reported that OpenAI fixed GPT-6 Sol the image encoding issue involving Luna
BIBIBI
AT A GLANCE
The breaking-news report stated that OpenAI fixed GPT-6 Sol the image encoding bug affecting Luna, and that Luna (Max) showed a significant improvement in the visual localization evaluation.
Article
OpenAI Fix GPT-6 Image bug, Luna visual localization doubled
Beating AI breaking news reported that OpenAI fixed GPT-6 Sol and GPT-6 Luna an image encoding bug. This issue reduced the models' ability to understand images. After the fix,API and Codex's visual tasks will improve, including the Computer Use。
OpenAI published test showed that GPT-6 Luna(Max)'s average IoU on the RefCOCOg visual localization evaluation rose from 28.8%to 60.0%, an increase of 31.2 percentage points. This test mainly evaluates whether a model can accurately locate a corresponding object or region in an image based on a description.
The two models September 22 officially went live only in API，OpenAI's changelog at September 25 also recorded this fix. The official source did not disclose Sol's specific improvement or state when the bug began or how many image requests it affected.OpenAI also unusually recommended that developers using image inputs rerun their evaluations and retry workflows previously affected.
Original article link https://m.theblockbeats.info/flash/369251
Key points
01
Beating AI reported that OpenAI fixed GPT-6 Sol an image encoding bug in GPT-6 Luna and
02
The original article stated that this issue reduced the models' ability to understand images. After the fix,API and Codex's visual tasks will improve, including Computer Use。
03
The report cited tests OpenAI published by GPT-6 Luna(Max)'s average IoU rose from RefCOCOg the visual localization evaluation from 28.8%to 60.0%, an increase of 31.2 percentage points.
04
The original article stated that the two models officially went live in September 22 API，OpenAI The changelog was updated on September 25 to record this fix.
05
The original article stated that OpenAI recommended that developers using image inputs rerun their evaluations and retry workflows previously affected.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This bug affected the process by which models encode and understand image content.RefCOCOg The evaluation examines whether a model can locate a corresponding object or region in an image based on a textual description; average IoU can be understood as the degree of overlap between the predicted region and the correct region. The figures in the report show that Luna (Max)'s post-fix localization results overlap more with the target regions.
Why it matters to readers
Those relying on image understanding, screen recognition, or Computer Use's API and Codex workflows may be affected. The original article recommended that developers reevaluate and retry their workflows, indicating that previous results may not represent performance after the fix.
If you are using image inputs or having an Agent recognize a screen, rerun your own tests and compare localization performance before and after the fix; the single evaluation result in the report does not mean that all visual tasks will improve by the same amount.
Risks and unknowns
Evidence not attached OpenAI The announcement, changelog, or original test is not attached, so the relevant claims cannot be independently verified from the current material.
OpenAI Not disclosed GPT-6 Sol specific improvement magnitude.
The time when the bug began affecting users and the number of affected image requests were not disclosed.
Luna (Max)'s improvement on the RefCOCOg evaluation is not known to be reproducible on other visual tasks.
Related Developments
Loading event timeline…
Related concepts
image encoding
This term is not in the glossary yet. Browse related concepts in the glossary.
