What it costs to run the OpenAI agents API

IC Guides · AI Agents · Build an agent

By IC Lab Desk · Edited by Yajush Gupta · How we report

OpenAI does not sell an agent by the seat or by the task. You pay per million tokens for the model you pick, plus separate fees for built in tools such as web search, file search and containers. OpenAI states that the Responses, Chat Completions, Realtime, Batch and Assistants APIs are not priced separately. Your bill is therefore the model rate multiplied by how many tokens your agent burns before it finishes, and that depends on how it loops.

Who it fits

Teams with engineers who want control over how the agent works. If you want a finished support or sales agent with a published outcome price, read quote only support agents or Fin outcome pricing.

Model rates, per million tokens

Model (standard, short context)	Input	Cached input	Output
gpt-6-astra	$10.00	$1.00	$50.00
gpt-6.1-sol	$2.00	$0.10	$10.00
gpt-6-luna	$0.10	$0.01	$0.50

Long context roughly doubles input, Astra goes to $20 in and $75 out, Sol to $4 and $15, Luna to $0.20 and $0.75. Cache writes are billed too, at $12.50, $2.50 and $0.125 per million for the three models. OpenAI also lists Batch, Flex, Fast and Ultrafast tabs with different rates, which we did not restate. Regional data residency endpoints carry a 10% uplift for models released on or after 5 March 2026, and FedRAMP endpoints a further 10% uplift. Priority processing was renamed Fast mode on 30 July 2026. Promotional pricing on GPT 5.6 Sol runs at least through 21 November 2026, so expect that rate to change.

Tool and infrastructure rates

Item	Price
Web search	$10.00 per 1,000 calls, plus search content tokens at model rates. The non reasoning preview version is $25.00 per 1,000 calls with content tokens free.
Containers (Hosted Shell, Code Interpreter)	Per 20 minute session, 1 GB is $0.03, 4 GB $0.12, 16 GB $0.48 and 64 GB $1.92. Billed by the minute with a 5 minute minimum.
File search storage	$0.10 per GB per day, first GB free
File search tool call	$2.50 per 1,000 calls (Responses API only)
ChatKit uploads	$0.10 per GB day after 1 GB free per account per month
Voice	gpt-live-1 is $0.05 a minute, billed per second, with backend model and tool usage charged separately. gpt-transcribe is $0.0045 a minute.

Worked examples (our arithmetic)

Take an agent that runs 8 steps. Assume it starts with 3,000 tokens of context, each step adds 1,000 tokens of tool output, and it writes 300 tokens of output per step. Because every step resends the history, input tokens across the run are 3,000 plus 4,000 plus 5,000 and so on to 10,000, which sums to 52,000. Output is 2,400 tokens.

Model	Input cost	Output cost	One run, no caching
gpt-6-luna	52,000 × $0.10 / 1M = $0.0052	2,400 × $0.50 / 1M = $0.0012	$0.0064
gpt-6.1-sol	52,000 × $2 / 1M = $0.104	2,400 × $10 / 1M = $0.024	$0.128
gpt-6-astra	52,000 × $10 / 1M = $0.52	2,400 × $50 / 1M = $0.12	$0.64

Same task, a hundredfold spread between the cheapest and dearest model. Caching helps, if 80% of input hits the cache on Sol, input falls to 41,600 × $0.10 / 1M plus 10,400 × $2 / 1M, or about $0.025, and the run drops to roughly $0.049. That ignores cache write fees, so treat it as a floor. A run that fails and retries three times costs four runs, so $0.51 on Sol without caching.

Add tools on top. Three web searches add $0.03 in call fees before the content tokens. A 1 GB container session is $0.03. At 10,000 runs a month, Sol with no caching and three searches is about $1,280 for tokens plus $300 for searches, near $1,580. These are our assumptions about the loop, not measurements of your agent.

What is not confirmed

OpenAI’s separate guide for hosted agents may carry its charges. The pricing page lists tools and containers, but we did not read the agents guide line by line. We also did not restate Batch, Flex, Fast or Ultrafast rates, which change the sums.

The page does not say whether a ChatGPT subscription includes any API credit, so treat them as separate bills.

Where costs surprise you

History resends, so cost per task rises faster than step count. Three failed attempts mean four times the cost. Prompts past the short context line cost about double on input.

Web search is $0.01 a call, and the retrieved text is billed again as input. A container bills at least 5 minutes even if it runs for seconds, and data residency adds 10%.

How to test it

1. Build the agent on 20 real tasks from your queue, not demo prompts.
2. Log input, cached and output tokens per finished task, and the number of tool calls.
3. Run the same 20 on a smaller model and compare the task success rate against the cost.
4. Set a hard spend limit and a per task step cap before you widen access.

Questions to ask before you buy

1. What does one finished task cost on our data, including retries?
2. Which caps and alerts stop a runaway loop?
3. Which tools and storage do we use, and at what rate?
4. Does our prompt length cross the long context line?
5. Do we need data residency, and is the 10% uplift in the budget?

Sources

• OpenAI, API Pricing, developers.openai.com/api/docs/pricing, checked 12 October 2026.

Worked examples are IC arithmetic from published rates and our stated assumptions. We did not test the product. See the OpenAI and Anthropic rows in the AI Agent Pricing Tracker. Send a correction.

Related guides Build or buy · Workflow automation pricing · Voice agent pricing · Claude Code usage limits · AI agent glossary · AI Agent Pricing Tracker · All AI agent guides