GPT Image 2.5 at $8 and $30 per million tokens vs Sume's cost

A launch report on AI Weekly (read 2026-10-07) prices both GPT Image 2.5 API models at $8.00 per million image input tokens and $30.00 per million image output tokens. Sume does not bill by token for images. It meters per image, sets usage.cost to the USD amount charged to your wallet, and always returns 0 for the token counts, so budget from the cost field, not from tokens.

The token prices on the page

The launch report gives the same prices for Flare and Sunburst. Sume's docs add the input-text rate from the provider rate card they cite: $5 per million input text tokens. They also say output estimates use OpenAI's ChatGPT Image 2.5 size and quality calculator, and that the provider rounds the total up to $0.0001.

What Sume shows instead

The endpoint record for each model has a pricing array with a cost_usd per billable line, and that amount already includes Sume's pricing. The response then reports usage.cost, the billed USD amount, with prompt_tokens, completion_tokens and total_tokens fixed at 0. The docs state that per-token usage data is not available yet.

How quality changes the reserve

Quality is the main spend lever. The docs list auto, low, medium, high, xhigh and max for GPT Image 2.5, and the default is high if you omit it. Quality auto reserves max, and auto size reserves the upper bound of output tokens. So if you want predictable reserves on a batch, pass explicit quality and an explicit size rather than auto.

curl "https://api.sume.com/v1/images/models/openai/gpt-image-2.5/endpoints" \
  -H "Authorization: Bearer $SUME_API_KEY"

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":"flat-lay of a leather wallet, soft window light","quality":"medium","aspect_ratio":"1:1"}'

Practical budgeting

Run five images at the quality you plan to ship, sum the usage.cost values, and multiply. Failed or cancelled generations are not charged, and a completed one is billed in full. Because you pay cost_usd × n, a four-image call costs four times the single-image line.

Budget checklist

To keep a batch inside a budget:

• Pass quality and an explicit size instead of auto.
• Multiply the cost_usd line by n for each call.
• Sum usage.cost from every response and compare it with your estimate.
• Do not compute spend from token counts; Sume always returns 0 there.