Simple, transparent pricing

Every plan gives you monthly credits in dollars. You spend credits on any model at the listed per-1K rate. For example, $59 in Starter credits buys 118,000 Fast records, 19,666 Balanced records, 7,375 Diverse records, or 2,360 Deep records. Mix and match however you like. Credits reset each billing cycle (they do not roll over).

conflict_v1 generates structured tension scenarios where two valid objectives compete (e.g., safety vs. autonomy, honesty vs. helpfulness). Each record includes responses from both drives, a resolution with confidence score, and an explicit override condition. It is designed for alignment researchers, safety teams, and anyone training models to reason about tradeoffs rather than apply rigid rules. Priced at preference_v1 rates.

Both preference_v1 and conflict_v1 require high-quality paired outputs. Balanced (GPT-4.1 Mini), Diverse (GPT-5.4 Mini), and Deep (Claude Sonnet 4.6) all support these schemas with reliable quality validation. The Fast tier does not support preference_v1 or conflict_v1 due to insufficient pair quality.

One record is one complete data item in your chosen schema. For instruction_v1, that's one instruction-response pair. For preference_v1, it's one prompt with chosen and rejected responses. For conflict_v1, it's one multi-drive tension scenario with resolution metadata. For eval_v1, it's one evaluation example.

Yes. All four models (Fast, Balanced, Diverse, Deep) are available on every plan, including Free. Your subscription gives you monthly credits, and you spend them on whichever model fits your needs.

Deep uses Claude Sonnet 4.6 for the highest possible per-record depth (1,955 chars avg output). At $25/1K records, it is the most expensive model but produces the most detailed, nuanced responses. Like every model, it is available on all plans. Just spend your credits on it.

If you have auto-recharge enabled and a payment method on file, your card is charged for the difference automatically. Otherwise, the job is blocked until you have enough credits. Upgrade your plan for more monthly credits, or enable auto-recharge in account settings.

Diversity % measures how unique the generated records are relative to each other, using trigram analysis. Higher diversity means less repetition and more varied training data. Diverse tier achieves 98%, Balanced 94%, Deep 82%, and Fast 80%. Choose the tier that matches your quality needs.

Every generation job supports custom system prompts, category-based distribution, response policies, hard negatives for adversarial training, train/validation output splitting, and custom metadata fields. All advanced options are included at no extra charge with any plan or pay-as-you-go tier.

No. You can use the simple model shorthand: 'fast', 'balanced', 'diverse', or 'deep'. We map these to the best model for each tier automatically. Advanced users can still specify exact provider/model combinations.