An OpenAI Community thread is dissecting a Responses API breakage: after gpt-image-2.6 shipped, requests to the image_generation tool with an explicit reasoning.effort started failing with HTTP 400 and the message "Complexity 'low' is not supported for this image model." The thread's author, henry_m, reproduced it on gpt-6-luna with the gpt-image-2 tool and posted the full script. What got under everyone's skin is that they changed nothing on their end: nobody sends a complexity parameter in the request — the API injects it itself.
henry_m, who opened the thread, boils the symptom down to one sentence: "the issue is that the reasoning effort is passed as 'complexity' at OpenAI's end without validating the model." In other words, the client asks for a low reasoning level, and the server forwards that value as image-generation complexity without checking which model is actually being called.
A participant going by _j probed the adjacent surfaces and mapped the boundaries: "there is no issue when not sending complexity to gpt-image-2 on the generation or edit endpoint." According to him, complexity is a parameter only for gpt-image-2.6, and the weirdness is in the docs themselves: they claim a default of "high." If the Responses API injects "low" on 2.6, expectations break. He also notes that "auto" doesn't exist for the parameter — only null.
sps confirms: "I can confirm that I'm seeing the same behavior on my end" — and says he passed the signal to the team along with thanks for the detailed repro. It's the only message in the thread that mentions an internal escalation rather than outside observations.
rus1243mor describes his own flavor of the same problem: since GPT Image 2.6 launched (his message dates it November 4), the image_generation tool fails on gpt-image-2.5-sunburst whenever reasoning.effort is present in the request. He says he never sends complexity himself, and the error comes back with "medium." His workaround: "omit reasoning.effort entirely; the model default (medium) is not forwarded as complexity."
suryak9 and Huzaifa_dev aren't buying the workaround. The former is terse: "Omitting reasoning.effort is not a solution as sometimes we need to explicitly increase reasoning effort." The latter unpacks the point: "The coupling is the actual problem here" — as long as the image tool lives inside a Responses call, a single reasoning.effort knob controls two unrelated things. He states outright that he's stopped using the hosted tool wherever he needs control over reasoning.
At the end, OpenAI_Support replies: the team shipped a fix "for a related issue where the reasoning setting could trigger an unsupported image complexity error." They suggest verifying by replaying the original request with the same models and reasoning setting, and if the error recurs, sending the exact message text, the time with timezone, the request ID, and account details privately.
There's common ground in the thread, but it's narrow: several independent participants reproduce the failure, none of them sends complexity by hand, and per the thread author and _j, the cause is reasoning effort being silently injected into someone else's parameter. Plus sps's confirmation and support's reply about a shipped fix.
The real conflict is over the workaround. For _j and rus1243mor, dropping reasoning.effort is a temporary hack to survive until the fix lands. For suryak9 and Huzaifa_dev, it's capitulation: they paid specifically for control over reasoning depth, and coupling two settings in one call breaks their use case. There's no follow-up verification after the fix in the thread — so whether the bug is actually closed can't be settled from the comments shared.
reasoning.effort: "low" alongside the image_generation tool, then the same request without the reasoning block — compare response codes.complexity explicitly to the image generation and edit endpoints: per _j, no error there.Disclaimer: this is a recap of a discussion, not editor-verified facts. Opinions belong to the participants; the numbers belong to whoever cited them.
Source: community.openai.com · post in Telegram