ChatGPT Image 2 vs 2.5 at medium quality, 16:9: price on Sume

A 16:9 image at medium quality costs $0.0105 billed on ChatGPT Image 2.5 and $0.0525 on ChatGPT Image 2 on Sume, so Image 2 is five times the price for the same ratio and tier. The list prices are $0.0084 and $0.042; Sume's billed amount is list times 1.25. If you are still sending openai/gpt-image-2 for widescreen thumbnails and slides, changing the model id to openai/gpt-image-2.5 is the single biggest saving on this route, and the request shape does not change.

Both models list the same nine aspect-ratio values and the same low, medium and high tiers, so a swap does not trip unsupported_parameter. Image 2.5 adds xhigh, max and auto on top.

Per-image and per-batch numbers

The 16:9 medium row is the one this post is about. Scale it to your own volume: the right-hand column is what you keep by moving the id.

Why the gap exists

ChatGPT Image 2 is priced from the provider's own per-image card. ChatGPT Image 2.5 is billed from output image tokens at $30 per million, rounded up to $0.0001, and the calculator scales the token count with the image's short side. A 16:9 image has a shorter short side than a 1:1 image of similar area, so it costs less. That is why the Image 2.5 widescreen number is lower than its square number, while Image 2's figure does not drop the same way.

The practical effect is that Image 2.5 gets cheaper as the image gets wider. For the full ladder at one ratio, see the stored posts on quality price by tier; this one covers only the medium tier.

Rounding is also worth knowing. Image 2.5's list figure is rounded up to the next $0.0001 before the 1.25 multiplier, so very small images can sit a hair above the raw token arithmetic, while Image 2's per-image card has no such step.

The swap

Change the model id and nothing else. The quality value medium is valid on both. Keep sending it explicitly: if you leave quality off, Sume's default is high, which costs much more than medium on both models.

Run a side-by-side before switching a customer-facing pipeline. The prices say nothing about whether Image 2.5 output is right for your brand; ten prompts on each cost $0.525 and $0.105 and take a few minutes.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Widescreen title slide, abstract teal gradient, no text",
    "aspect_ratio": "16:9",
    "quality": "medium"
  }'

A cost-check script

If you want the live numbers instead of the ones in this post, read each model's /endpoints record and compare cost_usd. The record lists the billable lines for the model; for the token-billed Image 2.5 it is built from the same token math, so the figure it returns for your request shape is the one to trust. Record it once a week if you run a large batch, because price tables change.

A simple guard is to keep the expected cost per image in config and alert when the usage.cost on a finished job is more than 10 percent above it. That catches a silently changed quality default, which is the most common reason a bill is higher than the plan.

When Image 2 is still the right row

Spending the same budget on Image 2.5 buys five times the images at this tier. A team with $50 of balance gets 952 Image 2 images at 16:9 medium, or 4,761 on Image 2.5. That is a throughput argument for draft and slide work, and it is separate from any view on which model's output is better.

Keep Image 2 if you have approved its output in production and you are not ready to re-approve. The saving is real, but a change in model is a change in look, and for a brand that has signed off a style that may cost more than the $0.042 per image.