Billing Units

Benched.ai Editorial Team

Billing units are the atomic quantities by which an AI service provider meters usage for pricing. Clear units let customers forecast cost and compare vendors.

Common Billing Schemes

Unit	Definition	Typical Service
Prompt tokens	Unicode tokens in request	Chat completion APIs
Generated tokens	Tokens in response	Chat completion APIs
Seconds of audio	Real-time audio processed	Speech-to-text
Images created	Single 512×512 output	Text-to-image
Jobs	One fine-tuning run	Managed training

Conversion Factors (2025 averages)

Unit	USD Cost	Notes
1k prompt tokens GPT-4o	$0.03	list price
1k generated tokens GPT-4o	$0.06	list price
60 s Whisper large	$0.006	0.1 $/min
1 SDXL Turbo image	$0.001	at volume tier

Design Trade-offs

Separate prompt vs completion billing rewards concise prompts but penalizes verbose outputs.
Bundled token pricing simplifies invoices but hides hotspot categories.
Per-job billing for training encourages experiments yet shifts risk of GPU overruns onto vendor.

Current Trends (2025)

Unified "compute credits" abstract over heterogeneous units, convertible at runtime.
Token adapters map BPE tokens to characters for cross-provider comparisons.
Regulators push for standardized spend dashboards after surprise-bill incidents.

Implementation Tips

Log raw billing units client-side; reconcile with vendor invoice to catch errors.
Budget alerts on projected month-end units prevent overrun.
Normalize tokens across encoding versions when comparing historical bills.