What is gpt-image-2.5 pricing, and should you pick Sunburst or Flare?

What are gpt-image-2.5 Sunburst and Flare?

gpt-image-2.5-sunburst and gpt-image-2.5-flare are OpenAI's image generation models, released to the API on September 8, 2026 (per the OpenAI API changelog). The same day, the consumer ChatGPT Images 2.5 update began rolling out to ChatGPT, ChatGPT Work, and Codex users across all tiers, according to the launch announcement quoted in the Hacker News discussion.

Both models accept text and image inputs and produce image outputs. You call them directly on the Image API endpoints v1/images/generations and v1/images/edits, or through the Responses API image generation tool, which adds multi-turn editing and file-based image inputs. The default snapshots are gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08, and both list inpainting as a supported feature (model pages, OpenAI docs).

They replace the older gpt-image-2 line at the top of OpenAI's image stack. OpenAI removed the legacy DALL-E 2 and DALL-E 3 model snapshots from the API on May 12, 2026, so the gpt-image family is the only image generation path left.

gpt-image-2.5 pricing: the exact per-token rates

OpenAI bills GPT Image 2.5 by tokens, not per image. The rates below appear on OpenAI's API pricing page (image generation section) and are repeated on each model page, so the ledger is double-checked:

• Image input tokens: $8 per million, $2 per million when cached.
• Image output tokens: $30 per million.
• Text input tokens (the prompt): $5 per million, $1.25 per million when cached.
• Text output: not billed, because both models output images only.

Sunburst and Flare share these rates, and OpenAI states GPT Image 2.5 uses GPT Image 2 token rates, so the per-token price did not change from the previous generation. What changed is capability: the 2.5 models add xhigh and max quality settings and support custom sizes up to 4K.

Two billing details are easy to miss. First, there is no fixed per-image price and no batch discount tier: OpenAI's model pages list batch processing as not supported for both image models, so plan for the real-time $30 per million output rate. Second, Responses API calls bill the mainline model's text tokens on top of the image cost, so a gpt-image-2.5 generation inside a conversation with a model like GPT-6 Astra carries two rate cards.

Sunburst vs Flare: which model should you use?

OpenAI positions the pair by workload, not by price. gpt-image-2.5-sunburst is described as the most capable model for image generation and editing, and the docs point to it for workflows where editing precision matters most, such as product shots, reference-based rework, and controlled image-to-image edits. gpt-image-2.5-flare is the fastest model for high-quality everyday generation, the default pick for volume work like social assets and quick illustrations where raw speed beats fine control.

Because the token rates are identical, the choice comes down to the job: pick Sunburst when the edit has to preserve the source image, pick Flare when you are generating fresh images at volume and want speed. Early community signals on the launch thread point the same way. One developer who says he generated roughly 50,000 images on gpt-image-2 reports average latency falling from about 104 seconds to 35 to 40 seconds with the 2.5 models. Benchmarks quoted in the same thread put both models at the top of the LM Arena text-to-image leaderboard (Sunburst 1421, Flare 1399, versus 1381 for gpt-image-2). Treat those as developer reports, not OpenAI claims.

Integration choice matters too. The Image API handles a single generation or edit from one prompt. The Responses API image tool is the route for conversational, multi-step editing, and OpenAI's docs say gpt-5 and newer mainline models support calling it.

Quality, size, and output settings that move the bill

Since billing is token-based, the settings you pass control the cost of each image. Both 2.5 models accept quality values low, medium, high, xhigh, max, and auto (auto is the default). xhigh and max are new in this generation: earlier GPT Image models stop at high. OpenAI's guide suggests low for quick drafts and higher settings for final assets, because higher quality consumes more output tokens.

Size is the other lever. Recommended dimensions are 1024x1024, 1536x1024, and 1024x1536, and both models accept custom WIDTHxHEIGHT strings where each edge is a multiple of 16, the aspect ratio stays between 1:3 and 3:1, no edge exceeds 3840 pixels, and total pixels land between 655,360 and 8,294,400 (4K). Resolutions above 2560x1440 are experimental per the docs.

Output controls: png is the default format, jpeg and webp are available, and jpeg renders faster when latency matters. Compression runs 0 to 100% for jpeg and webp. Transparent backgrounds work when you set background to transparent and request png or webp output. Finally, latency: complex prompts can take up to 2 minutes to process, and if you stream generation with partial images, each partial image adds 100 image output tokens, roughly $0.003 at the $30 per million rate.

Rate limits, org verification, and moderation

Image generation is rate-limited per usage tier. OpenAI's model pages list the default limits as 100,000 tokens per minute and 5 images per minute at Tier 1, rising to 8 million tokens per minute and 250 images per minute at Tier 5. A Tier 1 account generating at volume will hit the 5 images per minute ceiling long before token cost becomes the constraint.

Access carries two gates. OpenAI's docs state that API Organization Verification may be required before using GPT Image models, and image requests are filtered by a moderation parameter with two settings: auto (standard filtering, the default) and low (less restrictive). Blocked requests return an image_generation_user_error with a moderation_blocked code and coarse category labels such as harassment, self-harm, sexual, or violence, so handling those errors programmatically is straightforward.

Known limitations and the alternatives to weigh

OpenAI's image generation guide is candid about four limits. Text rendering is significantly improved but can still struggle with precise text placement and clarity. Visual consistency for recurring characters or brand elements can break across separate generations. Composition control weakens in structured or layout-sensitive scenes. And complex prompts can take up to 2 minutes, which matters if your product shows a spinner while users wait.

The alternatives depend on your buying context. If you only need occasional images inside a chat interface, ChatGPT plan access may beat API spend entirely, and our ChatGPT Plus vs Pro breakdown covers those tiers. If you are building an app, the older gpt-image-2 remains available at the same token rates for teams that do not need the new quality settings, and OpenAI's own coverage of ChatGPT Work tools lists where image generation sits inside its agentic products. Outside OpenAI, Google's Imagen family and open-weight FLUX models are the usual cross-vendor points of comparison, but their rate cards are outside OpenAI's docs, so compare them against your own workload before committing.

At a glance

GPT Image 2.5 model comparison (verified on OpenAI docs, September 9, 2026)

FAQ

Is gpt-image-2.5 available through the OpenAI API?

Yes. OpenAI released gpt-image-2.5-sunburst and gpt-image-2.5-flare on September 8, 2026. Both are callable on the v1/images/generations and v1/images/edits endpoints and through the Responses API image generation tool, which gpt-5 and newer mainline models can invoke.

What is the difference between gpt-image-2.5 Sunburst and Flare?

Sunburst is OpenAI's most capable image model for generation and editing and is recommended when editing precision matters. Flare is the fastest model for high-quality everyday generation. Token rates are identical, so the choice is workload-based, not cost-based.

How much does one gpt-image-2.5 image cost?

OpenAI does not publish a fixed per-image price. Image output bills at $30 per million tokens and image input at $8 per million, and token consumption scales with size and quality settings, so cost per image varies. Streamed partial images add 100 image output tokens each, about $0.003.

Related reading

ChatGPT Plus vs Pro pricing and limits, gpt-6-astra API pricing, ChatGPT Work tools inventory

Sources

• OpenAI API pricing (image generation section)
• OpenAI image generation guide
• GPT Image 2.5 Sunburst model page
• GPT Image 2.5 Flare model page
• OpenAI API changelog (September 8, 2026)
• Hacker News launch discussion