GPT Image 2.5 Cost per Image

GPT Image 2.5 cost per image is not a fixed fee: Sunburst and Flare both charge $8 per million image input tokens, $2 cached, $30 image output, $5 text input, and $1.25 cached text input. A defensible per-image number comes from each response's usage plus every billable edit, partial image, and retry, not an older GPT Image 2 lookup table.

GPT Image 2.5 Cost per Image: The Formula

The useful price is total billable token usage divided by accepted images. OpenAI GPT Image 2.5 Sunburst and Flare share the same published rates, so neither model is automatically cheaper per image. Their token consumption, settings, and number of attempts decide the final bill.

These prices were verified against OpenAI's live API pricing page and both live model pages on September 9, 2026.

Text output has no separate charge because these models output images. For a direct Image API request, calculate:

Request cost = standard text input + cached text input + standard image input + cached image input + image output.

Multiply each token count by its matching per-million rate. An original text-to-image request normally has no image-input term. An edit has one because the reference image must be processed. If the Responses API orchestrates the image call, add the mainline model's token usage too. OpenAI's image generation guide explicitly says Responses requests carry that extra usage.

The response's usage is the source of truth. Do not label a repeated prompt or reference as cached unless the reported usage does. Do not divide spend by requests if the team rejects half the outputs. A creative workflow buys accepted assets, not API calls.

Worked Per-Image Examples With Explicit Assumptions

A modeled example is useful only when its assumptions stay visible. None of the token counts below is presented as typical Sunburst or Flare usage. Replace them with the values from your own responses.

Fresh generation. Assume one request uses 200 uncached text tokens and 2,000 image output tokens, with no image input. Text costs 200 × $5 / 1,000,000 = $0.001. Output costs 2,000 × $30 / 1,000,000 = $0.060. The modeled request totals $0.061.

Reference-image edit. Assume 250 uncached text tokens, 6,000 uncached image input tokens, and 2,000 output tokens. The three components cost $0.00125, $0.048, and $0.060. The modeled edit totals $0.10925.

Repeated edit with reported caching. Keep the same token counts, but assume the response reports all text and image inputs as cached. The input components fall to $0.0003125 and $0.012; output remains $0.060. The modeled repeat totals $0.0723125.

At a thousand accepted assets, those assumptions produce three very different budgets: $61 for one-pass fresh generations, $109.25 for one-pass edits, or $183 when each accepted result takes three billable $0.061 attempts. The right forecasting unit is therefore:

Cost per accepted asset = total cost of all billable attempts / number of outputs that pass review.

That denominator catches the expensive behavior a token-rate table cannot: prompt retries, edit loops, abandoned variants, and streamed previews. A model with a higher average call cost can still be cheaper if it reaches approval in fewer attempts.

1. 1

   Pin the model

   Record the exact model ID. Use gpt-image-2.5-flare-2026-09-08 or gpt-image-2.5-sunburst-2026-09-08 when a stable snapshot matters, rather than silently mixing future revisions into one cost sample.

2. 2

   Fix the output settings

   Choose an explicit size and quality instead of auto during measurement. Both models support low, medium, high, xhigh, and max, and changing the setting can change token consumption.

3. 3

   Log every billable response

   Store model, task type, text input, image input, cached input, image output, partial-image usage, and any mainline-model usage returned by Responses.

4. 4

   Mark the creative outcome

   Tag each result accepted, edited, rerun, or discarded. This turns infrastructure spend into a design-production metric.

5. 5

   Compare accepted assets

   Sum all billable usage for each model and divide by accepted outputs. Promote the model and settings with the lower approved-asset cost that still clears the craft bar.

GPT Image 2.5 Sunburst Pricing

Sunburst has no premium token tariff. It uses the same $8 image input, $2 cached image input, $30 image output, $5 text input, and $1.25 cached text input rates as Flare.

The difference is the job it is built to do. OpenAI describes Sunburst as its most capable image generation and editing model, with editing precision as the priority. Its listed speed is medium. Vercel similarly positions it as the slower option with tighter control for detailed creative work.

That makes Sunburst the rational default for a product-image correction, character-preservation pass, or layout-sensitive revision where another rejected attempt costs more than waiting. It is not automatically the economical choice for bulk ideation. The break-even rule is simple: choose Sunburst when its average call cost multiplied by attempts per accepted asset is lower than Flare's result for the same brief.

GPT Image 2.5 Flare Pricing

Flare costs the same per token as Sunburst, so its economic case is speed and iteration throughput, not a discounted rate. OpenAI calls Flare its fastest model for high-quality everyday image generation and lists its speed as very fast.

Flare is the sharper starting point for mood boards, rapid composition exploration, social variants, and other work where a designer expects to compare several directions. It can also accept image inputs and perform edits. What the pricing page does not promise is fewer tokens or fewer retries for your brief.

Use Flare for the draft lane, then keep it for final output only if the accepted-asset log supports the choice. Moving every final automatically to Sunburst wastes time when Flare already clears review; keeping every preservation-sensitive edit on Flare can waste money when additional attempts erase the speed advantage.

GPT Image 2.5 Image Editing Cost

An edit costs more than its output alone because the source image becomes a billable input. This is where a headline price most often stops matching the invoice.

Vercel's release example makes the workflow concrete. The first request establishes a perfume bottle on pale stone in morning light. The follow-up passes that image to Sunburst, changes the cap to gold, and asks the model to preserve everything else. The second call now has three priced streams: its instruction text, its reference image, and its new image output.

Under the explicit edit assumptions above, the reference contributes $0.048 to a $0.10925 request. If the reference is reported as cached on a later revision, the modeled image-input component falls to $0.012. The output still costs $0.060 in the example, so caching reduces only the input side. It does not make a rerender free.

The craft bar is preservation, not a technically successful response. Check whether the bottle silhouette, label geometry, light direction, crop, and background survive the cap change. If they drift, the next corrective call is another billable attempt. For text in images, inspect every character at final size. For transparent assets, the guide requires PNG or WebP with the background set to transparent.

GPT Image 2.5 returns raster output, meaning a fixed grid of pixels, through PNG, JPEG, or WebP. If the deliverable needs editable vector paths for a logo or production mark, treat the generated result as direction and rebuild it in a vector tool. A clean-looking PNG is ship-ready for many campaign placements; it is mood-board-only for artwork that must remain structurally editable.

GPT Image 2 Pricing per Image Is Not a 2.5 Quote

The older GPT Image 2 pricing per image table cannot be relabeled as Sunburst or Flare pricing. OpenAI's guide lists GPT Image 2 image-output examples from $0.005 to $0.211 across its displayed sizes and low, medium, and high settings, then separately says the earlier-model details do not apply to Sunburst or Flare.

The current Sunburst model page is even more direct: token rates match GPT Image 2, but its calculator does not estimate GPT Image 2.5 token consumption. Same unit prices do not prove the models spend the same number of units.

There is one confirmed older-model discount worth separating cleanly. OpenAI's Batch column currently lists GPT Image 2 at $4 image input, $1 cached image input, $15 image output, $2.50 text input, and $0.625 cached text input per million tokens. If its consumption matched the modeled 200-text-token and 2,000-output-token request, the arithmetic would be $0.0305. That is a hypothetical older GPT Image 2 comparison, not a Sunburst or Flare price.

The 2.5 model pages list the Batch endpoint, but OpenAI's live pricing page does not publish 2.5 Batch rates. A forecast that silently halves Sunburst or Flare rates is not verified and should not ship.

GPT Image 2.5 API Pricing: Hidden Costs

The base token rates are only the first line of the invoice. Six operational choices move cost per accepted image.

Responses orchestration. When a mainline model decides when or how to call image generation, its own input and output usage is charged in addition to the image model. Use the direct Image API when a fixed generation step needs no conversational planning. Use Responses when that orchestration is worth its separate cost.

Automatic settings. Both models default to auto quality, and size can also be automatic. That is useful for general output, but weak for a controlled price comparison. The documented recommended sizes are 1024x1024, 1536x1024, and 1024x1536. Set size and quality explicitly while measuring.

Streaming previews. The API can return zero to three partial images. Each partial image received adds 100 image output tokens, which is $0.003 at the current output rate. Three received previews add $0.009 before the final image's other usage. A faster-feeling interface has a small but real preview bill.

Billable retries. Three attempts at the modeled $0.061 request produce one $0.183 accepted asset. Count only calls whose telemetry reports usage, but count every such call even when the creative team discards its image.

Prepaid credits. OpenAI's prepaid billing guidance lists a $5 minimum initial purchase, a $10 default, and auto-reload enabled by default during setup. Purchased credits expire after one year and are non-refundable. There is no self-serve annual plan discount to offset that expiry. Manual purchases also do not count toward the optional auto-recharge limit.

Balance and throughput limits. A prepaid balance is not an instantaneous hard stop. OpenAI says delayed usage can create a negative balance that comes out of the next purchase. Separately, both 2.5 model pages mark the Free tier unsupported and publish the same throughput ladder: Tier 1 at 100,000 tokens and 5 images per minute; Tier 2 at 250,000 and 20; Tier 3 at 800,000 and 50; Tier 4 at 3,000,000 and 150; Tier 5 at 8,000,000 and 250. A positive balance does not override those limits.

OpenAI API Pricing vs Vercel AI Gateway

Vercel AI Gateway does not lower the base Sunburst or Flare token rates. Vercel says it reflects provider pricing with no markup and charges no inference platform fee, including when you bring your own OpenAI key.

The choice is operational. Direct OpenAI access has fewer layers and exposes the authoritative model response. Vercel AI Gateway adds spend tracking, request traces, budgets, routing, failover, and configurable retries behind its openai/gpt-image-2.5-flare and openai/gpt-image-2.5-sunburst names.

That observability can improve the denominator by showing which prompts and routes lead to accepted assets. It can also hide waste if automatic retries are not included in the asset-level cost log. Choose the gateway for control and traces, not for an imagined token discount.

Consumer access is a different product with different limits. Keep it outside the API spreadsheet; the separate ChatGPT pricing guide covers that decision.

Which Model Should a Design Team Choose?

Start with Flare for fast exploration and Sunburst for preservation-sensitive or detail-critical edits, then let accepted-asset cost overturn that default when the data says otherwise.

For a social team generating many loose visual directions, time-to-first-options matters and Flare's speed positioning fits the job. For a brand team correcting one element inside an approved composition, Sunburst's tighter editing control is the relevant value. A studio working on consistent characters or product scenes should compare both on the exact reference-heavy brief, not on unrelated sample prompts.

The ship decision still belongs to craft review. Check typography, hands, edge integrity, product geometry, composition drift, and whether the output format matches the downstream edit path. The wider AI image editing model guide is useful when consistency matters more than staying inside one vendor family.

On Monday, pin one model snapshot, choose one explicit size and quality, attach asset IDs to every generation and edit, and start storing response usage beside the review outcome. At the end of the week, divide each model's spend by accepted assets. That number is the price your production plan can defend.

GPT Image 2.5 Pricing FAQ

How much does gpt-image 2 cost per image?

OpenAI's older GPT Image 2 guide lists image-output estimates from $0.005 to $0.211 across the displayed sizes and quality settings, before text or image input. Those figures do not price GPT Image 2.5 Sunburst or Flare.

How much do 1000 tokens cost?

For GPT Image 2.5, 1,000 standard image-input tokens cost $0.008, cached image input costs $0.002, image output costs $0.03, standard text input costs $0.005, and cached text input costs $0.00125.

Is gpt-image 2 unlimited?

No. GPT Image 2 has account-tier token and image-per-minute limits, and usage is metered. GPT Image 2.5 has the same published throughput ladder on its model pages.

Is gpt-image 2 available for free?

No standing Free API tier is supported on the official GPT Image 2 model page. The Sunburst and Flare pages also mark Free as not supported.

How good is GPT image 2?

OpenAI describes GPT Image 2 as a fast, high-quality generation and editing model with flexible sizes and high-fidelity image inputs. GPT Image 2.5 is now the current split: Flare prioritizes speed and Sunburst prioritizes precise editing, so compare accepted outputs on your own brief rather than treating the version number as a quality score.

Is there a 100% free AI image generator?

GPT Image 2.5 is not one. An account may receive promotional credits, but the model's Free rate-limit tier is unsupported and the public product is usage-priced.

Is there a free AI image generator with no limits?

Not in the GPT Image 2.5 API. Even paid access has token and image-per-minute limits tied to the account's usage tier.

Can you legally sell AI-generated art?

OpenAI's Services Agreement says that, between OpenAI and an API customer and to the extent permitted by law, the customer owns the output. The customer still bears responsibility for input rights and use of the output, and outputs may not be unique. That contract language is not a substitute for advice about copyright, trademark, publicity rights, or the rules in your jurisdiction.

Which AI is totally free to use?

There is no durable universal answer because product allowances and limits change. GPT Image 2.5 is not totally free, so evaluate any alternative against its current vendor terms rather than a free-tool list.

Does GPT Image 2.5 have a student discount?

No GPT Image 2.5 API student discount is published. OpenAI's current student offer applies to ChatGPT Plus billing, not API image usage.

What is the GPT Image 2.5 refund policy?

The image model has no separate refund plan. OpenAI says purchased prepaid API credits are non-refundable, expire after one year, and cannot be extended.

Did GPT Image 2.5 change image token prices?

No at the rate-card level: OpenAI says Sunburst and Flare token rates match GPT Image 2. Per-image spend can still change because token consumption can differ by model, quality, size, edit input, and retry count.

Want a cleaner way to audit the workflow around those calls? Get the AI business workflow audit checklist when you join the newsletter.