Batch API

Q: What is Flex processing in the OpenAI API?

Flex is a beta service tier, set with service_tier: "flex", that bills normal Responses or Chat Completions calls at Batch prices in exchange for slower answers and possible 429 errors, which are not charged.

The OpenAI Batch API runs up to 50,000 requests from one file within 24 hours at 50% off the normal price. Flex processing gives the same discount on ordinary calls that can be slow, Fast mode costs double for quicker answers, and prompt caching cuts the price of repeated prompt prefixes.

The OpenAI Batch API takes a file of up to 50,000 requests, finishes them within 24 hours, and charges 50% less than the same calls made one by one. It is the simplest way to cut an OpenAI API bill in half for work nobody is waiting on: classifying a backlog, embedding documents, running evaluations, generating images in bulk. Three related options trade price against speed on ordinary calls: Flex processing (also half price, but slower and sometimes refused), Fast mode (double price, quicker), and Ultrafast (six times the price, for GPT-6 Astra and GPT-6.1 Sol only). Prompt caching lowers the price of any prompt that repeats a long prefix. All four are explained below, with prices as of Nov 6, 2026.

Quick facts

Batch discount	50% lower cost on input and output
Completion window	24 hours, the only option
Batch size	Up to 50,000 requests, input file up to 200 MB
Batch creation	Up to 2,000 batches per hour
Flex	`service_tier: "flex"`, Batch prices, beta
Fast (formerly Priority)	2x Standard, renamed Jul 30, 2026
Cache read on GPT-6.1 Sol and Terra	0.05x input
Minimum cacheable prompt	1,024 tokens

How the Batch API works

Write a JSONL file, one request per line, each with its own id, method and endpoint. Up to 50,000 requests and 200 MB (Batch guide).
Upload the file to OpenAI.
Create the batch, naming the endpoint and the 24h completion window.
Poll the batch until it completes, or come back later; results are guaranteed within 24 hours.
Download the output file and match each result to its request by id.

The endpoints a batch can call, as of Nov 6, 2026: /v1/responses, /v1/chat/completions, /v1/embeddings, /v1/completions, /v1/moderations, /v1/images/generations and /v1/images/edits (Batch guide). Image models joined on Feb 10, 2026. Batches have their own queue limit, counted in input tokens queued per model, separate from the per-minute limits (rate limits guide); GPT-6 Astra's queue is 3M tokens on Build, 200M on Launch and 15B on Grow (GPT-6 Astra).

Batch prices for the flagship models

Model	Standard (in / cached / out)	Batch (in / cached / out)
gpt-6-astra	$10.00 / $1.00 / $50.00	$5.00 / $0.50 / $25.00
gpt-6.1-sol	$2.00 / $0.10 / $10.00	$1.00 / $0.05 / $5.00
gpt-6.1-terra	$1.00 / $0.05 / $5.00	$0.50 / $0.025 / $2.50
gpt-6-luna	$0.10 / $0.01 / $0.50	$0.05 / $0.005 / $0.25

Source: OpenAI API pricing, short context, per 1M tokens, read Nov 6, 2026.

Image models are half price in Batch too (GPT Image 2 image tokens $4.00 in, $15.00 out), but the pricing page lists no Batch row for the newest GPT-Image-2.6 models. See GPT Image.

Flex processing

Flex is for calls that can wait but do not fit a batch file. You make an ordinary Responses or Chat Completions call with service_tier: "flex" and pay Batch rates. The catch: responses are slower, and when capacity is short the API answers "429 Resource Unavailable", which is not charged. Flex is in beta. On the older models it is offered for short context only, on gpt-5.4-mini, gpt-5.4-nano, gpt-5.2, gpt-5.1, gpt-5, gpt-5-mini, gpt-5-nano and o3 (pricing).

Batch or Flex? Use Batch when you have many requests at once and can wait up to a day. Use Flex when requests arrive one at a time and a slow or retried answer is acceptable, such as background jobs in an app.

Fast and Ultrafast

Fast mode is the opposite trade: twice the Standard price for "up to 2.5x faster speeds and more consistent latency", a figure OpenAI states for gpt-5.6-sol. It replaced Priority processing on Jul 30, 2026, and service_tier accepts either "fast" or "priority". Fast covers the GPT-5.6 models, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna and GPT-6 Astra; Astra's Fast mode carries no latency SLA (Fast mode guide).

0:23 Ultrafast is offered only for GPT-6 Astra and GPT-6.1 Sol, at $60 / $300 and $12 / $60 per 1M input and output tokens, six times Standard. In OpenAI's Sep 29, 2026 Ultrafast video, GPT-6 Astra in Ultrafast generates tokens in the API "over seven times faster than standard tier" and over four times faster than Fast; Unite.AI's DevDay report put it at up to 6x faster (Unite.AI). These are OpenAI's speed claims; no independent latency measurement was found for this page.

Prompt caching

When a request starts with the same tokens as a recent one, OpenAI serves that prefix from cache at a much lower price (prompt caching guide, read Nov 6, 2026):

Minimum: 1,024 tokens of prompt for GPT-5.6 and later.
Cache reads: 0.1x the input price on GPT-5.6 and later, and 0.05x on GPT-6.1 Sol and GPT-6.1 Terra. OpenAI sums this up as "up to 95%" off.
Cache writes: 1.25x the input price for the default 30-minute lifetime, 2.5x for the 12-hour lifetime (Responses API only, since Oct 20, 2026).
Setting: prompt_cache_options.ttl = "30m" or "12h". Older models use prompt_cache_retention with in_memory (typically 5 to 10 minutes, up to an hour) or 24h; code moving from GPT-5.5 or earlier must switch to the new field (latest model guide).
Privacy: compatible with Zero Data Retention.

How to get more cache hits

Put the parts that never change first: system instructions, tool definitions, long documents.
Put the user's message and anything that varies last.
Use the 12-hour cache only when a prefix will be reused across hours; its write costs twice the 30-minute write.
Change reasoning effort with a configuration_update item rather than a new request, so GPT-6 keeps the cache (latest model guide).

OpenAI added Prompt Cache Diagnostics on Sep 8, 2026 to show why a request missed the cache. Caching and Batch stack: batch cached input on GPT-6.1 Sol is $0.05 per 1M tokens.

Which option to pick

Need	Use	Price vs Standard
Thousands of requests, results by tomorrow	Batch API	0.5x
Background calls one at a time, can be slow	Flex	0.5x
Normal interactive app	Standard	1x
Faster, steadier answers	Fast	2x
The fastest Astra or Sol	Ultrafast	6x
Long repeated prompts	Prompt caching, with any of the above	Reads 0.05x to 0.1x input

Source: OpenAI API pricing, read Nov 6, 2026.

All prices: OpenAI API Pricing. Limits per tier: API Rate Limits. Anthropic's Claude API offers the same 50% batch discount; see OpenAI API vs Claude API.

Frequently asked questions

How much does the OpenAI Batch API save? The OpenAI Batch API costs 50% less than standard calls on both input and output tokens; GPT-6.1 Sol, for example, drops from $2 and $10 to $1 and $5 per 1M tokens, as of Nov 6, 2026.

How long does an OpenAI batch take? OpenAI's Batch API completes each batch within a 24-hour window, the only window offered, so plan for results up to a day later.

What is Flex processing in the OpenAI API? Flex is a beta service tier, set with service_tier: "flex", that bills normal Responses or Chat Completions calls at Batch prices in exchange for slower answers and possible 429 errors, which are not charged.

What replaced OpenAI Priority processing? Fast mode replaced Priority processing on Jul 30, 2026. It costs twice the Standard price, and service_tier still accepts "priority" as well as "fast".

How does OpenAI prompt caching work? OpenAI caches prompt prefixes of at least 1,024 tokens; later requests that repeat the prefix pay 5% to 10% of the input price for it on GPT-5.6 and later models, after a write charged at 1.25x input for 30 minutes or 2.5x for 12 hours.