Guide
GPT Image 2.5 with OpenClaw
Choose Sunburst or Flare, pick how to pay, and ask your agent to create an image. This guide covers the OpenClaw Launch integration and its current limits.
Updated October 10, 2026. OpenAI announced Images 2.5 for ChatGPT, ChatGPT Work and Codex across all tiers on September 8 in its launch announcement. API details were checked against the OpenAI Sunburst documentation, Flare documentation and OpenRouter image catalog.
Sunburst or Flare?
| Model | Choose it for | OpenAI API ID |
|---|---|---|
| GPT Image 2.5 Flare | Faster generation and iteration | gpt-image-2.5-flare |
| GPT Image 2.5 Sunburst | Edits where preserving details matters | gpt-image-2.5-sunburst |
Both accept text and image inputs. The image model is separate from your agent’s chat model: select it in Image, rather than the text model menu.
Option 1: Use OpenRouter credits
- Open an OpenClaw instance on the dashboard.
- Open the Image model menu and select OpenRouter.
- Choose GPT Image 2.5 Flare (Low) for an initial draft, or a Sunburst tier. Low, Medium and High are separate choices for each model.
- Use your available OpenRouter balance, or add your own OpenRouter key on the API keys page.
- Ask your bot: “Create a square launch poster for a fictional plant shop called Fern. Use cream and forest green, with the headline ‘Room to Grow’.”
The proxy uses OpenRouter’s dedicated Images API. Pasting a bare 2.5 ID into an OpenRouter chat model field will not configure this route. Your selected quality is fixed by the tier, so asking for “max quality” in a prompt does not change a Low selection.
Option 2: Bring your OpenAI API key
- Create a key in the OpenAI API dashboard with API billing and access to the image model.
- Save it under OpenAI on the API keys page.
- In your OpenClaw instance’s Image menu, open OpenAI · Your key and choose Sunburst or Flare.
- Generate one small draft before a larger batch. This route bills your OpenAI account.
If the app reports an Agnes video conflict, choose another video model or use the OpenRouter image route. These direct OpenAI images share a provider slot with Agnes video in the current runtime.
What your ChatGPT subscription gives you
OpenAI says Images 2.5 is available to ChatGPT and Codex users on every plan. The subscription image route that agents sign in to does not let the caller choose a model: it ignores the requested model, size and quality (OpenClaw #148527), and its internal label still reads gpt-image-2-codex, which is why it was long described as GPT Image 2.
We tested it on October 10, 2026 through an OpenClaw bot and a Hermes bot, each signed in to a ChatGPT plan, and sent the same prompts to GPT Image 2.5 Flare and GPT Image 2 through the API. Every subscription image carried the same provenance stamp as GPT Image 2.5 Flare (GPT Image 2 carries a different one), and the images matched Flare’s output side by side. The subscription row in the Image menu is therefore shown as GPT Image 2.5 (ChatGPT plan).
What you cannot do there is choose. OpenAI picks the variant and the quality, and the xhigh and max levels cannot be selected. To pick Flare or Sunburst and a quality tier, use OpenAI · Your key or OpenRouter. If you selected GPT Image 2.5 under OpenAI · Subscription before September 28, 2026, nothing changed: it keeps generating with Images 2.5, billed to your plan.
Pricing and quality
For API billing, both models list $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. Actual cost depends on the request; an edit also includes its reference images. The API panes show “Variable”; hover over the API price to see the token rates. A fixed per-image estimate is not yet confirmed. See the linked model pages for current rates.
Do not carry over GPT Image 2 per-image estimates: OpenAI explicitly says its older calculator does not estimate 2.5 token use. Start at Low and increase quality when the result needs it. The API supports additional quality levels, but the OpenRouter picker currently offers Low, Medium and High.
Editing and current limits
For an edit, attach a reference and describe what should stay: “Keep the product shape, label and camera angle. Replace only the background with a warm studio wall.” Use direct OpenAI for reference edits on older OpenClaw runtimes; OpenRouter reference edits require a runtime that sends multipart image uploads.
Our OpenRouter proxy forwards aspect ratio, up to four outputs and up to five reference images. It does not expose every upstream control, including exact pixel dimensions or transparent backgrounds. Direct API capabilities are described in the OpenAI image generation guide.
If access fails, check API billing, model permissions and any required OpenAI organization verification. If generation is slow, wait for the current task before submitting it again. A provider listing establishes availability; account access and successful generation still depend on the provider.
Frequently asked questions
Is GPT Image 2.5 released?
Yes. OpenAI lists Sunburst and Flare, with snapshots dated September 8, 2026. There is no single model ID called gpt-image-2.5 in the documented pair.
Does this use my ChatGPT subscription?
Partly. The ChatGPT plan image option renders with Images 2.5, but OpenAI chooses the variant and quality there, so you cannot pick Flare or Sunburst. We confirmed this on October 10, 2026 by comparing subscription images with GPT Image 2.5 Flare and GPT Image 2 from the API. To choose a variant and a quality tier, use OpenRouter credits or your own OpenAI API key.
Can I use it on Hermes Agent?
Yes. To choose Flare or Sunburst, use your own OpenAI API key: save the key, then select Flare or Sunburst at Low, Medium or High under OpenAI · Your key in the Image menu. This needs Hermes v2026.9.14 or newer; an older bot shows an update message instead. A Hermes bot signed in to a ChatGPT plan also gets Images 2.5 from its GPT Image 2.5 (ChatGPT plan) option, without a choice of variant or quality.
Does the free trial include GPT Image 2.5?
Yes. Since October 4, 2026 the free GPT Image images on a trial are generated with GPT Image 2.5 Flare at Low quality, on OpenClaw and Hermes. The allowance is unchanged at 2 images. Medium, High and Sunburst need OpenRouter credits or your own OpenAI key.