How Together AI pricing works and how to avoid surprise bills

How prepaid credits and balance cut-offs work

Together AI does not currently offer free trials. Accessing the platform requires an initial purchase of at least $15. After that first transaction, subsequent top-ups have a lower minimum purchase requirement of $5. You buy credits by heading to your billing settings, selecting "Add credits" in the Credit balance card, entering the amount, and confirming the payment.

Credits bought with a debit or credit card usually show up in your balance within moments. If you choose Automated Clearing House (ACH) instead, the deposit waits until the transaction clears, which typically takes 1 to 3 business days. Avoid re-submitting payments while they're pending, or you might end up funding your account twice.

Prepaid balance credits don't currently expire, but there's a strict rule on how they apply to outstanding debts: credits purchased after an invoice is generated cannot be used to clear previous invoices or past-due balances.

For standard accounts, the $0 balance cut-off acts as the platform's primary backstop:

• Prepaid accounts: Once your balance reaches zero, API calls are rejected until you add new funds.
• Metering delay: Because metering has a short delay, some usage can arrive after your balance hits zero. You remain responsible for paying that overage, so the $0 cut-off doesn't guarantee an exact spending cap.
• Contracted accounts: Customers with an active Scale or Enterprise contract don't get cut off at zero; they continue to be billed under their contract terms. Qualifying monthly enterprise customers can configure an in-arrears cut-off at an agreed negative balance to pause access for the rest of the billing period.

If you're comparing provider spend limits across the market, check our breakdown of which cloud and AI providers let you set a hard spending cap.

Serverless inference pricing

Serverless inference is billed strictly on consumption. Text models charge per 1 million input and output tokens, image models charge per image or per megapixel (mp), and audio speech recognition models charge per audio minute.

Pricing ranges widely depending on model size and complexity:

For image models like FLUX.2 [max], prices cover the default step count, but additional costs apply if you configure your request to exceed those default steps. If you want to compare these token rates to multi-provider routing layers, see what OpenRouter charges on top of model prices or look at DeepInfra pricing explained for alternative host rates.

Dedicated endpoints, GPU clusters, and fine-tuning costs

Running dedicated resources shifts your billing from per-token charges to clock-time infrastructure costs. You must have a sufficient credit balance to cover the cost of creating a dedicated endpoint, fine-tuning job, or evaluation job before you launch it.

Dedicated inference and GPU clusters

Dedicated single-tenant GPU instances are billed per GPU per hour on demand:

• NVIDIA HGX H100: $5.49 per GPU per hour
• NVIDIA HGX B200: $8.99 per GPU per hour

If you run GPU Clusters, on-demand rates run at $3.99 for NVIDIA HGX H100, $5.99 for H200, $8.19 for B200, and $9.99 for B300 per GPU per hour. If your workloads tolerate interruptions, preemptible compute lowers those hourly rates to $1.99 for H100, $2.99 for H200, $4.09 for B200, and $4.99 for B300. Reserved terms across 3 to 180 days offer deeper discounts; a 91 to 180-day reservation drops an H100 down to $3.19 per hour, an H200 to $3.99, and a B200 to $6.79.

Code sandbox and storage

For agentic workloads executing generated code:

• Code Sandbox compute costs $0.0446 per vCPU per hour and $0.0149 per GiB RAM per hour.
• Code Interpreter runs at $0.03 per 60-minute session.
• Shared Filesystem storage is billed at $0.16 per GiB/month.

Post-training and fine-tuning

Post-training infrastructure is billed per minute for each active model resource. An NVIDIA HGX B200 costs $11.00 per GPU per hour, discounted to $9.00 per GPU per hour under a promotion valid until 12/14/2026.

Fine-tuning charges per 1M tokens based on the sum of training dataset tokens (training dataset size multiplied by number of epochs) plus any evaluation dataset tokens (validation dataset size multiplied by number of evaluations). Every fine-tuning job enforces a model-specific minimum charge. Models like Llama 3.1 8B, Llama 3.3 70B, and Qwen3.5 0.8B have a $4.00 minimum charge per job, while Kimi K2.6, Kimi K2.7 Code, GLM-5.1, and GLM-5.2 have a $60.00 minimum charge per job.

The auto-recharge trap and missing budget caps

Auto-recharge is convenient, but it's dangerous if you treat it like a spend control. Auto-recharge keeps purchasing credits as you consume them; it does not cap your spend.

To set up auto-recharge, open billing settings, locate the "Auto-recharge" panel, click "Set thresholds" (or "Edit limits"), and set two values:

1. When balance goes below: The dollar amount that triggers a purchase.
2. Bring credit balance up to: The target balance to fund.

Keep in mind that if you configure your threshold above your current balance, auto-recharge triggers immediately with a single purchase to bridge the gap. Auto-recharge is also only supported when your default payment method is a credit or debit card. If you switch your default payment method to ACH, auto-recharge turns off automatically, even if you keep a card on file.

Here is the catch for small teams: organization and project budgets and spend alerts are planned features, but they aren't currently available to configure. Until budgets launch, balance cut-offs are the only hard stops on the platform.

Furthermore, projects within an organization share the organization's single credit balance. A separate project doesn't get its own prepaid balance or spending cap, and turning off auto-recharge affects funding for every project across the board, including production. If a teammate leaves an inference loop running in a staging project, it drains the balance shared by your customer-facing apps.

Tracking usage via analytics and the beta billing API

To see your current billing period spend in the UI, go to your billing settings and click "See current monthly invoice usage" on the Monthly spend card.

For deeper analysis, organization admins can use the Cost analytics dashboard (external collaborators cannot view project cost analytics). You can view the chart by "Cost ($)" or "Units", and group data by Product, Line items, Project (beta), or API key (beta). Time ranges include Yesterday, Last 7 Days, or the default Last 30 Days, as well as custom ranges up to 31 days in UTC. You can export the breakdown table via "Download CSV", which outputs Date, Line Item, Dimensions, Quantity, Unit Price (USD), and Amount (USD).

If you prefer programmatic monitoring, Together AI provides a beta billing usage endpoint:

GET /v1/billing/usage

A few important caveats about this API:

• Off by default: The API is in beta and requires you to contact support to enable access for your organization.
• Broad scope: Any project's API key can read the usage data, and it returns numbers for the key's entire organization.
• Parameters: Passing organization_id is optional during beta, but will be required at general availability.
• Latency: Data for the current month can be up to 1 hour behind, and prior-month data can trail by up to 24 hours (all timestamps in UTC).
• Line items: Responses include product, quantity, unit price, and cost. Inference line items include project_id and api_key_id.

While cost analytics attributes usage to individual API keys, the keys themselves do not support independent spending limits.

Monitoring Together AI with Costbrake

Because Together AI lacks native budget caps, alerts, and per-key spending limits, you have to track usage externally to catch unexpected usage before your card recharges multiple times.

Costbrake monitors spend across cloud and AI providers, alerting you to sudden spikes, budget thresholds, month-end forecast overruns, and new charges. For Together AI, Costbrake reads billed dollars per hour, product, and API key via the GET /v1/billing/usage endpoint. Checks run hourly on the Free plan and every 15 minutes on Pro and Team. Usage is counted in dollars before credits are applied, so API consumption covered by prepaid credits still triggers alerts properly.

Costbrake triggers:

• Spike alerts: An hour costing more than 3x the normal 7-day baseline and at least $5 above average.
• Budget alerts: Threshold warnings at 50%, 80%, and 100% of your configured budget.
• Forecast alerts: Triggers when month-to-date spend plus the pace of the last 72 hours projects that you'll cross your budget by month-end.
• New-charge alerts: Triggers if a service appears for the first time this month at over $1.

Alerts route to email on all plans, with Slack, Discord, and signed webhooks available on Pro ($12/month), and PagerDuty on Team ($39/month).

Costbrake has no brakes for Together AI: it alerts, and you act in Together AI's dashboard. Together AI's billing usage API is in beta and off by default, so ask Together support to enable it for your organization first. Together AI has no read-only keys, so the key you supply can also call models; if it ever leaks, revoke it under your API key settings.

You can set up alerts via our agent setup by prompting your coding agent, or manage connections across OpenAI, Anthropic, and Together AI by reading our guide on how to monitor spend across AI APIs in one place.