GPT-5.5 API Cost: Per-Token Prices and Real Workload Math

GPT-5.5, OpenAI’s flagship, costs $5 per million input tokens and $30 per million output tokens as of June 2026, with batch jobs at half those rates; the live source is openai.com/api/pricing. In concrete terms: a mid-sized PR review runs about $0.21, a 100-document summary pass about $4.20, and a thousand agent tasks about $994 a month. This page works each of those numbers, then shows where the flat subscription alternative sits.

The June 2026 price sheet

Model	Input /1M	Output /1M	Batch input / output
GPT-5.5	$5.00	$30.00	$2.50 / $15.00
GPT-5.4	$2.50	$15.00	$1.25 / $7.50
GPT-5	$1.25	$10.00	$0.625 / $5.00
GPT-5 Mini	$0.25	$2.00	$0.125 / $1.00

GPT-5.5 sits at 4x GPT-5 and 20x Mini on both sides of the meter. Cached-input rates per model sit alongside these on OpenAI’s pricing page, and OpenAI adjusts prices as models rotate, so treat this as a June 2026 snapshot. The full lineup, discounts included, is in OpenAI API pricing explained.

The ratio worth memorizing: output costs six times input. GPT-5.5 charges $5 to read a million tokens and $30 to write them, so the bill follows what the model says, not what you send.

Job 1: reviewing a pull request

A mid-sized PR review: 15,000 input tokens (the diff, surrounding files, review guidelines) and 1,000 visible output tokens of comments. Reasoning models think before they answer, and the thinking bills as output; call it 3,500 reasoning tokens, for 4,500 billed output.

input    15,000 × $5/1M  = $0.075
output    4,500 × $30/1M = $0.135
per review                 $0.210
300 reviews a month        $63

The reasoning share is the swing variable: a review that thinks hard can double the output line without adding a word to the visible comments. That multiplier has its own writeup in reasoning tokens explained.

Job 2: summarizing 100 documents

100 documents at 6,000 input and 400 output tokens each:

input    600,000 × $5/1M  = $3.00
output    40,000 × $30/1M = $1.20
per run                     $4.20
nightly for a month         $126

The same run costs $1.05 on GPT-5 ($31.50 a month) and $0.23 on GPT-5 Mini ($6.90 a month). Bulk summarization rarely justifies the flagship; the tier-by-task guide is in the cheapest OpenAI model that still does the job.

Job 3: an agent session

Using the reference agent from the agent cost formula: 8 steps, 18.4 effective calls per task after retries, averaging 6,000 input and 800 output tokens per call.

per call   6,000 × $5/1M + 800 × $30/1M = $0.030 + $0.024 = $0.054
per task   18.4 × $0.054                = $0.99
1,000 tasks a month                       $994

The same agent runs $285 a month on GPT-5 and $57 on Mini. A thousand GPT-5.5 agent tasks a month is about $994 on the meter, which is the number to hold onto for the next section.

Where GPT-5.5 earns its premium

Three places, in our experience reading request logs: planning steps inside agents, where a wrong plan multiplies the cost of every step after it; debugging and review tasks where the cheaper tier’s miss rate costs human time; and long-context synthesis where weaker models drop threads. Everywhere else, route down. A stack that sends 20 percent of calls to GPT-5.5 and 80 percent to GPT-5 or Mini usually beats an all-flagship stack on cost by 3x or more with no visible quality change. The practical test is cheap to run: sample 50 production tasks, run them on the tier below, and count how many a reviewer can actually tell apart.

Two discounts before you switch lanes

Both of OpenAI’s standing discounts apply to GPT-5.5, and they map cleanly onto the jobs above. The Batch API halves both sides for work returned within 24 hours: the nightly 100-document run is exactly that shape, so its $126 month becomes $63 with no quality change. Prompt caching bills repeated prefixes at the per-model cached rates on the pricing page, and the agent’s stable system-plus-tools prefix is the natural candidate, since it repeats on all 18 calls of every task.

Worth being clear about what discounts do: they scale the meter, they do not change its shape. A bill that is half as steep still grows with every document, task, and retry, which is what the next section is about.

The subscription contrast

Per-token math is one way to buy frontier-model work. ChatGPT plans are the other: they include Codex, OpenAI’s coding agent, which runs programmatically by design and bills against flat plan windows instead of the meter. We run the official Codex CLI signed in with your own ChatGPT account and expose it as an OpenAI-compatible endpoint; the model surface is whatever Codex serves, and the Codex lane returns complete responses rather than streams.

As capacity estimates, never guarantees: Plus at $20 absorbs roughly $700 of API-equivalent work a month, Pro 5x at $100 roughly $3,500, Pro 20x at $200 roughly $14,000. The $994 agent month above sits inside the Pro 5x estimate at $100 for the plan plus our $129 fee, about $229 all-in versus $994 metered. The full crossover math, including when the meter wins, is in OpenAI API vs ChatGPT subscription cost.

Price your own workload at GPT-5.5 rates in the calculator; it shows the metered total next to the flat setup that covers it.