How much does the OpenAI batch API cost versus synchronous?

50% of the synchronous rate - half price - on both input and output tokens, across every OpenAI model. That is OpenAI's own published batch price, not a batchwatch estimate. What the price table cannot tell you is how long the queue takes, which is the one thing we measure.

Is the batch discount worth the wait?

For most work, yes. The saving is a flat 50%, and the wait is short for the fast models: gpt-5.6-sol clears in a measured median of 78s. It depends on your deadline, not on folklore - the "up to 24 hours" window is a ceiling, and the real median is far below it. Send your own limit to /v1/should-i-batch and we answer against it.

Where do these prices come from?

OpenAI's own published per-token rates, one row per model. We quote them as OpenAI's, never as ours, and we keep them honest with a scheduled drift check against a broad price feed. The batch column is simply the sync rate at 50% - OpenAI's own discount applied to both input and output tokens.

Are the queue times measured or estimated?

Measured. Each is an observed completion of a real batch job, timed on our server from submit to result, not a prediction and not the provider's marketing window. We publish a median only once a model has at least 8 completions; below that we show the count and say so, because a median on a handful of jobs is just the slowest of the handful.

How fresh are the numbers?

Live. The prices are checked against OpenAI's pages on a schedule, and the queue times are rendered straight from the rollup on every request - public figures delayed by 15 minutes, which is exactly what a paying customer buys out of. A static number in a pricing article is wrong within days; ours moves with the queue.

OpenAI API pricing & the batch discount, timed

The half-price question

50% of synchronous OpenAI's own published batch rate, on every input and output token

OpenAI's batch API costs 50% of the synchronous price: a job that costs $100 in real time costs $50 in batch. The catch the price never shows is the wait - the provider quotes a "within 24 hours" ceiling and stops. That ceiling is folklore repeated until it sounds like a number; the real median is far below it, and it is different for every model. Below is the price, and then the wait that comes with it - the half nobody else measures.

The per-model price, sync and batch

Model	Sync $/MTok (in / out)	Batch $/MTok (in / out)	Saving
gpt-5.5	$5 / $30	$2.5 / $15	50%
gpt-5.6-sol	$4 / $20	$2 / $10	50%
gpt-5.4	$2.5 / $15	$1.25 / $7.5	50%
gpt-5.6-terra	$2 / $12	$1 / $6	50%
gpt-5.4-mini	$0.75 / $4.5	$0.375 / $2.25	50%
gpt-5-mini	$0.25 / $2	$0.125 / $1	50%
gpt-5.4-nano	$0.2 / $1.25	$0.1 / $0.625	50%
gpt-5.6-luna	$0.2 / $1.2	$0.1 / $0.6	50%
gpt-5-nano	$0.05 / $0.4	$0.025 / $0.2	50%

Sync is OpenAI's own published $/MTok rate; batch is that rate at 50% (OpenAI's own discount, applied to input and output equally). Every model links to its measured queue-time page.

What the discount costs you in time

Model	Median wait (p50)	p90	n
gpt-5.6-sol	78s	10 min	520
gpt-5-nano	88s	3 hours	508
gpt-5.6-luna	5 min	5 hours	561

Queue times rendered live from the rollup; newest measurement 2026-08-30. Prices verified against OpenAI's own pages (see how we measure).

Questions this answers

How much does the OpenAI batch API cost versus synchronous?: 50% of the synchronous rate - half price - on both input and output tokens, across every OpenAI model. That is OpenAI's own published batch price, not a batchwatch estimate. What the price table cannot tell you is how long the queue takes, which is the one thing we measure.
Is the batch discount worth the wait?: For most work, yes. The saving is a flat 50%, and the wait is short for the fast models: gpt-5.6-sol clears in a measured median of 78s. It depends on your deadline, not on folklore - the "up to 24 hours" window is a ceiling, and the real median is far below it. Send your own limit to /v1/should-i-batch and we answer against it.
Where do these prices come from?: OpenAI's own published per-token rates, one row per model. We quote them as OpenAI's, never as ours, and we keep them honest with a scheduled drift check against a broad price feed. The batch column is simply the sync rate at 50% - OpenAI's own discount applied to both input and output tokens.
Are the queue times measured or estimated?: Measured. Each is an observed completion of a real batch job, timed on our server from submit to result, not a prediction and not the provider's marketing window. We publish a median only once a model has at least 8 completions; below that we show the count and say so, because a median on a handful of jobs is just the slowest of the handful.
How fresh are the numbers?: Live. The prices are checked against OpenAI's pages on a schedule, and the queue times are rendered straight from the rollup on every request - public figures delayed by 15 minutes, which is exactly what a paying customer buys out of. A static number in a pricing article is wrong within days; ours moves with the queue.

See it live

Every queue time here is measured, and it is all public.

The live dashboard shows every model with its full distribution, n and confidence grade. /compare ranks every provider's batch API side by side, and /v1/coverage is the same data as JSON, no key required. Send your own deadline to /v1/should-i-batch and we tell you, for your limit, whether the discount is worth the wait.