CalcSays
BULK CLASSIFICATION · BATCH SAVINGS

Bulk Classification Batch Savings

A bulk classification job is ~95% waitable (tune below) — so the 50% batch discount lands as roughly 48% off the whole bill. Here's the exact number and which model saves most.

For engineers weighing the batch API — models the real saving after only your waitable traffic can take the 24h discount, not the brochure's flat 50%.

Model prices from OpenRouter · updated 2026-07-13

01 Your setup

Model

$2/M in · $10/M out · Anthropic batch 50% off (24h)

02 Real-time vs batch

Batching saves $998/month — 48% overall, not the 50% brochure figure.

All real-time
$2,100
standard prices, every request
With batch
$1,103
95% waitable at −50%

50% discount × 95% waitable = 48% overall

Push the waitable share up (queue overnight jobs, bulk work) and the saving scales toward the full 50%.

03 Where the savings come from

Real-time baselineevery request at standard price$2,100
Waitable share batched (95%)50% off the 95% that tolerates 24h−$998
Blended monthly$1,103
Tweet
Why not the full 50%? The discount is per-request and only for a 24-hour turnaround, so it applies to the waitable slice of your traffic, not all of it. Overall saving = waitable share × discount. The way to save more is to move more work into the waitable bucket — real-time chat and agent steps can never batch.

04 Biggest batch savers at this workload

1. o1-proOpenAI$71,250/mo (48%)
2. GPT-5.4 ProOpenAI$15,675/mo (47%)
3. GPT-5.5 ProOpenAI$15,675/mo (47%)
4. Claude Opus 4.7 (Fast)Anthropic$14,963/mo (48%)
5. GPT-4OpenAI$12,825/mo (48%)
6. GPT-5.2 ProOpenAI$11,970/mo (48%)
7. o3 ProOpenAI$9,500/mo (48%)
8. GPT-5 ProOpenAI$8,550/mo (48%)
9. Claude Opus 4Anthropic$7,481/mo (48%)
10. Claude Opus 4.1Anthropic$7,481/mo (48%)

Top 10 of 315 models by absolute batch saving at this workload and waitable share. Only OpenAI, Anthropic and Google publish batch tiers; the rest show no saving.

📄 Get the full agent cost audit
A multi-page report on your exact setup: where cost spirals, how much context-trimming / state-resets / model-routing each saves, and which model+framework combo is cheapest at your real failure rate. Generated from your numbers.

Related cost calculators

Batch API Savings Calculator →Claude Sonnet 5 batch savings →GPT-5.5 batch savings →LLM API Cost Calculator →Prompt Caching Savings Calculator →

Standard prices from OpenRouter, snapshot 2026-07-13, synced daily. Batch discounts hand-verified against provider docs (2026-07-13): OpenAI Batch API · Anthropic Batch API · Google Batch API. Blended = requests × per-request × [waitable × (1−discount) + (1−waitable)]. All math runs in your browser.

How the math works

Anthropic's batch API is 50% off for a 24-hour turnaround (verified 2026-07-13). The brochure number assumes every request can wait — but latency-sensitive traffic can't. Only your waitable share takes the discount, so the real saving on the whole bill is smaller.

Same baseline, one identity: at 95% waitable, batching cuts the $2,100/month real-time bill by $998 to $1,103 — that's 48% overall, exactly 95% waitable × 50% discount, not the headline 50%.

So the lever isn't "turn on batch", it's "how much of your traffic can tolerate 24h". Push the waitable share up (queue overnight report generation, bulk enrichment, evals, back-catalog processing) and the saving scales linearly toward the full 50%. Anything user-facing stays real-time and saves nothing.

A second lever people skip: for the non-waitable slice, a cheaper model in real time may beat batching an expensive one — worth checking on the model comparison pages. Prices sync daily from OpenRouter; the batch discount is hand-verified against Anthropic's official docs and re-checked quarterly.

Frequently asked questions

Is the batch API worth it for a bulk classification job?

At this page's defaults — 1,000,000 requests a month, 95% able to tolerate 24h — batching cuts $2,100 to $1,103/month: $998 saved, 48% overall. That's the 50% headline discount applied to only your waitable share. Worth it if a real slice of traffic can wait; near-useless if it's all user-facing.

Why isn't the batch discount just 50%?

Because 50% is the per-request discount, and it only applies to requests that can wait 24 hours. If 95% of your traffic is waitable, your overall saving is 95% × 50% = 48%. The way to save more isn't a better discount — it's moving more traffic into the waitable bucket.

What workloads can actually use batch?

Anything without a human waiting on the response: overnight report generation, bulk data enrichment and classification, embedding backfills, offline evals, synthetic-data pipelines, back-catalog reprocessing. Real-time chat, agent steps, and interactive tools cannot — they're the non-waitable share this calculator holds at full price.

Which models give the biggest batch savings here?

At this workload: o1-pro ($71,250/month saved), GPT-5.4 Pro ($15,675) and GPT-5.5 Pro ($15,675) — refreshed daily. Absolute savings track absolute spend, so expensive models with lots of waitable volume save the most dollars.

What formula does this use?

Blended monthly = requests × per-request cost × [waitableShare × (1 − discount) + (1 − waitableShare)]. Savings = real-time − blended; overall savings % = waitableShare × discount. The discount is the provider's published batch rate (50% for the big three, 24h turnaround); per-request cost uses the live standard prices.

Are these prices current?

Standard prices sync daily from OpenRouter; batch discounts are hand-verified against each provider's official batch documentation (last checked 2026-07-13) and re-audited quarterly. All math runs client-side with tested code.