Pricing

One price per job. No subscription.

From ~$0.15 per million training tokens on production-size Llama 3.1 8B LoRA — measurably cheaper than Fireworks ($0.50) and Together ($0.48) on equivalent workloads, with no per-token middleman and full-precision QLoRA at 70B available on the same rate card.

You pay one all-in price per tune — the dashboard quotes it in dollars and in $/M training tokens before you commit. $5 minimum, no upfront bond — save a card on file once, then we charge it off-session when each job completes. Per-job cap starts at $500 for new accounts and unlocks to $5,000 as you build payment history.

Estimate your job

Approximate. The dashboard runs a more precise estimate against the actual dataset.

Estimated price
$5
All-in. Final number is quoted before submit.
Estimated duration
~24 min
Includes cold-start + average spot preemption overhead.

This job hits our $5 minimum— at this size you’re paying the floor, not the per-token rate. Without the floor it would be $0.15/M across ~15.7M tokens. Increase dataset size or epochs to see effective pricing.

Storage included: ~2.0 GB for 30 days. Extend later at $0.06/GB-month plus a $0.50 handling fee, up to 24 months. $5 minimum applies on small jobs.

Typical jobs

All-in price including managed-GPU compute, storage (30-day retention), and orchestration. Final number is quoted before submit.

Use caseYou pay
Quick experiment$5
Domain assistant~$10
Multilingual chatbot~$15
Mid-size production~$45
Large model fine-tune~$160
Long-context heavy job~$500

Prices scale roughly with model parameters × dataset size × epochs. You always see the exact number before you commit.

How we compare

Ballpark all-in price for a 70B QLoRA tune on a 100 MB dataset, 3 epochs (~75M training tokens). Competitor numbers are estimated from each provider’s public rate for an equivalent workload as of May 2026. Fireworks and Together tier their per-token rates by model size — the $0.50/M and $0.48/M rates quoted elsewhere on this site apply to models ≤16B; 70B falls into a higher tier.

ProviderEstimated total
Yachay~$160
Fireworks AI~$225
Together AI~$240
Replicate~$200–400

On 70B managed tuning we land ~29% below Fireworks and ~33% below Together at their 70B-tier rates, and up to ~60% below Replicate’s top per-second H100 pricing. The advantage grows with model size — we right-size the GPU per model, run compute on cloud Spot capacity with auto-resume on preemption, and skip the per-token middleman markup. On smaller (≤16B) jobs the gap compresses but our $/M still beats the same-tier ~$0.50/M (see the FAQ for the 8B comparison).

$5 minimum

The smallest job still settles at $5 — covers Stripe fees and a few minutes of GPU amortization. Lowered from $10 (May 2026) so trial-size LoRA jobs price honestly against industry $/M-token rates.

Card on file, no upfront charge

Save a card once via Stripe. We charge it off-session for the realized cost when each job completes — never before. No bond, no Stripe redirect on every submit.

Tier-based per-job cap

New accounts cap at $500 per job, auto-promoting to $2,000 after 3 clean jobs and $5,000 at $10K lifetime spend. Need more? Email hello@condorbox.ai for Enterprise ($25K cap).

What’s included in the price

  • · Cloud GPU time for the tune job, right-sized to the model (L4 → A100 → H100)
  • · Dataset upload + adapter storage (30 days, signed-URL access)
  • · Per-job orchestration, monitoring, and honest cancel billing — mid-run cancels charge only the realized compute, not a flat fee
  • · Billing through your existing Condor umbrella Stripe customer

Storage extensions

Default retention is 30 days after completion. Extend from the job’s dashboard page at $0.06/GB-month plus a $0.50 handling fee per extension, with a $1 minimum and 24 monthsmaximum. The exact charge is shown on the Extend button before you commit. Adapters auto-delete when retention lapses; we don’t archive to cheaper tiers — keep what you need, let the rest expire.

Refunds

Infrastructure failures on our side never produce a charge — the success-finalize path is the only place an invoice fires, and we bail out before it when the trainer fails or no adapter lands. Since billing is charge-at-completion against your card on file, “no invoice” literally means “no charge.” The dashboard shows the failure reason on the job page. If a charge did clear and the job is subsequently marked failed for an infrastructure reason, the refund auto-issues with the Stripe refund ID shown in the dashboard; if it doesn’t appear on your card within 10 business days, email hello@condorbox.ai with your job ID. Dataset-side failures that passed the linter but produced no usable adapter are reviewed case by case at the same address.

Start a tune

Billed through your existing Condor umbrella account. No separate signup.