Open-model inference API with no free allowance. Pay per token, and new accounts need a deposit first.
Sniffed 17d ago Card required#llm#inference#open-source-models#gpu#fine-tuning#openai-compatible#api
There is no free credit here, so treat it as a paid API from day one. Together AI hosts open and partner models behind an OpenAI-compatible API, with serverless inference, batch pricing, dedicated GPU endpoints, GPU clusters, sandboxes and fine-tuning. The pricing page I read lists no free allowance, and G2 and Trustpilot reviewers from 2025 and 2026 say new accounts start read-only until you make a deposit. A few models are listed at $0.00 per million tokens, and some cheap ones such as GLM-5.3-Flash run $0.15 in and $0.50 out. G2 shows 4.3 from 91 reviews, mostly incentivized, praising latency and the easy switch from OpenAI. Trustpilot shows 2.7 from 8, mostly one-star reports about charges, cancellation and paying before testing. Reviewers also mention 429 errors on bursty traffic and thin documentation.
Free tier quotas
Free allowance
None listed on the pricing page
New accounts
A G2 reviewer in October 2026 said new accounts start read-only and need an initial deposit before the Playground works
Models listed at $0.00
Ternary Bonsai 27B is $0.00 in and out per million tokens; Tev1 4B Experimental has free output
Cheap serverless example
GLM-5.3-Flash at $0.15 in and $0.50 out per million tokens
Another budget example
gpt-oss-120B at $0.15 in and $0.60 out; DeepSeek V4 Flash 0731 at $0.14 in and $0.28 out
Cached input
Cheaper cached-token rates on many chat models, for example $0.03 for GLM-5.3-Flash
Batch API
A separate batch price view is shown for a subset of models
Code Interpreter
$0.03 per 60-minute session
Fine-tuning minimum
Each job has a minimum charge, for example $4.00 on small models
Last sniffed Oct 8, 2026. Always confirm on the official pricing page.
👍 Good boy
+G2 rates it 4.3 from 91 reviews, with 59% giving 5 stars and 38% giving 4 stars. All of the most recent reviews on that page carry an incentivized (G2 invite) label.
+A G2 reviewer building voice agents said time to first token and tokens per second were excellent and that new models arrive quickly, while noting Cerebras was still faster for their use.
+Several G2 reviewers said the OpenAI-compatible API made switching easy. One said an app already using OpenAI needs only a few code changes, and an enterprise architect said internal teams plugged in within a day.
+A G2 reviewer said running open models without managing GPUs removed a lot of painful work, and that fine-tuning an open model on your own data and then serving it on the same platform worked well.
+The pricing page covers a wide range: cheap models such as GLM-5.3-Flash at $0.15 in and $0.50 out, large ones such as Kimi K3 at $3.00 in and $15.00 out, plus image, video, audio and embedding models in one account.
+Capacity options are priced in public. Dedicated H100 endpoints are $5.49 per GPU hour, on-demand H100 clusters $3.99 and preemptible H100 $1.99, and provisioned throughput is $0.05 per PTU minute.
+G2's AI summary lists ease of use and accessible open-source models at 2 mentions each among the top pros, with free services, ease of implementation and integrations at 1 each.
👎 Watch out
−There is no free inference. A G2 reviewer in October 2026 wrote that there is no free tier for inference and that usage-based pricing makes budgets hard to predict. Another said new accounts are read-only until a deposit, and wanted a small free credit instead.
−Trustpilot shows 2.7 from 8 reviews, and 6 of the 8 are critical. They describe a $160 monthly charge that could not be cancelled (July 2026), paying before being able to test, and a $1 card authorization that was not refunded (October 2024). The profile is unclaimed and the sample is small.
−A G2 reviewer who runs voice agents said request-per-minute limits cause 429 errors during traffic bursts, and wanted higher limits even for lower-tier users.
−Documentation draws repeated criticism. G2's summary lists poor documentation, payment issues and required technical expertise as cons, and one reviewer said it is definitely not for beginners.
−A G2 reviewer said on-demand GPU clusters can occasionally be tight during peak hours and that support response times can be hit or miss if you hit a billing snag.
−An enterprise G2 reviewer said provisioning private deployments on custom clusters can take longer than expected and the console progress bar does not always match the real status.
−A Trustpilot reviewer in November 2025 said the website shows two serverless model lists, a long one and a short one, and that a model they paid for was not available on serverless. I could not verify that claim.
The catch
Together AI does not list a free allowance. The pricing page I read shows per-token serverless prices, a batch price view, provisioned throughput, dedicated inference, GPU clusters, sandboxes, storage, post-training and fine-tuning, and no free credits.
Two things are free in a narrow sense. Ternary Bonsai 27B is listed at $0.00 in and out per million tokens, and Tev1 4B Experimental has free output with a $0.04 input price. Reviewers say you still need a funded account. A G2 reviewer in October 2026 wrote that new accounts start in read-only mode and require an initial deposit before the Playground works, and another wrote that there is no free tier for inference. I set the card requirement to required for that reason, since reviewers describe paying before testing, but I did not see the sign-up flow myself.
This entry is in the directory because the product is a staple for open-model inference, not because it has a free tier. Expect to pay from the first call. Representative prices per million tokens: GLM-5.3-Flash $0.15 in and $0.50 out, gpt-oss-120B $0.15 in and $0.60 out, Kimi K3 $3.00 in and $15.00 out. Fine-tuning jobs carry a minimum charge, and the page also shows a promotion on B200 post-training that runs until December 14, 2026.
Trustpilot shows 2.7 from 8, with several one-star reports about charges and cancellation. The sample is small, and the profile is unclaimed.
What’s included
✓OpenAI-compatible serverless chat API
✓Batch API pricing for a subset of models
✓Image, video, audio, transcription, embeddings, rerank and moderation models
✓Provisioned throughput billed in PTUs
✓Dedicated inference on H100 and B200 GPUs
✓On-demand and reserved GPU clusters
✓Code sandbox and code interpreter sessions
✓LoRA and full fine-tuning
✓Post-training on dedicated GPUs
✓Shared filesystem storage at $0.16 per GiB a month
Free tier value
Generosity
B
Estimated value
No free allowance; cheapest paid chat models start near $0.14 per million input tokens
Feature parity
Serverless, batch, dedicated endpoints and fine-tuning are priced separately.
Screenshots
HomepagePricing page
Paid plans
Serverless inference, pay per token
Price per million tokens varies by model, with cached-input discounts on many. Examples: GLM-5.3-Flash $0.15 in and $0.50 out, Llama 3.3 70B $1.04 in and out, Kimi K3 $3.00 in and $15.00 out.
Provisioned throughput
Reserved capacity in PTUs at $0.05 per PTU minute for MiniMax M3, Kimi K3 and GLM-5.2. Tokens per minute per PTU depends on the model.
Dedicated inference
Single-tenant GPUs per hour on demand: H100 $5.49 and B200 $8.99. H200 and newer parts are by quote.
GPU clusters
On-demand H100 $3.99, H200 $5.99, B200 $8.19 and B300 $9.99 per GPU hour. Preemptible H100 is $1.99. Reserved terms lower the rate.
Fine-tuning
Priced per million training tokens with a minimum charge per job. Qwen3.5 9B supervised fine-tuning is $0.34 per million tokens with a $4.00 minimum.
Common questions
Does Together AI have a free tier?
The pricing page lists none. A few models are priced at $0.00, and reviewers say new accounts need a deposit before the Playground works.
Is the API compatible with OpenAI's?
G2 reviewers say yes and describe switching with an endpoint change and a few code edits.
Why do I get 429 errors?
A G2 reviewer reported request-per-minute limits being hit during traffic bursts. A Trustpilot review from October 2024 said Tier 1 promised 600 requests per minute but the account stayed at 10.
What do reviewers say about support and billing?
Opinions split. G2 reviewers praise onboarding support, while Trustpilot reviewers describe charges and cancellation problems. Trustpilot has only 8 reviews.
Can I fine-tune models?
Yes. The pricing page lists LoRA and full fine-tuning per million training tokens with a minimum charge per job.
Is it good for beginners?
Reviewers say not really. G2's summary lists poor documentation and required technical expertise as cons.
Change history
Checked October 22, 2026
Pricing page lists no free allowance. G2 shows 4.3 from 91 reviews and Trustpilot 2.7 from 8.
