Copilot can now draw: Images from chat, billed on top of your AI credits

GitHub has added image generation and editing to Copilot chat, according to its changelog. The feature, in public preview, runs on GPT Image 2.5 Flare and is billed separately from the chat model.

How the feature works and who gets it

Users describe an image in plain language in any Copilot chat session, whether it is a fresh request or part of a coding task. According to GitHub, the reasoning model you have selected calls a dedicated image-generation tool. The result appears as an inline preview and is saved into the project workspace.

GitHub frames this as removing a detour. Until now, a developer who needed an icon, mockup or illustration had to leave Copilot, explain the request again in another tool, generate the file and bring it back by hand.

The preview is open to Copilot Pro, Pro+, Business and Enterprise subscribers. GitHub names three places to try it: VS Code, the GitHub Copilot app and Copilot CLI. The changelog does not mention a free tier, so Copilot Free users appear to be excluded, though the post does not say so directly.

On cost, GitHub says image generation is billed separately from the chat model. Charges are based on image tokens, which count against the AI credits included in a plan. The changelog gives no per-image price, no token counts per image and no usage caps, so a team cannot yet estimate the bill from the announcement alone.

Admins can switch it off, and watermarks are optional

Organization and enterprise owners can disable the feature by turning off the image generation model in Copilot model policies. That makes it a model-level setting, handled the same way as other models in the Copilot picker.

GitHub also offers a visible “Made with AI” watermark for images that Copilot generates and edits. It is opt-in and off by default. The changelog does not say whether the images also carry embedded provenance metadata, nor who controls the watermark setting, an individual user or an administrator. Companies that need labeled assets will want those answers before turning the feature on widely.

An OpenAI image model inside a coding tool

The most interesting design choice is the tool-calling setup. Copilot does not hand the prompt to a separate image product. The chat model decides when to call an image tool, which means the request can draw on the surrounding conversation and project context, such as a component you are building. GitHub does not publish how much of that context reaches the image model, so the benefit is plausible but not demonstrated.

The model named in the post is GPT Image 2.5 Flare. OpenAI’s developer documentation describes it as its fastest model for high-quality, everyday image generation, with token rates matching GPT Image 2. Speed matters in a chat workflow, where a long wait breaks the flow GitHub says it wants to protect.

Flare is one of two API variants of GPT Image 2.5. According to a third-party guide, the other is Sunburst, which it recommends when image quality comes first. The same guide reports that Sunburst leads Flare on an Arena text-to-image ranking, 1421 to 1399. That is one unofficial comparison, and we found no independent benchmark of either model inside Copilot. By implication, GitHub chose the faster variant over the one that scores higher.

Another listing, on OpenRouter, prices Flare at $8.00 per million input tokens and $30.00 per million image output tokens. GitHub’s own billing is through AI credits, so those API rates do not translate directly into what Copilot subscribers will pay.

Part of a broader Copilot billing shift

The pricing model fits a pattern in recent GitHub announcements. Our report on GitHub Enterprise including Copilot for every licensed user described AI that is billed by use, and the image tokens here draw on the same pool of AI credits. The Copilot app itself has been gaining features quickly, from multi-repo Projects and a computer use mode to fixing code scanning alerts.

The preview label also matters. GitHub has not said when image generation will reach general availability, whether pricing will change, or whether other image models will join the picker. For now, the announcement establishes where the feature works, who can use it and how admins can block it. It leaves open how good the output is, how much it costs per image and how provenance is handled.