AIStall
EditionBy The AI Stall editorsOctober 14, 2026

The Fifty Percent Discount You Have to Ask For

The Fifty Percent Discount You Have to Ask For

Three platforms, one number: send a pile of model requests asynchronously instead of one at a time and all three charge half. OpenAI calls it a 50 percent cost discount compared to synchronous APIs. Anthropic's batch page says costs drop by 50 percent. Google says batch costs 50 percent of standard. The discount is the easy part. What differs is how much you can put in one batch, what happens when the clock runs out, and how long the answers stay downloadable. All three pages were read on October 12, 2026. This follows our note on the cached-input price, which covers the other half of the cost question.

The headline is the same on all three platforms

Freshdesk's public homepage, captured automatically on October 13, 2026. Not a logged-in account; the page may have changed since.
Freshdesk's public homepage, captured automatically on October 13, 2026. Not a logged-in account; the page may have changed since.

Google's pricing page shows the halving as separate rows rather than a rule. Gemini 3.5 Flash-Lite is listed at $0.30 input and $2.50 output standard, and $0.15 and $1.25 batch, checked October 12, 2026.

Anthropic's batch documentation prints a table in the same shape. Haiku 5.5 batch is listed at $0.05 per million input tokens and $0.25 output for prompts up to 100,000 tokens, rising to $0.25 and $1.25 above that.

OpenAI's pricing page lists batch rows too: gpt-6-astra at $5.00 input and $25.00 output, gpt-6.1-sol at $1.00 and $5.00, gpt-4.1 at $1.00 and $4.00. Each is half the standard row.

ChatGPT Business's public homepage, captured automatically on October 10, 2026. Not a logged-in account; the page may have changed since.
ChatGPT Business's public homepage, captured automatically on October 10, 2026. Not a logged-in account; the page may have changed since.

So the choice between platforms is not about the size of the discount. It is about whether your job fits inside the limits each one sets.

The envelopes are different sizes

This is where the pages diverge, and the gap is wide enough to decide an architecture.

Anthropic's limit is twice OpenAI's on request count and slightly larger on payload. Google does not publish a request ceiling that we could read, only a file size, so a very large job needs a test run rather than an assumption.

What happens when the window closes is the part to read

All three target 24 hours, and none of them guarantees it. Anthropic says most batches finish in less than one hour. Google says the target turnaround is 24 hours, but in the majority of cases it is much quicker.

The failure behaviour is specific and worth knowing before a deadline. OpenAI says batches that do not complete within the window move to an expired state, unfinished requests are cancelled, and completed responses are still available in the output file.

Crucially, OpenAI adds that you are charged for tokens consumed by any completed requests before expiry. A half-finished batch is a half-sized invoice, not a refund.

Anthropic expires batches that do not complete within 24 hours, and makes results accessible when all messages have completed or after 24 hours, whichever comes first. It also notes that the failure of one request does not affect the others.

Results do not sit there forever

Anthropic states that batch results are available for 29 days after creation. After that the batch is still visible but the results are no longer downloadable, and the same 29 days is given as the retention period for request and response data.

Google says batch job results are stored and available for download for six weeks before being permanently deleted. That is roughly double Anthropic's window.

Either way, a pipeline that writes results into its own database on the same day is safer than one that treats the vendor as storage. Retention rules move, and they are set by the vendor, not by you.

The wider retention picture, including what each vendor keeps for abuse monitoring, is a separate question from batch storage, and the answers are longer than 29 days.

Batch and caching pull in opposite directions

Both mechanisms cut the same bill, and they interfere. Anthropic says cache pre-warming with max_tokens set to zero is not supported inside a batch, because an ephemeral cache entry written during batch processing would likely expire before the follow-up request runs.

That is the five minute default time to live colliding with a job that may take an hour. Teams planning to do both should read the caching floors and clocks in our caching note first.

Google's documentation points the other way for one surface: the Interactions API supports implicit caching only, which means explicit cache control and batch work live on different endpoints.

Which jobs actually fit

Batch suits work where nobody is waiting for the answer. Anthropic's own list names evaluations, data analysis and bulk content generation, and that is a fair description of the shape.

It does not suit anything a person is watching. Chat assistants sold by the seat, such as ChatGPT Business and Claude Team, are priced per person precisely because the user is waiting.

Cheap small models narrow the case further. When a model already costs ten cents per million input tokens, as Haiku 5.5 does, halving it saves five cents and costs you a day.

Bottom line

The 50 percent is real and identical across OpenAI, Anthropic and Google. Treat it as the default for any job without a person attached, and read the expiry clause before trusting it with a deadline.

For readers buying seats rather than tokens, none of this appears on the bill. Our guide to AI assistant plans for small teams and the plan comparison cover that side, and issue zero explains why we link the source for every figure.

Sources