning, cold starts and DEGRADED replicas don't bill. Token volume is irrelevant here: the model affects cost only through the GPU count it needs.

Fine-tuning meters tokens processed, defined as (n_epochs × training tokens) + (n_evals × validation tokens). Disable packing and it recalculates as dataset length multiplied by max_seq_length, which moves the number a long way. Cancelled jobs pay for completed steps only, and failed jobs get fully refunded.

How credits work

Together is fully prepaid, and credits are the only currency. You buy a balance, minimum $5, and every service draws from it: API calls, dedicated deployments, fine-tuning and evaluation jobs. There's no free trial and no signup grant. Credits don't expire, and Together commits to advance notice if that changes, which is more than most prepaid vendors say.

Auto-recharge tops the balance back to a target when it falls below a threshold you set, as a single transaction on your default payment method. It works only when that default is a card: setting a US bank account as default turns auto-recharge off automatically. One sharp edge, credits bought after an invoice is generated can't clear that invoice or any past due balance.

What happens when you hit the limit

The balance hitting zero is the limit, and Together suspends API access until you add credits. That's a hard stop rather than an overage charge, which is what prepaid should mean and frequently doesn't. Creating a dedicated endpoint or a fine-tuning job needs enough balance up front to cover the cost.

Rate limits behave differently. They're dynamic, applied per model rather than per account, and grow with sustained reliable traffic. The old Build Tier 1 to 5, Scale and Enterprise labels are retired, so there's no tier to buy into. Every serverless response returns headers carrying the current limit.