Industry News
The price of GPT Image 2.6 remains unchanged, but the cost of the low-tier plan has increased by four times.
The pricing of GPT Image 2.6 is exactly the same as that of 2.5: $5 per million tokens for text input, $8 for image input, and $30 for image output.
However, for the same prompt and the same 1024×1024 icon, the cost of the low-tier version of the bill increased from $0.0061 to $0.0250.
The unit price remained unchanged; what increased was the number of output tokens required per image. On the night of its release, a developer conducted thorough testing on Flare.
| quality grade | 2.5 tokens per fee | 2.6 Tokens and Fees | The time required to increase the value from 2.5 to 2.6 |
|---|---|---|---|
| low | 196 / $0.0061 | 824 / $0.0250 | 11.8s → 5.9s |
| medium | 439 / $0.0134 | 1449 / $0.0437 | 8.3s → 5.9s |
| high | 1758 / $0.0529 | 1758 / $0.0530 | 17.2s → 7.9s |
| xhigh | 3122 / $0.0939 | 3122 / $0.0939 | 25.3s → 11.8s |
| max | 7024 / $0.2110 | 7024 / $0.2110 | 45.1s → 18.0s |
The token capacity of the “low” tier is approximately 4.2 times that of the previous generation; the “medium” tier offers about 3.3 times more capacity. For the “high” tier, both the capacity and speed are roughly twice as high as those of the previous generation.
Therefore, with the update to version 2.6, there was no overall increase in prices; instead, only the lower price ranges were raised. Those who usually use the “high” settings for generating maps will see a significant improvement in speed, while those who rely on the “low” settings to produce large numbers of small-scale maps will be surprised by the resulting costs.
The speed of promotion can increase by nearly 50% – it all depends on which model you choose to use.
OpenAI claims that the generation speed has increased by approximately 50%, and the Flare data even indicates an even greater improvement. However, Sunburst is another story: within the same set of tests, its logo generation took about 17% longer than that of its predecessor, while the generation of its mascot images took approximately 27% more time.
The results of another set of tests were even more disappointing: When the three static images were processed together, version 2.5 took 39.75 seconds to process and cost approximately 0.045 US dollars, while version 2.6 required 48.74 seconds and incurred costs of around 0.125 US dollars. Moreover, no significant difference in quality was observed between the two versions.
Both sets of data represent small sample sizes collected on the day of release. One notebook was used to run the tests throughout the night, but it’s impossible to determine the exact server load at that time. Don’t regard these results as definitive conclusions; instead, use them as a reminder: before going live, run the tests again using your own test parameters and pay attention to the “usage” value displayed in the response results.
Those tiny print items on the bill that are easily overlooked…
For batch processing, Version 2.5 of Flare still allows discounts, with the cost amounting to only $15 per million outputs; however, this option is not available in Version 2.6. In summary, for tasks that require nighttime batch processing, the cost associated with Version 2.5 is approximately one-fourth of that incurred with Version 2.6.
For GIF files, a format with 12 frames and medium quality costs approximately $0.056 per frame, while a format with 16 frames and high quality costs around $0.108 per frame. The more frames there are, the higher the cost will be.
SVG. It’s probably best not to include this in the production process for now. On the night of its release, someone made 15 requests, and 14 of them resulted in a 502 error code. In comparison with Recraft V4.1, out of 48 valid requests only 9 were successful; nearly all cases involved prompts that contained text, and these prompts failed to display properly. Aside from this…output_formatIt also requires something that was not mentioned in the generation guidelines.svg.modelParameters; billing is also divided into two parts.
Marking editing: You can select a specific area and ask the model to make changes in that region. This feature is also available in version 2.5, and the cost is approximately 0.022 US dollars; the same operation in version 2.6 costs around 0.052 US dollars.
Version locking: Currently, version 2.6 does not include snapshot dates. Therefore, if there is any change in the behavior of the model, you will have to detect it yourself.
There is also a free option: the newly added complexity parameter. When set to the low, medium, or high level, it results in 1,132 tokens and a cost of $0.0347 in all cases. However, the default value is “high”, and the documentation warns that setting it to a higher level may cause unwanted changes. Therefore, you might consider manually lowering this value when necessary.
Based on the required amount, I would prepare it in this way.
Thousands of small images are generated every day: Keep using the 2.5 setting and process them in batches.
For print outs at resolution 300dpi or higher: it is advisable to use the 2.6 Flare setting. It offers faster processing at the same cost – so there is absolutely no reason not to choose it.
For multi-round editing tasks that require high consistency, both versions of Sunburst should be tested. Version 2.6 is slightly slower, but it’s worth waiting for if the additional stability it provides is truly worthwhile.
For teams that generate dozens or even hundreds of images in just one month and have no development resources at all, this table is basically irrelevant. Trying to save a few cents by accessing APIs, implementing retry mechanisms, or constantly monitoring usage statistics will only result in higher labor costs in the long run. In such cases, using GPT Image 2.6 on PixPix is far more convenient: you can use Flare to quickly explore different design directions and Sunburst to finalize and make continuous adjustments. You can see exactly how many credits are required for each step right within the platform, without having to manually convert tokens yourself.
There’s a somewhat naive way to save money: when making edits, only modify one part at a time. If you try to include five requirements in the same prompt, and something goes wrong, you’ll have to start over with the entire document – and each round of editing will be charged as if you were working on the whole document. This tip applies whether you’re using ChatGPT, APIs, or PixPix.
This table is unlikely to remain valid for too long. OpenAI’s documentation currently contains conflicting information; the rules regarding token consumption and batch processing capabilities may change at any time. It is recommended to check the official price page again before making a decision to switch to this system.