Prime Intellect.
Build AI for a particular job.
Connect your app to hosted AI models, test them on tasks, or train a model for a job you repeat. This is a developer platform. It needs code and a spending plan; it is not a ready-made assistant for your inbox or errands.
Getting started: TechnicalConnect an API key and write code for model calls. Training also needs Python tasks, a way to judge answers, suitable model/GPU choices and confirmed beta access. Our setup guide, not a performance score.
Good for
Developers making a model work well on a specific task. Prime’s Ramp customer story describes training FastAsk to find answers inside Ramp Sheets workbooks. This is an attributed training workflow, not our test or proof of every gateway model’s performance.
Start with
A Prime account, a suitable API key and developer skills. Training beta must be enabled for your account or team, with enough spendable balance and an allowed GPU count. Access to one component does not establish access to every model or training job.
Cost
Usage-based. Current Prime-hosted GLM 5.3 lists $1.40 per million input tokens and $4.40 per million output tokens. Training is billed by GPU time at the rate shown before dispatch. Sandboxes and stored images can add separate costs. No general hosted free quota, signup grant or fixed monthly subscription is confirmed.
Plans and limits explainedModel calls, training and running costs.
- Free access not confirmed
- No general ongoing free hosted quota is confirmed. Public documentation and open libraries do not include free model, compute or storage use.
- Trial not confirmed
- A standard free trial or signup grant was not established. Conditional training credit grants can exist; amount, eligibility and expiry are not promised to every account. Signup card requirements are unconfirmed.
- Pay-as-you-go
- Current Prime-hosted GLM 5.3: $1.40 input and $4.40 output per million tokens. Training bills per GPU-hour at the rate confirmed before dispatch; beta access and a spendable balance are needed. No fixed monthly or annual platform bill is confirmed.
- Extra costs & limits
- Sandboxes, images and checkpoints have separate running/storage charges. Sandbox launch rates expire after 22 December 2026. Team API billing needs the documented header. Tax can apply. Terms permit recurring orders; refill defaults, credit expiry, minimum purchase and a cash cap are unconfirmed.
Prices in USD unless noted. Check the billing period; local prices, taxes and offers can differ.
Plan details
These are separate usage meters, not one unlimited subscription. Prices are in USD. Read your actual account offer before paying.
- Free access and trialNot confirmed
- No general free hosted quota or standard signup-credit amount is established by the reviewed current pages. Conditional training grants and an overdraft allowance can exist; they are not promised to every account. Free access to documentation or software is separate from hosted model, compute and storage bills.
- Model API exampleGLM 5.3 · $1.40 in / $4.40 out
- Per million tokens in the native catalog, checked 21 October 2026. Exact ID: z-ai/glm-5.3, Prime Hosted. Other models have their own rates, features and routes. This is API usage, not a ChatGPT or Claude app subscription and not an overall model recommendation.
- On-demand trainingPer GPU-hour · rate shown before dispatch
- Training beta must be enabled. The CLI shows available GPU rates, then confirms the selected rate before launch. Billing starts when the training workers are ready and stops when it finishes or pauses. It is prorated to the second with no minimum. Queueing, scheduling and image pulls are not billed. A multi-GPU run costs more than one GPU.
- Optional sandbox example$0.046 / running hour
- Example only: 1 vCPU, 2 GiB memory and 5 GiB disk. CPU is $0.02/vCPU-hour, memory $0.0125/GiB-hour and disk $0.0002/GiB-hour. These launch rates are valid through 22 December 2026. Images and checkpoints add $0.10/GiB/month storage, billed hourly, plus a one-day-storage upload fee.
- Payment, credits and limitsCheck the account and order
- The terms allow recurring charges for a recurring order until canceled; that does not prove automatic top-ups on every account. Free-credit expiry, signup card requirements, minimum top-up, automatic refill defaults and a total spending cap were not confirmed. Tax can apply. An API request normally bills the key owner; using a team balance needs the documented team header.
Source checks: 21 October 2026. Current GPU rates and eligibility are account-specific; we did not sign in or run the CLI. Do not reuse prices from the retired shared training service. Running resources, storage and model calls are separate costs.
Pricing checked 21 Oct 2026.
Try something small.
- Connect a small app to a model through its API
- Check a model against a few questions with known answers
- Train a specialist model after confirming access and compute costs
Getting started.
Start with the public model catalog and documentation. Check the model ID, price and where it runs. For an API test, create an eligible account and a key with Inference permission, set a small request limit and send harmless sample text. Training and sandbox access have separate steps and bills.
A training experiment can use several GPUs and many runs. Confirm the bill before starting. Prime-hosted models run on Prime infrastructure; gateway models go to another provider. Check the chosen route’s features and data rules. The older shared LoRA training service is retired. We have not opened an account, trained a model or made a paid request.
Classify three harmless notes.
Try this task
- You’ll need
- Eligible Prime Inference access, a configured client and a small request limit. Check the selected model’s price and account balance first.
- A request to copy
Classify each note as Question or Reminder. Return one row for each note. Do not use tools. Notes: 1. Where is the drawing club? 2. Bring a green pencil on Saturday. 3. What time does the club start?
- What to expect
- Three rows: Question, Reminder, Question. No outside tools or account actions should be needed.
- How to check it is done
- Check that all three notes are classified correctly, that no tool was used and that the usage charge matches your request. This checks a small model call, not a trained agent or a quality ranking.
Our suggested first task based on the API documentation. We have not run it. It uses the selected model’s paid usage or an eligible account grant; no free grant is promised.
Our suggested prompt, based on the linked source. We haven’t tested this task ourselves.
Source checked October 21, 2026.Checked 21 October 2026. Prices and access can change. Character art is our unofficial illustration.
The current linked terms name Prime Intellect Inc., registered in Delaware with a Dover registered-office address. This establishes the operator country, not its operational headquarters, product access or data residency.
