How long a bulk LLM job takes β and what it costs β with parallel workers.
A bulk jobβs runtime is total tokens Γ· combined throughput, so more workers finish faster β until the providerβs TPM/RPM ceiling bites. This tool shows time and cost together. Check the ceiling on the rate limit calculator and the discount on the batch API savings calculator.
The Batch Processing Time & Cost Calculator estimates two things for a bulk LLM job: the wall-clock time it takes to finish and the total cost of the run. It multiplies your item count by the tokens generated per item to get the total token volume, then applies your price per 1M tokens to produce the dollar figure. Runtime comes from dividing that token work across your workers: the per-worker speed in tokens per second and the number of parallel workers together set throughput, so higher concurrency shortens the elapsed time even though the token bill stays the same.
The key trade-off is that concurrency cuts time but not cost. Doubling workers roughly halves the wall-clock duration, yet you still pay for every token generated. Before scaling up, check whether the tokens-per-item estimate is realistic, since output length is the single biggest driver of both time and spend. Trim verbose outputs first, then raise worker count to hit your deadline.
Total tokens = items Γ tokens per item. Divide by combined throughput (speed per worker Γ number of parallel workers) to get wall-clock seconds. Ten workers at 60 tok/s deliver 600 tok/s, so an 8M-token job takes about 3.7 hours β before rate limits throttle you.
If the job is not time-sensitive, yes. Batch endpoints (OpenAI, Anthropic, Gemini) run jobs asynchronously within a 24-hour window at roughly half the per-token price. You trade immediate latency for a large discount β ideal for embeddings backfills, evaluations and offline enrichment.