Start$0to start
No subscription. Add credits by card and pay per token.
Fine-tuningfrom $0.15per 1M training tokens
Download the adapter you train.
Batch inferencefrom $0.029per 1M input tokens
The same prices on your fine-tuned model.
Prices per model
| Model | Fine-tuning, per 1M training tokens | Batch input, per 1M tokens | Batch output, per 1M tokens | Batch on your fine-tune | Minimum |
|---|---|---|---|---|---|
qwen3.8-27b | — | $0.03 | $1.00 | — | $0.20 per batch |
nemotron-3.5-lightning-30b-a3b | — | $0.10 | $0.20 | — | $0.20 per batch |
gemma4-26b-a4b-it | — | $0.15 | $0.30 | — | $0.50 per batch |
qwen3.5-9b | $0.35 | $0.095 | $0.145 | Download only | $0.50 per fine-tune$0.20 per batch |
gpt-oss-20b | $0.45 | $0.029 | $0.12 | Download only | $0.50 per fine-tune$0.20 per batch |
qwen3-embedding-0.6b | — | $0.009 | — | — | $0.15 per batch |
qwen3-8b | $0.33 | $0.06 | $0.14 | Same as base | $0.50 per fine-tune$0.20 per batch |
qwen2.5-0.5b-instruct | $0.15 | $0.03 | $0.05 | Same as base | $0.50 per fine-tune$0.20 per batch |
qwen2.5-1.5b-instruct | $0.15 | $0.029 | $0.049 | Same as base | $0.50 per fine-tune$0.20 per batch |
qwen2.5-7b-instruct | $0.29 | $0.049 | $0.115 | Same as base | $0.50 per fine-tune$0.20 per batch |
llama-3.1-8b-instruct | $0.29 | $0.049 | $0.115 | Same as base | $0.50 per fine-tune$0.20 per batch |
mistral-7b-instruct-v0.3 | $0.29 | $0.049 | $0.115 | Same as base | $0.50 per fine-tune$0.20 per batch |
bge-m3 | — | $0.009 | — | — | $0.15 per batch |
Batches on your fine-tuned models use the base model's batch rates and minimum, with no surcharge.
What a job costs
A support-ticket classifier
| Step | Cost |
|---|---|
| Fine-tune qwen2.5-1.5b-instruct on 500 tickets of 400 tokens for 3 epochs: 0.6M training tokens at $0.15 per 1M, raised to the minimum per fine-tune. | $0.50 |
| Then classify 50,000 tickets with your fine-tuned model, 400 tokens in and 10 out each: 20M input tokens at $0.029 per 1M, and 0.5M output tokens at $0.049 per 1M. | $0.60 |
| Total | $1.10 |
Distil a large model into a small one
| Step | Cost |
|---|---|
| Run the teacher, qwen3.8-27b, on 500 prompts of 400 tokens, with answers of 250 tokens: 0.2M input tokens at $0.03 per 1M, and 0.125M output tokens at $1.00 per 1M, raised to the minimum per batch. | $0.20 |
| Fine-tune the student, qwen2.5-7b-instruct, on 450 of those prompts and answers, 650 tokens each, for 3 epochs (50 are held out to check it): 0.8775M training tokens at $0.29 per 1M, raised to the minimum per fine-tune. | $0.50 |
| Then run your fine-tuned student on the 50 held-out prompts: 0.02M input tokens at $0.049 per 1M, and 0.0125M output tokens at $0.115 per 1M, raised to the minimum per batch. | $0.20 |
| Total | $0.90 |
There is no separate distillation fee: a distillation is two batches and a fine-tune, each charged at the prices above, minimums included.
Estimate a job
Total$1.10
- Fine-tuning
- $0.50
- Batch
- $0.60
An estimate at the prices above, with each job's minimum. A job is charged for the tokens it actually uses.
How billing works
- One credit is worth 1 USD. Buy a credit pack, or enter any amount from $5 to $1,000 per purchase.
- Prices exclude taxes such as VAT or GST, which Creem adds at checkout where they apply.
- Every fine-tune is quoted when you create it, before it starts, and its charge is never more than the quote. Only the results a batch delivers are charged. A batch that fails through our fault is not charged.
- If a job fails because of Gridline or before training starts, you aren't charged. If you cancel a job after training starts, you pay only for the tokens trained so far, with no minimum and never more than your quote. If training finishes but your account already holds the most fine-tuned models allowed, the job fails and you pay for the tokens trained, with no minimum.
- Purchased credits don't expire. Rate changes apply only to jobs created after the change. A job keeps the rates it was quoted.
Questions
- Do I pay to host my fine-tuned model?
- No. There is no endpoint to keep running: you run batches on your fine-tuned model at its base model's prices.
- Can I download my model?
- Yes. When a fine-tune succeeds, you can download its adapter with your token.
- Why is there a minimum per job?
- Every job has a cost to start, whatever its size, and the minimum covers it: $0.50 per fine-tune, and from $0.20 per completed batch ($0.15 for embeddings).
- What if a job fails?
- If a job fails because of Gridline or before training starts, you aren't charged. A batch that fails through our fault is not charged.
- Do you offer real-time inference?
- Not today: Gridline runs fine-tuning and batch jobs. You can download your adapter and serve it yourself.
- Is there volume pricing?
- The prices on this page apply to every account. For large or regular workloads, get in touch. Talk to the founder.
Credits
- One credit is worth 1 USD.
- Credits are prepaid. They are non-transferable and can be spent only on Gridline.
- Purchased credits don't expire.
- Buy a credit pack, or enter any amount from $5 to $1,000 per purchase.
- Prices exclude taxes such as VAT or GST, which Creem adds at checkout where they apply.
- Prices below are in US dollars. Every charge is deducted from your credits.
Credit packs
A pack's price in US dollars equals its credits. Packs carry no bonus credits.
| Pack | Credits | Price |
|---|---|---|
| Pack 1 | 10 | $10 |
| Pack 2 | 25 | $25 |
| Pack 3 | 100 | $100 |
Fine-tuning
Charge = training tokens × epochs ÷ 1,000,000 × the base model's price per 1M training tokens, and at least its minimum per job.
- Training tokens are counted with the base model's tokenizer.
- Every job is quoted when you create it, before it starts. The quote shows the training tokens, the epochs and the most the job can cost.
- When the job succeeds, the charge is deducted from your credits. It is never more than the quote.
- If a job fails because of Gridline or before training starts, you aren't charged. If you cancel a job after training starts, you pay only for the tokens trained so far, with no minimum and never more than your quote. If training finishes but your account already holds the most fine-tuned models allowed, the job fails and you pay for the tokens trained, with no minimum.
- On the 7B and 8B bases, QLoRA runs on a 48 GB GPU at any rank up to 64 and is charged the LoRA rate, so it costs the same as bf16 LoRA. There, bf16 LoRA (the default) trains faster and fits more training tokens in a job; QLoRA is the default on the smaller bases.
Preference fine-tuning (DPO)
- A DPO job is charged the same way at its own price per 1M training tokens. Its training tokens are the prompt plus both answers, every epoch, and it pays at least the base model's minimum per job.
- A DPO job you cancel after training starts isn't charged.
| Base model | DPO price per 1M training tokens |
|---|---|
qwen2.5-7b-instruct | $0.725 |
llama-3.1-8b-instruct | $0.725 |
Batch inference
Charge = input tokens ÷ 1,000,000 × the price per 1M input tokens + output tokens ÷ 1,000,000 × the price per 1M output tokens, and at least $0.20 for a batch that completes ($0.50 on gemma4-26b-a4b-it, $0.15 for an embeddings batch).
- Embeddings are charged for their input tokens only, at the price per 1M embedding tokens.
- Only the results a batch delivers are charged. A batch that fails through our fault is not charged.
- Batches on your fine-tuned models use the base model's batch rates and minimum, with no surcharge.
Rate changes
Rate changes apply only to jobs created after the change. A job keeps the rates it was quoted.