Batch vs Flex vs Standard vs Priority — same model, four prices.
Latency is the lever: async Batch work should never pay standard price, and only truly interactive paths need Priority. Sizing a single model? Rank every model on the 300+ model comparison, and check async savings on the batch API savings calculator.
The LLM Service Tier Pricing Calculator estimates your monthly cost of running a large language model workload across the four service tiers most providers offer: Batch, Flex, Standard, and Priority. You enter your expected input tokens per month and output tokens per month (in millions), along with the standard input and standard output prices per 1M tokens. From those figures it computes the total spend at the standard rate and then applies each tier's typical multiplier, so the two biggest drivers are your monthly token volume and the ratio of output tokens to input tokens — since output pricing is usually higher, output-heavy workloads cost noticeably more.
The key trade-off is price versus latency and reliability. Cheaper tiers like Batch and Flex trade lower per-token cost for slower or non-guaranteed processing, while Standard and especially Priority cost more but return results faster and more predictably. The practical tip is to match the tier to the job: route large, non-urgent jobs — overnight summarization, bulk classification, evaluations — to Batch or Flex to capture the discount, and reserve Standard or Priority for user-facing, real-time requests. Because savings scale with volume, run your real monthly token numbers through the calculator before committing, as a small per-token difference compounds into a meaningful figure at scale.
Most big providers sell the same model at several latency tiers. Batch runs your requests asynchronously (often within 24h) for about half price; a Flex or service tier is cheaper but slower or variable; Standard is the default; a Priority tier costs a premium for guaranteed fast responses.
Whenever the work is not user-facing in real time — bulk summarisation, embeddings, evals, data labelling, overnight jobs. Batch is typically ~50% off, so moving async workloads off the standard tier can roughly halve that part of the bill for zero quality loss.