#

Key takeaways

• OpenAI logged a fix for GPT-6 Sol and GPT-6 Luna on September 25, 2026.

• The issue involved image encoding and degraded image understanding.

• OpenAI says the fix improves visual task results in the API and Codex.

• The company recommends rerunning evaluations and retrying affected workflows.

• The source does not give impact numbers or a root-cause detail.

OpenAI logs a visual-task maintenance fix

OpenAI's API changelog lists a fix dated September 25, 2026. It applies to GPT-6 Sol and GPT-6 Luna. The entry says an image-encoding bug had degraded image understanding.

The changelog frames this as a maintenance update, not a new model announcement. It also says the fix improves results on visual tasks in the API and Codex. The source keeps the scope narrow and does not expand beyond that wording.

What the changelog says changed

The source identifies three points clearly. First, the affected models are GPT-6 Sol and GPT-6 Luna. Second, the problem area is image encoding. Third, the observed effect was weaker image understanding.

OpenAI does not describe the underlying defect. It does not name image formats or give a numerical measure of the degradation. It also does not say how many users or requests were affected.

What users are told to do

OpenAI recommends rerunning evaluations and retrying workflows that may have been affected when image inputs are involved. That guidance is practical and limited. The changelog does not present it as a mandatory action for all users.

The recommendation matters most for teams that rely on image-based evaluation or automation. If a workflow depends on visual inputs, the source suggests checking results again after the fix. The changelog does not say whether every image task was affected.

What the source does not say

The entry leaves several questions unanswered. It does not specify when the problem began. It does not say how long it lasted. It also does not state whether a particular model release was required to receive the fix.

The source also avoids broader product claims. It mentions visual tasks in the API and Codex. It does not make a claim about ChatGPT or other products. It also does not estimate customer impact or performance gains.

Operational note for readers

For teams using these models, the safest reading is simple. Recheck any evaluation that used image inputs. Retry workflows if results may have been influenced by the bug.

Because the changelog gives no numeric impact, avoid assuming the size of the improvement. The source supports only a limited conclusion: image understanding was degraded, and OpenAI says the fix improves visual-task results.

Morocco relevance

The source reports no Morocco-specific deployment or availability information. For readers in any market, the global lesson is to rerun image-based evaluations after a documented model fix.

Source context

This report is based only on OpenAI's API changelog entry from September 25, 2026. It reflects the wording in the source and avoids adding unsupported detail.