Pricing
A price for every layer of the stack.
Platform plans for teams, rate cards for training, datasets, and compute. Rent what we’ve built, bring your own, or commission something new — our open research stays free for everyone.
- Currency
- All prices in USD
- Billing
- Monthly or annual · cancel anytime
- Startups
- Up to $25k in credits
Platform plans
Start on your own. Bring the team when it’s time.
Researcher is free for individuals. Team bills per active seat, with a 14-day trial and no card up front. Enterprise is an annual contract scoped with us.
Team bills per active seat and prorates changes; annual billing takes two months off. Open research needs no plan at all.
Plan details
The spec sheet, not the sales pitch.
Model training services
Training, priced like an engineering line item.
Two axes set the price of every run: the compute it uses, metered per GPU-hour, and the parameters you need, billed per billion. Fine-tune, shrink a model you already have, or scope a custom program.
A run bills as compute used at the Compute GPU rates plus the per-params service fee; custom programs are quoted after a scoping call. The weights are yours either way. Dataset and training engagements start at a $40,000 minimum contract.
Model datasets
Rent ours, bring your own, or build your own.
High-quality training data three ways, all priced per billion params of the model you train: license the corpora our pipeline keeps refreshing, run your own data through it, or commission a corpus built for your domain.
Dataset pricing scales with model size — a 7B model pays 7× the per-B rate, a 70B model 70×. Rented corpora carry a non-exclusive term license; anything we process or build for you is yours, including the generation harness. Dataset and training engagements start at a $40,000 minimum contract.
Compute
The clusters we train on, by the hour.
The same GPU fleet behind our research, priced per GPU-hour with no instance minimums. Reserve for a year and the rate drops about 30%.
Estimates assume a 30-day month and steady daily usage. Reserved rates require a 12-month commitment; spot capacity is available on request.
Research
The research is free. The researchers are not.
Everything we publish stays open, whatever your plan. Rent the lab’s time when you want it pointed at your problem.
Sponsorship buys attention, not conclusions — sponsored findings publish openly whether or not they flatter the sponsor.
Billing questions
The fine print, answered.
Can I try Team before paying?
Yes — every Team workspace starts with a 14-day free trial, no card required. When it ends you can downgrade to Researcher without losing your data.
How does per-seat billing work?
You’re billed monthly for active seats. Add or remove people any time; changes prorate automatically on your next invoice.
Do you offer startup pricing?
The Startup Program includes up to $25k in product credits over 12 months for eligible early-stage teams — often more valuable than a discount.
What happens to my open-research access if I cancel?
Nothing. Datasets, protocol specs, and published research are open to everyone regardless of plan — that’s the point of publishing them.
Who owns a model you train for me?
You do — weights, checkpoints, and eval results. Training engagements are work for hire; we keep the methods, you keep the model.
How is compute billed?
By the GPU-hour, with no instance minimums — on-demand meters exactly the hours you use. Reserved capacity bills monthly against a 12-month commitment at roughly 30% below on-demand.
Not sure which plan fits?
Tell us about your workload and we’ll tell you honestly — including when the free tier is all you need.