OpenAI doubled the per-token price of its flagship and says you’ll make it up by using fewer tokens. GPT-5.5 pricing is the fact. The saving is a claim, and it’s one you can check against your own work in an afternoon. Run 20 of your tasks on both models and judge by what each finished job cost, not by the rate card.
Key points
- What does GPT-5.5 pricing look like? According to OpenAI’s launch page , the standard API rate is $5 per million input tokens and $30 per million output tokens.
- How many tokens does it have to save? A lot, and more than most buyers expect. OpenAI says GPT-5.5 is more token efficient and delivers frontier coding intelligence at half the cost of competitors, citing Artificial Analysis.
- Who ran the benchmarks, and what did real traffic show? OpenAI’s own table has GPT-5.5 at 82.7% on Terminal-Bench 2.0 against 69.4% for Opus 4.7, and 93.6% on GPQA Diamond against 94.2% for Opus 4.7.
- How do you test it on your own work? Pick 20 real tasks from last month, not made-up ones.
What does GPT-5.5 pricing look like?
According to OpenAI’s launch page, the standard API rate is $5 per million input tokens and $30 per million output tokens. GPT-5.4, launched in March, was $2.50 and $15. GPT-5.5 Pro costs $30 and $180. Batch and Flex processing cost half the standard rate, and Priority processing costs 2.5 times it.
The API context window is 1 million tokens, which OpenAI added in an update dated April 24 after the main announcement said it was coming soon.
Anthropic’s Claude Opus 4.7, released about a week earlier, lists at $5 input and $25 output. On sticker price alone, GPT-5.5 costs the same on input and 20% more on output. Anthropic also warns that Opus 4.7’s new tokenizer can map the same text to between 1.0 and 1.35 times as many tokens, so list prices alone don’t settle that comparison either.
How many tokens does it have to save?
A lot, and more than most buyers expect. OpenAI says GPT-5.5 is more token efficient and delivers frontier coding intelligence at half the cost of competitors, citing Artificial Analysis. Here’s the arithmetic so you can swap in your own numbers. It’s our illustration, not a measurement.
Say a task sends 3,000 tokens in and gets 1,000 back, and you run 1,000 of them.
- GPT-5.4: 3 million input tokens at $2.50 is $7.50, and 1 million output tokens at $15 is $15. Total $22.50.
- GPT-5.5 with identical token counts: $15 plus $30. Total $45.
- GPT-5.5 if output falls 40%: $15 plus $18. Total $33.
- GPT-5.5 if output falls 50%: $15 plus $15. Total $30.
Even a halved output leaves the bill 33% above GPT-5.4’s. To land at $22.50 on this mix, output would have to fall by 75%, because the 3,000 input tokens don’t shrink when the model gets more efficient. It’s a car that gets better mileage but needs dearer fuel. The mileage has to improve a lot before you save anything.
That doesn’t make OpenAI’s claim false. Reasoning models spend a lot of tokens thinking, and on long, hard jobs the saving can be large. It does mean the claim shouldn’t be accepted as a general rule.
The mix of input and output changes the break-even, so run your own shape. Flip the example to a task that sends 500 tokens in and gets 2,000 back, as in drafting work. On GPT-5.4 that’s $0.00125 for input plus $0.03 for output, about $0.031. On GPT-5.5 at the same counts it’s $0.0625. Here output only has to fall about 52% to match GPT-5.4, because output is where most of the money goes. Long-answer work has an easier road to break-even than long-prompt work, such as summarizing a big document. Know which of the two your team does most before you read anyone’s efficiency claim.
Then scale it. At 50,000 tasks a month on the first mix, GPT-5.4 costs $1,125 and GPT-5.5 at identical token counts costs $2,250. The $1,125 monthly difference is the size of the bet you’re making if you switch on the strength of a launch page. Small for some, a salary for others, and your finance lead will want to know which before the invoice arrives.
Who ran the benchmarks, and what did real traffic show?
OpenAI’s own table has GPT-5.5 at 82.7% on Terminal-Bench 2.0 against 69.4% for Opus 4.7, and 93.6% on GPQA Diamond against 94.2% for Opus 4.7. On SWE-Bench Pro, a coding test, it shows GPT-5.5 at 58.6% and Opus 4.7 at 64.3%. So OpenAI’s own chart has a rival ahead on one coding test.
Every figure there is OpenAI’s, chosen and reported by OpenAI, on tasks someone else designed. Your invoices, contracts and customer emails aren’t in them.
Then there’s a data point from after the launch. The Decoder later reported on OpenRouter’s April 2026 usage logs and found real-world effective cost per million tokens had risen between 49% and 92% from GPT-5.4, depending on input length, even though responses were shorter on long inputs. That’s a third-party reading of one platform’s traffic, and your pattern may differ. It does match the arithmetic above.
What about GPT-5.5 Instant, the new ChatGPT default?
On May 5, 2026, OpenAI made GPT-5.5 Instant the ChatGPT default, replacing GPT-5.3 Instant. Paid users keep GPT-5.3 Instant for three months. OpenAI says the new default produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law and finance, and 37.3% fewer inaccurate claims on conversations users had flagged for errors.
Both figures come from internal evaluations, and OpenAI doesn’t give the starting error rate or the sample size.
A relative cut with no starting rate can’t tell you how many errors you’ll see. If the old model was wrong on 10% of claims, a 52.5% cut leaves about 4.75%. If it was wrong on 1%, you land near 0.5%. Those starting rates are our assumptions, not OpenAI’s data.
In the API, OpenAI ships the new default as chat-latest. The name suggests it moves whenever the default changes, so pin a named model for any workflow whose output has to repeat. The 20-task test below works for this model too: grade answers you already know, with and without the source file attached.
How do you test it on your own work?
Pick 20 real tasks from last month, not made-up ones. Run each on GPT-5.4, GPT-5.5 and whichever rival you’d consider, with the same prompts. Log input tokens, output tokens, dollars and whether the answer was good enough to send. Divide total dollars by the count of acceptable answers, and that’s your cost per finished task.
Then repeat on your hardest five tasks, because that’s where a reasoning model’s token savings should show up.
The same cost-per-finished-task logic applied when we looked at Cursor Composer 2 pricing, and our Sonnet 4.6 pricing piece shows how a launch headline can overstate a saving. For the test method on a previous OpenAI model, see our GPT-5.4 computer use piece. If you’re budgeting for a premium tier, our Claude Mythos leak piece covers how to plan for one.
A fair objection is that list price is a distraction, because a better model that gets it right the first time saves your staff hours. It does, and that’s why we recommend cost per finished task over cost per token. Where we part ways is on assuming it. A better model can be worth double. It can also be worth nothing extra on the tasks you run most. The test tells you which, and it costs less than one wrong guess on a monthly bill. Keep the log. When the next model launches in a few weeks, the same 20 tasks give you an answer in an hour.
The price comparisons here are arithmetic on list prices as published on April 24, 2026, with an assumed request size, and OpenAI’s launch page changed after publication. If your own test shows GPT-5.5 finishing tasks that GPT-5.4 fumbles, or using a fraction of the output tokens on your hardest jobs, our caution becomes a recommendation to switch. Watch for independent measurements of its token use on business-style prompts, and for any competitor response on output-token prices. Until you have your own 20-task result in a spreadsheet, keep the current model as the default and treat the new one as a trial. The upgrade will still be there next week, and so will the price.
Frequently asked questions
How much does GPT-5.5 cost?
OpenAI lists GPT-5.5 at $5 per million input tokens and $30 per million output tokens. GPT-5.5 Pro is $30 and $180.
Is GPT-5.5 more expensive than GPT-5.4?
Per token, yes. GPT-5.4 listed at $2.50 and $15, so GPT-5.5’s rate is double. OpenAI says GPT-5.5 uses fewer tokens, so the cost per task depends on your workload.
How do I compare AI model costs fairly?
Run the same 20 real tasks on each model, record tokens and dollars, and divide total spend by the number of acceptable answers.
Archive entry dated April 24, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
