GPT Image 2.5: Cost per Accepted Image
A practical guide to GPT Image 2.5 pricing: how to calculate the true cost of an accepted image by accounting for token consumption, revisions, and retries until reaching an approved result, alongside choosing between Flare and Sunburst.
Contents
Final Image Cost and Model Selection
In the official GPT Image 2.5 API, there is no single universal fixed price per image. The real cost equals the cumulative expenses across all generation attempts, paid edits, and discarded variations, divided by the number of accepted outputs.
Choosing the right model:
gpt-image-2.5-flareis a fast model designed for everyday generation and rapid drafting.gpt-image-2.5-sunburstis intended for detailed creative work where editing precision matters; generation takes longer.
As of 2026-09-12, official token rates are identical for both models. However, identical token pricing does not guarantee an equal total project budget: retry counts and acceptance rates can vary depending on the task. No comparative tests for latency, quality, or acceptance rates were conducted for this article.
OpenAI Pricing and Calculating Request Cost from Tokens
OpenAI released GPT Image 2.5 on 2026-09-08 (OpenAI announcement). The API provides two models: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Both models support quality tiers low, medium, high, xhigh, max, and auto (Image Generation guide).
The current Image Generation guide includes a dedicated GPT Image 2.5 calculator. It estimates image-output tokens and image output cost for an explicit model, image size, and quality setting: low, medium, high, xhigh, or max.
The auto setting depends on the generated image itself. Furthermore, the calculator excludes text input, image input, and streaming partial images (each partial image adds 100 image-output tokens).
Therefore, the full request cost and the total cost per accepted image must still be determined using actual usage data, including input tokens, cache, retries, and revisions.
As of 2026-09-12, official OpenAI pricing consists of five components:
| Token Category | Per 1M Tokens | Per 1 Token |
|---|---|---|
| Uncached text input | $5.00 | $0.000005 |
| Cached text input | $1.25 | $0.00000125 |
| Uncached image input | $8.00 | $0.000008 |
| Cached image input | $2.00 | $0.000002 |
| Image output | $30.00 | $0.00003 |
Each token is counted strictly in one category: it cannot be assigned to both cached and uncached categories at the same time.
Formula for the cost of a single request:
C_request = (T_uncached * 5.00 + T_cached * 1.25 + I_uncached * 8.00 + I_cached * 2.00 + O_img * 30.00) / 1 000 000Calculating Cost per Accepted Image
The true unit cost of an approved image is calculated with the following formula:
C_accepted = Total Cost of All Attempts / N_acceptedThe numerator includes expenses from every paid attempt, including initial generations, rejected variations, and paid edit requests. The denominator is the number of approved images. If no images are accepted (N_accepted = 0), the unit cost is mathematically undefined.
Example calculation for a sequence of image edit requests:
| Calculation Stage | Parameters and Calculations | Amount |
|---|---|---|
| Uncached text input | 500 tokens × $5.00 / 1 000 000 | $0.0025 |
| Uncached image input | 5 000 tokens × $8.00 / 1 000 000 | $0.0400 |
| Image output | 2 000 tokens × $30.00 / 1 000 000 | $0.0600 |
| Cost per attempt | $0.0025 + $0.0400 + $0.0600 | $0.1025 |
| Series of 100 attempts | 100 × $0.1025 | $10.25 |
| Cost per accepted image (80% acceptance rate) | $10.25 / 80 approved images | $0.128125 |
Token counts and the acceptance rate are hypothetical illustrative assumptions. Prompt caching was not utilized in this calculation. No paid generations, token measurements, or acceptance evaluations were performed for this article.
Per-Request Billing: The BetterToken Example
As an alternative billing model, consider per-request pricing on the BetterToken platform.
According to the BetterToken catalog on 2026-09-12, gpt-image-2.5-flare and gpt-image-2.5-sunburst are available under the image-generation and openai endpoints. The catalog specifies quota_type = 1, the group GPT Image 4K, and a price of model_price = $0.05 per request. These are dynamic values; always verify current rates on the BetterToken pricing page before launching jobs.
At the current catalog rate of $0.05 per request, 100 paid requests total $5.00:
- 100 paid requests: 100 × $0.05 = $5.00.
- With 80 accepted images, the cost per approved output is: $5.00 / 80 = $0.0625.
No paid requests were performed for this article.
Token-based billing in the OpenAI API and per-request pricing in the BetterToken catalog represent two fundamentally different accounting models that cannot be directly summed or conflated. Differences in calculated numbers do not imply identical generation quality, feature parity, or price permanence.
Running a Test Batch and Tracking Metrics
Before scaling generation workflows, run a controlled test batch on a sample of 10–20 representative tasks. To ensure comparable results, fix the target resolution, keep the quality setting constant (low, medium, high, xhigh, max, or auto), and maintain a consistent number of reference images per request.
Your tracking log should record:
- Model (
gpt-image-2.5-flareorgpt-image-2.5-sunburst),qualitytier, and output resolution. - Number of reference images, actual token consumption from
usageor billed cost. - Result status (accepted, sent for revision, or rejected).
API keys and confidential source assets must never be stored in tracking logs.
Before running a batch, verify current rates at bettertoken.ai/pricing. After completing a small test run, confirm the actual model, request statuses, and deductions in BetterToken Workspace.
