AIStall
EditionBy The AI Stall editorsNovember 2, 2026

The Fifty Percent You Get for Waiting

The Fifty Percent You Get for Waiting

Three AI platforms sell the same tokens at half price if you let them answer tomorrow. The headline discount is identical. What you give up for it is not.

This note reads the batch pricing and the batch limits published by OpenAI, Anthropic and Google, as the documentation showed them on October 30, 2026.

Cursor's public homepage, captured automatically on October 23, 2026. Not a logged-in account; the page may have changed since.
Cursor's public homepage, captured automatically on October 23, 2026. Not a logged-in account; the page may have changed since.

All three print 50 percent. The differences sit in what the discount covers, how long you wait, and what happens to a job that does not finish in time.

The flagship rate cards halve exactly

The OpenAI batch guide describes 50 percent lower costs. The flagship pricing table carries a Batch tab, and every figure on it is exactly half the figure on the Standard tab.

GitHub Copilot Business's public homepage, captured automatically on October 24, 2026. Not a logged-in account; the page may have changed since.
GitHub Copilot Business's public homepage, captured automatically on October 24, 2026. Not a logged-in account; the page may have changed since.

The Anthropic pricing page states that the Batch API carries a 50 percent discount on both input and output tokens. Its batch table reads $5 and $25 on Claude Fable 5.1, against $10 and $50 on standard.

Claude Opus 5.5 runs $2 and $10 in batch against $4 and $20 standard. Claude Sonnet 5.5 is $1 and $5 against $2 and $10. Claude Haiku 5.5 is $0.05 and $0.25 against $0.10 and $0.50.

The Gemini batch documentation puts the figure at 50 percent of the standard interactive API cost. The Vertex AI table prints the result: Gemini 4 Argon at $1.00 input and $5.00 output in batch mode.

Those batch figures are half the global standard rates of $2.00 and $10.00 that the same table shows above them. The arithmetic checks on all three vendors, which is rarer than it sounds.

The discount reaches the cached rows on two platforms, not the third

OpenAI halves the caching columns along with the token columns. Cached input on gpt-6-astra falls from $1.00 to $0.50, and the 30-minute cache write from $12.50 to $6.25.

Anthropic is explicit that its cache multipliers stack with other pricing modifiers, including the Batch API discount and data residency. A cache hit inside a batch costs half what the same hit costs outside one.

Google goes the other way. Its batch page says that if a request in your batch results in a cache hit, you pay the standard context caching rates.

For a job built on one large shared prefix, that is the difference between two discounts and one. The size of the gap depends on how much of the prompt repeats, which is the subject of the cached-input price note.

The turnaround figure is a deadline, not an average

OpenAI advertises a clear 24-hour turnaround time, and the create call takes a completion_window of 24h. Jobs that miss it move to an expired state, and unfinished requests inside them are cancelled.

Completed requests in an expired batch still land in the output file. OpenAI says you will be charged for tokens consumed from any completed requests, so a half-finished job produces a half-sized invoice.

Anthropic says most batches complete within one hour. Results become available when every message has completed or after 24 hours, whichever comes first, and batches expire if processing does not complete in that window.

Its billing on expiry is kinder. The documentation says an expired request reached its 24-hour expiration before it could be sent to the model, and that you will not be billed for these requests.

Google publishes a 24-hour turnaround objective and a longer hard stop. JOB_STATE_EXPIRED means the job was running or pending for more than 48 hours, and the job will not have any results to retrieve.

So the same word covers three behaviours: billed for what finished, not billed at all, and nothing to collect. That is the first thing to check before a batch job carries anything that matters.

One batch is capped three different ways

The caps are not published in comparable units, which makes a like-for-like reading impossible until you convert your own workload into each vendor's unit first.

Anthropic caps the processing queue by tier as well. Its rate limits page lists 200,000 queued batch requests on Start, 300,000 on Build and 500,000 on Scale, against 1,000, 2,000 and 4,000 requests per minute.

Requests per batch stay at 100,000 on all three tiers. Moving up a tier buys queue depth, not a bigger single job, which is the same pattern described in the note on how account tiers price throughput.

On one platform the batch pool is genuinely separate

OpenAI states that Batch API rate limits are separate from existing per-model rate limits, and that using the Batch API will not consume tokens from your standard per-model rate limits.

It also says the Batch API currently has no output-token limit. The queued prompt token ceiling per model is not printed in the guide at all; the page points to the Platform Settings page inside the account.

That is a readable instruction rather than a readable number. Anyone sizing a nightly job has to open their own console to find it, as set out in the monthly cap note.

Anthropic shares a pool instead. Its documentation says rate limits apply both to Batches API HTTP requests and to the number of requests within a batch still waiting to be processed.

Two lines in the small type that change a budget

Anthropic warns that because of high throughput and concurrent processing, batches may go slightly over the spend limit configured in the Claude Console for your workspace.

That matters to anyone treating a console spend limit as a hard stop. On this path the vendor says plainly that it is not one.

The second line is about caching. Anthropic recommends the one-hour cache duration for batches, because a five-minute entry written during batch processing would likely expire before the follow-up request runs.

Results also have a shelf life. Anthropic keeps batch results downloadable for 29 days after creation. After that the batch is still visible in the console and the results are not.

When half price is the wrong trade

The last item is the vendor's own caution. Anthropic says a single 300k-token generation can take over an hour to complete, and tells readers to plan submissions with the 24-hour window in mind.

Long prompts carry a second effect on top of the batch discount, because some rate cards reprice past a published size. That threshold is covered in the long-context price note.

Bottom line

The 50 percent is real, uniform and easy to verify. Every batch column on all three rate cards is exactly half its standard neighbour, and the three vendors say so in words as well.

The fine print is where they part. One halves the cache rows, one stacks the discount explicitly, and one leaves cache hits at full price inside a batch.

For context budgets that interact with all of this, see what a million-token window actually holds. Batch work is where a large window gets cheap enough to use fully.

Coding assistants that could run evaluation jobs this way are graded in the Cursor review and the GitHub Copilot review, alongside the shortlist of AI assistant plans.

What this brief prints and what it cuts is set out in issue zero.

Sources