Image route guide · checked Nov 13, 2026

GPT Image 2 pricing per image and per million tokens

GPT Image 2 is mostly sold per image, and one Kie.ai tier is sold per million tokens. The tier, resolution, and route set the actual rate, so keep token-billed rows out of a per-image comparison.

Short answer: GPT Image 2 lists $0.18 per million tokens, not per image, for Kie.ai's gpt-image-2-6-flare. Per-image rows: $0.047 on Replicate, $0.053 on fal and Together AI (medium 1024x1024, checked 2026-08-22), and $0.025/$0.03 on Kie.ai economy and 2.5 flare tiers.

GPT Image 2 rates by provider and unit: APIDir accepted observations from August 2026 and provider page rates retrieved November 13, 2026. Per-image and per-token rows are not directly comparable.
RouteRateUnitSource date
gpt-image-2 on replicatemedium 1024x1024$0.047per imageAPIDir accepted observation, 2026-08-22.
gpt-image-2 on falmedium 1024x1024$0.053per imageAPIDir accepted observation, 2026-08-22.
gpt-image-2 on together-aiserverless default$0.053per imageAPIDir accepted observation, 2026-08-22.
gpt-image-2-6-economy on kie-aieconomy variant$0.025per imageKie.ai pricing page (5 credits per image), retrieved 2026-11-13.
gpt-image-2-5-flare on kie-aiflare variant$0.03per imageKie.ai pricing page (6 credits per image), retrieved 2026-11-13.
gpt-image-2-6-flare on kie-aiflare variant$0.18per 1M tokensKie.ai pricing page (36 credits per 1M tokens), retrieved 2026-11-13. Per-image cost unavailable from this page.

Per-image routes, plus one token-billed tier

The accepted observations cluster tightly: $0.047 on Replicate and $0.053 on fal and Together AI for a medium 1024x1024 image. Kie.ai's per-image tiers retrieved in November are lower: $0.025 for the economy variant and $0.03 for the 2.5 flare variant. Those five rows are the ones you can compare per image.

Kie.ai also lists gpt-image-2-6-flare at $0.18 per million tokens, 36 credits on the same meter. That is not an $0.18 image. A per-image figure needs the token count the route reports for your resolution, which the page does not publish, so this guide leaves that row as a token rate and keeps it out of the per-image budget. Measure one request at your production settings before you convert it.

Edits and variations have their own meters

Image routes often expose text-to-image and edit endpoints under the same model. Edits may be billed per image, per input, or per output unit, and a single edit call can produce multiple outputs. Read the endpoint's meter rather than assuming the text-to-image rate applies. A workflow that chains three edits per shipped image costs three times the per-image rate before retries.

Count outputs, not prompts. A request that returns two variations is two billable images on a per-image route. If your UI shows four options per request, the bill is four images whether or not the user keeps them.

Resolution and quality settings are the other hidden multiplier. A route that prices per image may still charge different rates for different output sizes or detail levels, and the catalog row you read may describe only the default. Capture the exact configuration in the same record as the rate, then run one measured request at your production settings and compare the charge with the estimate.

  • Separate text-to-image and edit endpoints in the budget; they can use different meters.
  • Count every returned image, not every prompt, when the UI offers variations.
  • Record the resolution and quality tier beside the rate; a default-configuration price is not your price.
  • Recheck failed-generation billing before relying on a retry-heavy workflow.

Retries and acceptance rates

Image review is subjective and rejection rates are high for brand work: hands, text, layout, and unwanted artifacts cause rework. A 75% acceptance rate needs 1.33 attempts per accepted image. At $0.047 that is $0.063 per accepted image; at $0.03 it is $0.04. The multiplier is the same, but the absolute difference is what decides whether the premium tier is viable.

Track acceptance per tier. A cheaper tier with a lower acceptance rate can cost more per accepted image once retries are counted. That is the number to compare, not the sticker price.

Storage and delivery deserve their own line once volume is real. Ten thousand outputs at a few hundred kilobytes each is a manageable object-store footprint, but signed URLs, transforms, and egress can add a recurring cost that outlives the generation batch. Price them per accepted image and add the result to the unit cost, so the comparison reflects the product rather than the API call alone.

Worked estimate: 10,000 images

Ten thousand accepted images at an 80% acceptance rate requires 12,500 billable attempts. At $0.025 that is $312.50; at $0.03 it is $375.00; at $0.047 it is $587.50; at $0.053 it is $662.50. The spread across per-image routes is the entire budget conversation, and none of it depends on prompt engineering. The token-billed flare tier cannot enter this table until you have a measured token count per image.

Add storage and delivery if the product serves the images. Those costs are usually small relative to generation at these rates, but they are not zero at ten thousand outputs, and they belong in the same monthly line so the unit economics are honest.

What to verify before a batch

Confirm four facts for the route you will call: the tier name, the output resolution, the billing unit, and whether failed generations are charged. A per-million-token row needs a measured token count before it can become a per-image budget. Then run a small batch and reconcile attempts against the invoice. A rate that cannot be reproduced on an invoice is a marketing number.

Keep the retrieved date with the rate. The accepted rows in this guide are from August 2026 and the provider rows from November 2026; the movement between them is a reminder that image pricing changes with model versions and provider tiers. Recheck before each campaign rather than at annual planning.

Estimate a GPT Image 2 batch

Enter the per-image price for a per-image tier, your measured acceptance rate, and any fixed monthly cost. A per-million-token tier needs a measured token count per image and is not entered here.

Estimate a GPT Image 2 batch

Enter the per-image price for a per-image tier, your measured acceptance rate, and any fixed monthly cost. A per-million-token tier needs a measured token count per image and is not entered here.

Images that pass review and ship.
One billable image from a per-image tier. Do not enter a per-million-token rate here.
Accepted images divided by all billable attempts.
Storage, delivery, or review tooling.
Reset example

Planning result

Review the example inputs, replace each rate with the exact current value from the linked source, then calculate. APIDir does not send these values to a provider.

Questions developers ask

How much does GPT Image 2 cost per image?

APIDir accepted $0.047 per image on Replicate and $0.053 on fal and Together AI at medium 1024x1024 in August 2026. Kie.ai listed per-image tiers at $0.025 and $0.03 in November 2026, and bills its gpt-image-2-6-flare tier at $0.18 per million tokens, not per image.

Why do tiers differ so much?

Economy and flare variants trade detail, speed, or output quality for price, and one tier switches from a per-image meter to a per-token meter. A product that needs consistent high-detail output must budget the tier it can actually measure.

Are image edits billed separately?

Often yes. Check the endpoint's meter; an edit call may be billed per image, per input, or per output, and a multi-output request can multiply the charge.

Do failed image generations cost money?

Confirm with the provider's terms and an invoice. Low-quality but successful generations are normally billable, which is why acceptance rate belongs in the estimate.

Sources and check dates

Every rate in this guide was retrieved from the linked page on Nov 13, 2026. Prices can change; recheck the source before purchasing.