ChatGPT Images 2.5 adds sketch tools; API gets Sunburst and Flare

What changed in ChatGPT Images 2.5, how Sunburst and Flare differ, and why matching API token rates do not establish the cost of a finished image.

Quick answer

OpenAI announced ChatGPT Images 2.5 on September 8, 2026, adding templates, mobile sketch input, image comments, and prompt sharing. Its API documentation distinguishes Sunburst for precision-focused editing from Flare for fast everyday generation. These are provider descriptions, not results from our own image-quality or speed tests.

Download Chat AI Opens the official App Store or Google Play for your device.

New starting points and editing controls in ChatGPT

Templates are under Images → Templates. On mobile, users can type @, choose Sketch, draw a starting idea, and add a description. Opening a generated image full-screen offers editing and commenting controls. Prompt sharing lets another person make a version from the same brief. OpenAI also reports improvements in detail, editing precision, and generation speed without quantifying them in this release note.

Sources: OpenAI

Sunburst and Flare serve different priorities

OpenAI positions Sunburst for tasks where editing precision is the priority, while Flare emphasizes speed for everyday image creation. Both accept text and image inputs and return images. Developers can select either through the Image API or the Responses API image-generation tool. The documented IDs are gpt-image-2.5-sunburst and gpt-image-2.5-flare; each also has a dated snapshot ending in -2026-09-08.

Sources: OpenAI, OpenAI

Token prices are not a per-image quote

Both model pages list $5 per million text-input tokens, $8 per million image-input tokens, and $30 per million image-output tokens. Cached input is listed separately at $1.25 for text and $2 for images. OpenAI says the token rates match GPT Image 2, but its older calculator does not estimate GPT Image 2.5 token consumption. Matching rates therefore do not establish equal final-image costs. API limits depend on the account's usage tier.

Sources: OpenAI, OpenAI

Evaluate the output you need, not just the model label

For a small evaluation, keep the source image and requested change constant. Record whether untouched details survive, whether requested text is correct, elapsed time, and actual billed usage. This is a suggested evaluation method, not a test we have run. Chat AI readers can use the same review checklist with their available image tools; this announcement does not verify GPT Image 2.5 access in Chat AI. Consult the current Chat AI catalog before assuming provider releases are included.

Frequently asked questions

What readers usually ask

Does the new Sketch tool work on desktop?

The September 8 release note describes Sketch on mobile; it does not establish desktop support.

Is Flare proven faster than Sunburst in this article?

No. Flare's speed emphasis comes from OpenAI's documentation. We have not run a controlled comparison.

Can I use the GPT Image 2 calculator for 2.5?

OpenAI explicitly says that calculator does not estimate GPT Image 2.5 token consumption. Check actual usage instead of treating its estimate as a quote.

Evidence

Sources

  1. ChatGPT release notes: September 8OpenAI · Primary source
  2. GPT Image 2.5 Sunburst model documentationOpenAI · Primary source
  3. GPT Image 2.5 Flare model documentationOpenAI · Primary source