OpenAI Previews GPT-Image-2 Transparency While Users Report GPT-5.6 Routing Bugs

OpenAI is previewing transparent background support for GPT-Image-2 through its API and in ChatGPT. [2][3] Users can activate the feature by using the "background=transparent" parameter. [3] The company recommends leaving background descriptions out of prompts, as the model may otherwise still generate one. [3] Baking the alpha channel in during generation yields better results than conventional background removal, particularly on complex areas like thin fibers or transparent glass. [3] The Cookbook details four use cases for the feature: merchandise artwork, design elements, PowerPoint diagrams, and product shots. [3]

One developer tested the feature by printing a generated cactus illustration and placing it on his laptop. [2] The feature eliminates the need to manually cut out backgrounds in an image editor or regenerate whole images for different backdrops. [2] Because the model produces raster graphics, it does not guarantee pixel-perfect accuracy for exact numbers. [3] To implement the API setup, developers need an API key, Python, and the OpenAI and Pillow libraries. [3] OpenAI notes this is a preview feature, so users should expect rough edges as the rollout expands. [2]

Separately, OpenAI announced a 20 percent price reduction for GPT 5.6 Sol across the API, Codex credits, and ChatGPT Work. [6] Alongside the price drop, ChatGPT users have documented a model-routing mismatch where GPT-5.6 Sol requests are handled by GPT-5.5-mini. [5][7] One administrator for a ChatGPT Business workspace reported that tests performed three minutes apart showed Chrome incorrectly falling back to GPT-5.5-mini, while Microsoft Edge maintained correct routing. [5] In this instance, no rate-limit, usage-limit, or fallback notices appeared in the Chrome browser. [5] The user submitted two sanitized HAR files, request IDs, timestamps, and a comparison report to OpenAI Support. [5]

Another report verified the issue using network telemetry, showing an outgoing request for "gpt-5-6-thinking" returning a server resolution of "gpt-5-5-mini". [7] In the separate Plus account report, the server-derived telemetry also contained a plan type of "plus" and a plan type bucket of "paid". [7] This routing bug was observed on August 21, 2026, persisting for multiple hours on a ChatGPT Plus account in Germany. [7]

The integration of native alpha channel support into GPT-Image-2 addresses a long-standing friction point in generative image workflows, allowing design assets to be immediately layered into presentations and websites without secondary processing. However, the operational updates are overshadowed by the silent model-routing failures affecting paying subscribers. When users explicitly select a flagship reasoning model like GPT-5.6 Sol, an unnotified fallback to a smaller, less capable model undermines trust in the platform's reliability and transparency. This behavior specifically penalizes enterprise and Plus users who rely on the advanced reasoning capabilities they are paying for. What the sources don't address: Whether the browser-specific model downgrades are the result of an intentional load-balancing strategy during peak traffic, or a strict frontend parsing bug in the ChatGPT Web client.