Summary: Together AI is a prepaid inference and training platform: you buy credits, then spend them on serverless tokens, dedicated GPUs, fine-tuning, and related services. There is no free trial as of 2026-11-06. Official docs set a $15 first credit purchase (then $5 top-ups); a support KB article still says a flat $5 minimum — both are quoted below. This post maps the credit rules, product lanes, sample serverless rates, and the limits that matter when you compare Together with other inference APIs.
What Together AI sells (and how billing works)
Together AI positions itself as an open-source-model cloud: serverless chat and multimodal inference, provisioned throughput, dedicated endpoints, GPU clusters, sandboxes, post-training, and fine-tuning. The public pricing page at together.ai/pricing lists per-model serverless rates (per 1M tokens for chat), per-GPU-hour rates for dedicated inference and clusters, and per-1M-token fine-tuning prices with minimum charges.
Billing itself is documented under Credits. Credits are the prepaid unit for API requests, dedicated inference, fine-tuning, evaluations, and other Together services. The platform is fully prepaid for non-contract accounts: you need a positive balance before calls succeed, and API access is suspended when the balance hits zero. Card purchases usually land in moments; ACH (U.S. bank) purchases clear in about 1–3 business days and can go up to $50,000 per transaction ($100,000 for Scale/Enterprise contracts per the payment-methods docs).
Official screenshots of the pricing page, credits docs, homepage, and free-trial KB are attached at the end of this post.
No free trial — and the $15 vs $5 purchase conflict
Both the live Credits docs and the Knowledge Base agree on one point: Together AI does not currently offer free trials.
- docs.together.ai Credits (authoritative product docs, verified 2026-11-06): “Access to the Together platform requires a minimum $15 credit purchase.” Under Buy credits, the same page clarifies: minimum is $15 for your organization’s first purchase and $5 after that. It also notes that free credits (if any appear on a balance) are not enough for platform access — purchased credits at that minimum are required.
- support.together.ai KB “Does Together AI offer a free trial?” (last updated ~5 months before this check): “Access to the Together platform requires a minimum $5 credit purchase.” It does not split first vs subsequent purchases.
For budgeting, treat the live Credits docs as the rule you should plan against ($15 in, then $5 top-ups). Mention the KB’s flat $5 wording so you are not surprised if an older FAQ or UI copy still says $5 only — re-check Billing settings before you buy.
Prepaid credits do not currently expire per the Credits docs. Auto-recharge can top up when the balance falls below a threshold you set, but only when a credit/debit card is the default payment method. Setting ACH as default turns auto-recharge off automatically.
Product lanes compared
Use this table as a map of Together’s own product lines — not a competitor bake-off. Numbers below are from together.ai/pricing and the dedicated-endpoints pricing docs on 2026-11-06.
| Capability | Serverless Inference | Dedicated Endpoints | Fine-tuning |
|---|---|---|---|
| How you pay | Per token (and modality-specific units for image/audio/video). Credits debit as you call the API. | Per ready GPU replica, billed by the minute while serving (not per token). On-demand H100 listed at $5.49/GPU-hour; B200 at $8.99/GPU-hour; reserved via sales. | Per 1M tokens in the training (× epochs) and optional eval sets, with a per-job minimum charge that depends on the model (examples on the pricing page start around $4 and go up to $60+ for large models). |
| Entry / balance gate | Positive prepaid balance required; docs: first purchase $15, later $5. No free trial. | Same credit balance; you need enough balance to cover creating/running the deployment. | Same; enough balance to cover the job (including the minimum). |
| Rate / capacity limits | Docs: most users hit no hard rate limit; under load you may see 429 or 503. Best-effort TPS/TTFT. Batch API: up to 50% lower cost for async work. US-only endpoint us-api.together.ai: 1.25× standard token rates for listed models. |
No hard per-token rate limit like serverless; cost scales with ready replicas. GPU quota can still return 429 if you over-request replicas. | Job-based; billed for tokens processed, subject to the model’s minimum charge. |
| Best fit | Prototyping, variable or bursty traffic, many open models without owning GPUs. | Steady high throughput, custom models, predictable latency; stop or scale-to-zero when idle to cut spend. | LoRA or full fine-tunes (SFT / DPO listed on pricing) when you need a production-tuned open model. |
| Sample chat rates (per 1M tokens) | Examples from the Chat table: GLM-5.3-Flash $0.15 / $0.50 in/out; gpt-oss-120B $0.15 / $0.60; Llama 3 8B Instruct Lite $0.14 / $0.14; Kimi K3 $3.00 / $15.00 (cached input lower). Always re-check the live table — model lists move. | n/a (hardware-hour, not token). | Examples from pricing: Llama 3.1 8B SFT $0.34 / DPO $0.84 (min $4); Llama 3.3 70B SFT $2.03 / DPO $5.08 (min $4); larger MoE/frontier rows list higher rates and higher mins. |
Provisioned Throughput (PTUs) sits between serverless and dedicated: you reserve capacity in throughput units with an estimator on the pricing page. GPU Clusters list separate on-demand and reserved hourly rates (for example H100 on-demand $3.99/GPU-hour on the clusters table — distinct from the dedicated-inference H100 on-demand $5.49). Do not mix those two tables when you estimate a bill.
Serverless sample rates and what “$0 model” still means
The Chat tab on the pricing page shows Input / Output (and often cached input) per 1M tokens. A few verified rows from 2026-11-06:
- GLM-5.3-Flash — $0.15 input ($0.03 cached) / $0.50 output
- MiniMax M3.1 Flash — $0.15 input ($0.015 cached) / $0.60 output
- Qwen3.5 9B — $0.17 / $0.25
- gpt-oss-120B — $0.15 / $0.60
- Llama 3 8B Instruct Lite — $0.14 / $0.14
- Kimi K3 — $3.00 input ($0.30 cached) / $15.00 output
- Ternary Bonsai 27B — $0.00 / $0.00 on the table; Tev1 4B Experimental lists $0.04 input and Free output
Even when a row shows $0 or Free output, the Credits docs still require a purchased credit balance to use the platform. A $0 listed rate is not a free-trial account. Image, audio, video, and transcription tabs use different units (per image, per video, per audio minute, per 1M characters) — use the matching tab, not the chat token table.
If you need U.S.-only processing for supported models, point the client at https://us-api.together.ai/v1 and budget 1.25× the standard serverless rates. For non-interactive volume, the Batch API can cut cost by up to 50% versus real-time serverless on supported workloads.
Spend controls, cut-offs, and common API errors
From Spend controls and Error codes:
- Hard cut-off at $0 — prepaid accounts lose API access at zero balance; access resumes after you add credits. Metering has a short delay, so a little usage can still arrive after $0 — you remain responsible for that usage.
- 402 Payment Required — balance at the spend cut-off, or the model requires a positive balance. Add credits; if you already paid, wait up to about 5 minutes for the balance to update.
- 401 — missing/invalid API key. 403 — model or feature blocked (message names the reason). 429 — throttle or reduce bursts on serverless, or lower GPU replica requests on dedicated. 503 / 529 — back off and retry.
- Organization is the billing boundary: projects share one credit balance. Auto-recharge off/on affects the whole org. Per-project spend budgets are documented as coming soon — not available to configure yet.
For another prepaid / free-tier inference comparison on this blog, see Groq API Free Tier: Rate Limits, Costs and When to Upgrade and OpenRouter Free Models: Limits, Fees and When to Pay. If credit and auth errors look like provider routing issues, Hugging Face Inference Errors Fixed: 401, Credits, Providers covers a related pattern.
Editor's note
Who this is for: teams comparing open-model inference APIs who can fund a small prepaid balance and want one vendor spanning serverless, dedicated GPUs, and fine-tuning. Especially useful if you already prototype on OpenAI-compatible clients (Together documents OpenAI-compatible endpoints).
Who this is not for: anyone expecting a no-card free tier or unlimited $0 playground. Together’s own docs say there is no free trial right now. Also a poor fit if you need hard per-project spend caps today — those budgets are still “coming soon.”
Gotchas and fixes from official docs:
- Calls fail immediately after signup. Buy the documented first credit pack ($15 per Credits docs), wait for the balance to show, then retry. Do not re-submit a pending card purchase.
- Balance looks positive but API still blocked. Credits docs: check whether the balance is free credits vs purchased credits — platform access needs the minimum purchased pack.
- 402 after a top-up. Wait up to ~5 minutes for balance propagation (error-codes docs), confirm the default payment method, then retry.
- Auto-recharge never fires. Default payment method must be a card, not ACH. Switching default to ACH disables auto-recharge.
- 429 / 503 under load. Spread bursts, exponential backoff; for guaranteed capacity use provisioned throughput or dedicated endpoints.
- Dedicated bill higher than expected. You pay for ready replicas by the minute. Scale to zero / stop idle deployments; cold-start and non-ready time is not billed per DMI pricing docs.
FAQ
Q: Does Together AI have a free tier or free trial?
A: No. As of 2026-11-06 both the Credits docs and the support KB state that Together does not currently offer free trials. You need a prepaid credit purchase before API use. Some model rows may show $0 token prices, but that is not the same as a free account.
Q: Is the minimum purchase $5 or $15?
A: The live Credits documentation says $15 for the organization’s first purchase and $5 afterward. The Knowledge Base free-trial article still says a flat $5 minimum. Prefer the Credits docs when you budget, and confirm the amount shown in Billing → Add credits before you pay.
Q: Do unused Together credits expire?
A: The Credits docs say prepaid balance credits do not currently have an expiration date. Credits bought after an invoice is generated cannot clear past-due invoice balances — those must be paid separately.
Start with a small purchased balance, pick a cheap Flash-class serverless model for smoke tests, turn on auto-recharge only after you trust your traffic shape, and move to dedicated or provisioned capacity when 429s or latency SLAs force the issue. Re-check together.ai/pricing and billing credits before you lock a production budget — model rows and purchase floors change.
Last verified: 2026-11-06
Sources: together.ai/pricing · together.ai · docs — Credits · docs — Spend controls · docs — Payment methods · docs — Rate limits · docs — Serverless overview · docs — Geo-specific inference · docs — Dedicated pricing · docs — Error codes · KB — free trial / billing
Images: official Together AI public pages (pricing, Credits docs, homepage, free-trial Knowledge Base), screenshots captured 2026-11-06 for this post.
댓글
댓글 쓰기