Compare LLM Prices per Accepted Job
For a practical LLM price comparison, divide each option's full monthly cost by the number of jobs that pass the same quality and latency checks. Compare OpenAI and Anthropic API quotes, a GPU you operate, and any managed agent platform at both typical and peak volume.
Price the same workload on each API
Create a fixed set of representative jobs and define what qualifies as an accepted result before running it. For document extraction, require the correct fields, a valid output format, and delivery within your deadline. Record input and output tokens separately for every attempt, including retries. Record cached usage and tool charges separately wherever they apply.
The OpenAI pricing excerpt lists gpt-6.2-sol Standard at $2 per million input tokens and $10 per million output tokens. For Anthropic's Claude API, verify the exact model, input and output rates, and any batch or cache terms in the provider's API pricing table. Run the same jobs on both APIs, apply their respective rate cards, and divide each total by its accepted jobs. A cheaper run that misses your pass threshold fails the comparison.
Define a customer deadline before collecting acceptance results. A late but correct answer can have no operational value when a human has already resolved the case.
Pricing | OpenAI APIMeasure the GPU hours you would pay for
For self-hosting, Google Compute Engine lists GPUs including L4 and A100. Verify the regional rate of a VM configuration that fits your chosen model, including attached resources, storage, and data transfer. Load a model you can serve and replay your representative jobs at expected concurrency. Record billed instance hours, including idle time, plus the accepted job count, tail latency, and operator time for maintenance and incident recovery.
In an illustrative quote, a VM at $0.60 per hour running 300 billed hours costs $180. Add $20 for other infrastructure and two hours of operator time at $60 per hour for a $320 monthly total. At 20,000 accepted jobs that is $0.016 per job. Higher utilization may lower that ratio, while extra replicas for peak demand may raise it.
Use billed amounts from the same trial population. Include discarded attempts and retries in the cost numerator even when they earn no accepted delivery.
Cloud GPUs (Graphics Processing Units) | Google CloudCalculate when a deadline changes the bill
The OpenAI excerpt lists gpt-6.2-sol Batch at $1 per million input tokens and $5 per million output tokens. In an illustrative month with 20,000 jobs, assume 3,000 input tokens and 500 output tokens per job, no caching, retries, or tools, and a passing result every time. That gives 60 million input tokens and 10 million output tokens. The listed Standard rates yield $220 and the listed Batch rates yield $110.
That $110 difference buys you a reason to test mixed routing. Put deadline-tolerant jobs through Batch and run deadline-sensitive jobs through Standard. Measure elapsed time from submission until a validated result is available. Include fallback attempts when a batch result misses the deadline. Compare cost per accepted job again after counting those attempts.
Pricing | OpenAI API