2026-09-25 Changes confirmed medium confidence

OpenAI Fixes Image-Encoding Bug in GPT-6 Sol and Luna

Developers using visual inputs are being told to rerun evaluations after OpenAI corrected model input handling across API and Codex workflows.

Edited by Tyronne Panaino

OpenAI said on September 25 that it fixed an image-encoding bug that had degraded image understanding in GPT-6 Sol and GPT-6 Luna. The OpenAI API changelog says the correction improves visual tasks in the API and Codex, including computer use, and advises customers with image inputs to rerun evaluations and retry affected workflows.

The practical implication is narrower than a new model release but still important for teams that evaluated visual behavior before the fix. Earlier results may reflect the encoding defect as well as the models' underlying capability. That does not mean every image request failed or that every prior score is invalid; OpenAI's entry does not quantify the affected share or publish a measured improvement.

What OpenAI changed

The correction is described as an image-encoding fix for two named models: GPT-6 Sol and GPT-6 Luna. OpenAI connects the defect to reduced image understanding and says the update improves visual work across its API and Codex surfaces.

The changelog does not identify a new model name or ask developers to change an endpoint. Instead, its operational advice is to repeat evaluations and retry workflows that involved images. For application teams, that makes revalidation more useful than assuming the same historical result will persist after the correction.

How to revalidate image workflows

A controlled rerun should keep the test set, prompts, tool configuration and scoring method unchanged where possible. Comparing the earlier and current outputs can show whether a workload was sensitive to the corrected input path. Teams should retain both result sets and record when each run occurred so that a quality change is not confused with a changed test.

Computer-use workflows deserve the same treatment because OpenAI explicitly includes them in the affected visual-task surfaces. The source does not state that every computer-use run was degraded, so the appropriate response is targeted evaluation rather than a blanket conclusion about reliability.

If an application retried failed tasks while the bug was present, its review should include both final success rates and the number of attempts required. That is an editorial recommendation for measuring operational impact, not a metric supplied by OpenAI.

What remains unknown

The short changelog entry does not provide an affected time window, request volume, severity rating, image-format breakdown, benchmark delta or independent reproduction. It also does not separate the effect by Sol versus Luna. Those omissions limit any claim about how broadly the problem changed results.

The evidence is authoritative for OpenAI's own correction and guidance, but it is still a single first-party account. Internal confidence is therefore medium. Independent before-and-after tests, or a more detailed incident note, would be needed to characterize the defect's practical scale.

What to watch next

The clearest next checkpoint is workload-level evidence from teams that rerun the same image evaluations after the fix. Additional OpenAI documentation could also clarify the affected period and whether any request patterns were more exposed than others.

Status

Confirmed. OpenAI documents the image-encoding correction and recommends fresh evaluation for image-input work; the extent of the earlier degradation remains unspecified.