If I had a production line that could generate thousands of small images every day and kept running on the lower-tier settings of GPT-Image 2.5, I’d probably have to “move” it somewhere else this week.
On November 4th, OpenAI released GPT Image 2.6, and the pricing list remained unchanged. Two days later, Google launched Nano Banana 2 Lite—a name that clearly indicated its focus on cost-effectiveness and speed. It seemed like Google was the better option in this case. I gathered all the available data and did some calculations, but my conclusion wasn’t so clear-cut after all.
The reason for moving: the price of “low-tier items” has quietly quadrupled.
The pricing for version 2.6 is the same as that for version 2.5: $5 per million tokens for text input, $8 for image input, and $30 for image output. The only difference lies in the number of output tokens required for processing each individual image.
On the night of the release, some developers used the same prompt and the same 1024×1024 icon to test each tier on Flare one by one. For the “low” tier, the number of tokens increased from 196 to 824, and the price per token rose from 0.0061 US dollars to 0.0250 US dollars; for the “medium” tier, the price increased from 0.0134 US dollars to 0.0437 US dollars. As for the “high” tier, the prices of the two higher tiers were almost identical, both being 0.0529 US dollars.
Therefore, it is precisely those who purchase large quantities of products that are affected by such billing practices.
What’s more troublesome is the batch processing aspect. With version 2.5 of Flare, it was possible to perform batch tasks at a discounted rate, with the cost amounting to only $15 per million outputs. However, this option is not available in version 2.6. For night-time batch processing tasks, the cost associated with version 2.5 is approximately one-fourth of that in version 2.6.
What cards do I have in my hand with Nano Banana 2 Lite?
Google has provided very few details—only three: the latency for image generation via text-based commands is 4 seconds; the cost per image at a resolution of 1K is 0.034 US dollars. These features are already available on the AI Studio, Gemini API, and Gemini Enterprise Edition Agent platforms, and will soon be added to the Search AI Mode and Gemini App as well. According to Google’s usual approach regarding the “Lite” version, sacrificing speed and cost inevitably means compromising image quality and the ability to process more complex instructions. After all, this version was never designed to be used for creating high-quality poster visuals in the first place.
Third-party quotes are currently in a state of chaos. Some reselling platforms are offering prices of $0.02 per token, while comparison websites show token prices that differ by an order of magnitude from those offered by other sources. I will only accept the price of $0.034 mentioned in Google’s official announcement; I will wait until the official pricing page stabilizes before considering any other figures.
Place the two bills one on top of the other.
| A square chart with a size of approximately 1K per sheet. | Price (US dollars) | speed |
|---|---|---|
| GPT Image 2.5 Flare low，走 Batch | Approximately 0.003 | Batch processing; no time constraints apply. |
| GPT Image 2.5: Flare effect is reduced, and real-time processing is available. | 0.0061 | 11.8 seconds |
| GPT Image 2.6 Flare low | 0.0250 | 5.9 seconds |
| Nano Banana 2 Lite | 0.034 | About 4 seconds |
| GPT Image 2.6 Flare medium | 0.0437 | 5.9 seconds |
The most counterintuitive aspect of this situation is that the Lite version is actually over 30% more expensive than the low-end model in the 2.6 series.
The name “Lite” doesn’t necessarily mean that it is the cheapest option in this comparison. What really gives it an advantage is its 2.6 medium configuration: it is over 20% cheaper and also runs about 2 seconds faster. If you were previously using the medium configuration, switching to Lite will definitely save you money; however, if you were already on the low configuration, changing to Lite would actually mean paying more. As for the 2.5 configuration that uses batch processing, neither of these new models covers it. As long as your business can accept results being generated the next day, the most cost-effective option is to stick with your current setup for now.
What is the value of that speed in 2 seconds?
OpenAI claims that the speed of its 2.6 version has increased by approximately 50%, and actual tests have shown that Flare performs even better: the response time for low-capacity scenarios has been reduced from 11.8 seconds to 5.9 seconds, while for high-capacity scenarios, it has decreased from 17.2 seconds to 7.9 seconds.
Sunburst is another story. In the same set of tests, it took approximately 17% more time to generate the initial data images compared to its predecessor, and about 27% more time to produce the mascot images. The comparison of two versions that each contained three static images was even more unfavorable: Version 2.5 took a total of 39.75 seconds to generate these images, costing approximately 0.045 US dollars; whereas Version 2.6 required 48.74 seconds, resulting in costs of around 0.125 US dollars. There was no obvious difference in quality between the two versions. Both sets of test results were obtained using a single laptop on the night of release, so it was impossible to determine the actual server load under real-world conditions. In my opinion, for scenarios where users are closely monitoring the progress bar while interacting with chatbots, a difference of 4 seconds or 6 seconds is definitely noticeable, and in such cases, Lite has an obvious advantage. However, in background batch processing scenarios, this 2-second difference is completely insignificant.
The places where both sides crashed separately
The problem with version 2.6 lay in the fact that “it seemingly looked correct”. Someone used version 2.5 and version 2.6 to process a set of e-commerce materials respectively. In all 6 static images, the specified text in 24 locations was accurately displayed, and the dimensions were also correct. However, the actual design of the product had changed: in version 2.5, the side handles were mistakenly drawn as top handles, while in version 2.6, this was reversed. Additionally, a white mark that no one actually needed appeared on the badge. Such errors did not trigger any warning systems, nor could they be detected by text verification tools; they could only be identified by manually comparing the images with the original design. It is best not to use SVG files for production purposes until these issues have been resolved. On the night of the release, someone made 15 requests, and 14 of them resulted in a 502 error code, meaning that almost all the prompts containing text failed to display correctly. Version 2.6 also did not include date stamps, so any changes in product design would only be noticed once they actually took effect.
At present, there are no third-party measurements available to evaluate the weaknesses of Lite, and this fact in itself represents a risk.
Google didn’t provide any scores regarding image quality or text rendering; it only provided data on latency and cost per unit. Given that the Lite version is positioned as having “lower reasoning capabilities,” I would assume that it would fail to handle certain multi-condition instructions. For example, in a command like “Change the background to wood grain, but keep the color of the bottle unchanged,” the latter part of the instruction would likely cause issues. Before launching the product, I would test it by making five consecutive changes to its design using images from our own products—changing only one element each time—and see at which round the design would start to deviate from the original intent. This approach would be far more reliable than any promotional material.
Something that only one company is capable of doing.
2.6 It is possible to generate GIFs that can be used directly within loops. By default, the API uses 9 frames per second, with each frame lasting 200 milliseconds; therefore, a sequence of 1.8 seconds would consist of 16 frames in total. The cost is relatively high: approximately $0.056 for 12 frames in the “medium” quality setting, and around $0.108 for 16 frames in the “high” quality setting. This option is sufficient for creating emoji or mascot animations, although there may be a slight noticeable transition at the start and end of the loop.
Both approaches of marking and editing seem feasible, but the price has increased with version 2.6: For selecting a specific area to modify within the model, the cost was approximately 0.022 US dollars for version 2.5 and has risen to around 0.052 US dollars for version 2.6.
The unique advantage of Lite lies in its integration with various channels. It is compatible with both the search AI Mode and the Gemini App, which means that ordinary users can easily generate images by simply using the search bar without having any need to understand APIs. For those involved in content distribution, this feature is far more significant than a small difference in price per unit.
My way of categorizing it
For a large number of small images that can be processed overnight: keep the setting at 2.5 and use batch processing; neither of the two new models needs to be adjusted in this case.
It turns out that using the 2.5 medium setting for real-time rendering used to be the recommended approach. The Lite option deserves a thorough evaluation—it’s inexpensive and fast, but the image quality must meet your requirements.
Those that require multiple rounds of editing and have high demands for product consistency: When comparing Sunburst in version 2.6 with that in version 2.5, Lite should not be considered at this time.
If you want an animated version, there’s no other option available; only version 2.6 is available.
For teams that don’t write any code and still generate dozens or even hundreds of documents in just one month, such tasks can basically be ignored – you can simply use tools like ChatGPT or Gemini App to complete them.
Relevant comparisons: “Canva Craft vs. GPT-Image-2.5: Although Craft ranks fourth, its own API documentation recommends using other models for images that contain text.” “FLUX 3 Image Klein vs. Qwen-Image-2.1: The version claimed to be capable of processing 8GB of data actually has a text encoder alone that occupies 8.88GB of memory.”