New model, same price, everything looks great, you'd say, let's upgrade. But nope! There's a new trend now, cheating with output tokens. The price looks the same, but they've found a way to increase the price by increasing output tokens.
Here's how I noticed. OpenAI released GPT Image 2.6 and I told my team to switch everything to it. They came back and said it costs three times more. Same price per token. It uses more tokens: 439 output tokens on 2.5, 1,449 on 2.6, same prompt, 1024x1024 medium.
Anthropic does the same thing too. Claude Fable 5.2 just came out, a fifth cheaper per token than Fable 5.1. I definitely saw a jump in price from Fable 5 to Fable 5.1, and now surely 5.2 will be more expensive too, I'm predicting that. It is: 3.4x. The same MacBook drawing at high effort cost $3.05 on 5.2 and $0.90 on 5.1.
If you don't know how to raise the price, but you need to, AI labs have figured it out.
Check it yourself, it costs about six cents. Run the same prompt on both models and look at usage.output_tokens.
New model, same price, everything looks great, you'd say, let's upgrade. But nope! There's a new trend now, cheating with output tokens. The price looks the same, but they've found a way to increase the price by increasing output tokens.
Here's how I noticed. OpenAI released GPT Image 2.6 and I told my team to switch everything to it. They came back and said it costs three times more. Same price per token. It uses more tokens: 439 output tokens on 2.5, 1,449 on 2.6, same prompt, 1024x1024 medium.
Anthropic does the same thing too. Claude Fable 5.2 just came out, a fifth cheaper per token than Fable 5.1. I definitely saw a jump in price from Fable 5 to Fable 5.1, and now surely 5.2 will be more expensive too, I'm predicting that. It is: 3.4x. The same MacBook drawing at high effort cost $3.05 on 5.2 and $0.90 on 5.1.
If you don't know how to raise the price, but you need to, AI labs have figured it out.
Check it yourself, it costs about six cents. Run the same prompt on both models and look at usage.output_tokens.
This is crazy, and the output quality stayed almost thesame.
I had to zoom into an image to ensure I did not miss any tiny details added to 2.6 that was not present in 2.5.
ScAM Altman