Do OpenAI, Anthropic and Google Gemini all give the same Batch API discount?

Yes, verified directly against each provider's current docs: all three discount both input and output tokens by exactly 50% versus their synchronous API rate, for every model tier checked, as of 2026-10-29. OpenAI states a '50% cost discount compared to synchronous APIs,' Anthropic's Message Batches API reduces costs by 50 percent, and Gemini Batch Mode is 'priced at 50% of the standard interactive API cost.'

Is the 24-hour batch window the same across providers?

No. OpenAI's completion window parameter can only be set to 24h, and batches move to an expired state at that point. Anthropic batches expire if processing does not complete within 24 hours. Gemini markets a 'target turnaround time' of 24 hours, but its documented JOB_STATE_EXPIRED only triggers after the job has been running or pending for more than 48 hours -- double the number it advertises.

Do you get charged for a batch request that expires before finishing?

For OpenAI and Anthropic, no -- both are explicit that you are only billed for requests that actually completed; anything still unprocessed when the batch expires is not charged. Google's Gemini docs do not state a billing rule for expired or incomplete batch requests at all, which is itself worth knowing before you rely on an assumption either way.

Does this affect how I should build batch job retry logic?

Yes. If your pipeline assumes a job is dead and safe to resubmit once 24 hours have passed, that assumption is correct for OpenAI and Anthropic but wrong for Gemini, where a job can still be legitimately running at hour 30 or 40 inside its real 48-hour allowance. Cancelling and resubmitting Gemini batches at the 24-hour mark risks duplicating work that was never actually stuck.

Tres API por lotes prometen una ventana de 24 horas. Uno de ellos significa 48.

29 de octubre de 2026 · Metodología del sitio y notas de precios · 6 minutos de lectura

Nuestro Explicador de API por lotes tiene una línea en sus preguntas frecuentes que ha estado sin control durante meses: "los proveedores generalmente dan un objetivo de finalización, como dentro de las 24 horas." Por lo general, nunca había sacado los documentos actuales de los tres proveedores el mismo día y había comparado los números exactos uno al lado del otro. Así que hoy lo hice. El descuento resultó ser la parte aburrida y tranquilizadora. La fecha límite es donde se pone interesante.

El descuento realmente es idéntico

Comprobado directamente de los precios actuales y los documentos por lotes de cada proveedor, 2026-10-29: la API por lotes de OpenAI ofrece "un descuento de coste del 50% en comparación con las API sincrónicas." La API de lotes de mensajes de Anthropic reduce "los costes en un 50 por ciento. El" Modo por lotes Gemini de Google "tiene un precio del 50% del coste estándar de la API interactiva." Tres competidores, tres equipos de precios independientes y los tres aterrizaron exactamente en el mismo número, se aplicaron de la misma manera tanto a los tokens de entrada como a los de salida.

Proveedor	Ejemplo → de lote estándar	Descuento
AbiertoAI	Misma tasa por token, reducida a la mitad	50%
antrópico	Misma tasa por token, reducida a la mitad	50%
Géminis 2.5 Pro	$ 1.25 → $0.625 in / $ 10.00 → $5.00 out por 1M	50%

Nuestro Calculadora de ahorro de API por lotes ya asume un 50% fijo por defecto. Esa suposición se mantiene. No se necesita ninguna corrección allí.

La "ventana de 24 horas" no es la misma 24 horas

Esta es la parte que no se sostiene. Los tres proveedores hablan de un cambio de aproximadamente un día, y asumí que eso significaba el mismo número contractual en todas partes. No es así.

OpenAI: el parámetro de la ventana de finalización "solo se puede establecer en 24h" y "cada lote se completa en 24 horas." Golpea esa marca y el lote pasa a un estado caducado: firme, número único, sin ambigüedad en los documentos.

Antrópico: "Los lotes caducan si el procesamiento no se completa dentro de las 24 horas. -El mismo número, el mismo punto de corte duro.

Gemini: "The target turnaround time is 24 hours, but in majority of cases, it is much quicker." Read only that sentence and you'd assume Gemini matches the other two. But the actual state-machine definition, a few paragraphs over in the same documentation, says something different: JOB_STATE_EXPIRED triggers only after the job "has been running or pending for more than 48 hours." Twenty-four hours is the marketing target. Forty-eight is the number that actually governs when Google will call a job dead.

Why that gap matters more than it sounds

Nobody builds retry logic around a provider's marketing copy on purpose. But if you've only ever skimmed the headline sentence of each provider's batch docs -- which is exactly what I'd done before today -- "24 hours" is the number that sticks, for all three. Build a pipeline that cancels and resubmits any batch job still open after 24 hours, and on OpenAI and Anthropic you're safe: a job still open at hour 25 really is stuck. On Gemini, you'd be cancelling and resubmitting jobs that are still inside their real allowance, possibly duplicating work and burning a second 50%-discounted run on top of a first one that was never actually going to fail.

On the billing side, there's actually good news and it's consistent across providers. OpenAI is explicit that when a batch expires, "you will be charged for tokens consumed from any completed requests" -- meaning anything that didn't finish isn't billed at all. Anthropic's batch reference says the same thing from the other direction: an expired request status means "you will not be billed for these requests." Both charge you for exactly what ran and nothing for what didn't. Gemini's docs don't state a billing rule for expired or incomplete requests either way, which isn't a red flag by itself, but it's a gap worth knowing about before you assume Gemini definitely matches the other two there too.

What I'm changing

I added a footnote to the batch explainer flagging that "24 hours" is a target for Gemini, not a hard cap, and that the real enforced limit is 48. Beyond that one page, the actual habit I'm changing is bigger: when three competitors use the same round number in their marketing copy, I now treat that as three separate numbers to verify, not one number confirmed three times. They agreed on the discount. They didn't agree on the deadline, and the only way to know that was to stop trusting the headline sentence and go find the state-machine definition underneath it.

Batch API explained — updated with the 24h-target vs 48h-actual note for Gemini.
Calculadora de ahorro de API por lotes — the 50% default is confirmed accurate across all three providers.
Batch processing time calculator — model your own turnaround assumptions against real per-provider limits.

New to batch pricing in general? Start with the guías gratuitas de costos de API.

Despliéguelo usted mismo: DigitalOcean — $200 de crédito gratis ↗ · VPS Hostinger ↗

Checked directly against developers.openai.com/api/docs/guides/batch, platform.claude.com/docs/en/build-with-claude/batch-processing, and ai.google.dev/gemini-api/docs/batch-mode, 2026-10-29. Reference figures, not live quotes -- confirm current terms on each provider's own docs before relying on a deadline for production retry logic.