OpenAI fixed a vision bug in GPT-6 Sol and Luna three days after launch — and neither model has a dated snapshot to tell the two versions apart

A three-line changelog entry with a large blast radius

OpenAI’s API changelog for 25 September 2026 says this:

“Fixed a bug in image encoding that degraded image understanding in GPT-6 Sol and GPT-6 Luna. This update improves results on visual tasks in the API and Codex, including computer use.”

The entry goes on to recommend that developers with image input use cases rerun their evaluations and retry affected workflows.

Start with the credit, because it is deserved. OpenAI could have shipped this quietly. Encoding fixes are the kind of change that routinely lands with no note at all, and a vendor that writes down “our model was worse than it should have been for three days, go re-measure” is behaving better than the industry baseline. The entry is specific about the defect, specific about which models, specific about which workloads, and explicit about what developers should do.

Now the timeline. Sol and Luna were released on 22 September, at $2/$10 and $0.10/$0.50 per million tokens — a flat 50% cut against the GPT-5.6 tier. A halved price on a frontier-class model is the single most reliable trigger for evaluation work in this industry. Within 72 hours, benchmark write-ups, vendor bake-offs, migration spikes and “is the cheap one good enough now” analyses were all in flight simultaneously.

That 72-hour window is precisely the interval in which the models’ image understanding was degraded.

The part with no workaround

If this were only a bug and a fix, it would be unremarkable. Software has defects; vendors fix them. What makes it structural is what OpenAI’s own model reference pages show.

The page for gpt-6-sol lists one snapshot identifier: gpt-6-sol. It instructs developers to “use gpt-6-sol in your API requests”. There is no gpt-6-sol-2026-09-22. The page for gpt-6-luna is identical in this respect — one string, no dated aliases. Both models carry a 1,050,000-token context window and accept text and image input.

So the model that degraded your image understanding on 23 September and the model that does not degrade it on 26 September are the same string. Not similar strings. The same one.

Three capabilities that mature engineering practice assumes are therefore unavailable:

• You cannot hold a version steady. The standard way to qualify a new model is to pin the current one, test the candidate, then switch deliberately. With a floating-only identifier, the thing you are holding steady can move underneath you mid-qualification.
• You cannot reproduce a prior result. A benchmark number from 23 September describes a build that no longer exists and has no name. Nobody — not you, not OpenAI, not an independent evaluator — can re-run it.
• You cannot roll back. This particular change was an improvement on average, which makes rollback sound irrelevant. It is not. A fix to image encoding changes how images become tokens, and prompts tuned against the old encoding can regress even as aggregate quality rises. Teams that spent three days adjusting prompts to compensate for degraded vision have just had their compensation invalidated, in the opposite direction, with no way to revert while they re-tune.

This is a genuine departure from the convention OpenAI itself established. The GPT-3.5 and GPT-4 generations shipped dated snapshots as a matter of course — gpt-3.5-turbo-1106, gpt-4.1-nano-2025-04-14 — with the undated alias offered as a convenience for people who wanted the latest. Those dated strings are load-bearing infrastructure: they are what makes a deprecation calendar expressible at all, which is why this week’s shutdown list is written in snapshot IDs rather than in model families. The GPT-6 tier has dropped them.

The question OpenAI’s launch numbers now carry

OpenAI’s launch materials led with agentic and computer-use results. Among them, per this desk’s record of the launch, an OSWorld 2.0 score of 60.5% at xhigh effort and an AutomationBench 1.0.6 score of 33.2% at xhigh.

OSWorld is a desktop-automation benchmark. The model is shown screenshots and asked to operate a computer. Its entire perception of the task is images — exactly the pathway the encoding bug degraded, and exactly the workload the changelog names when it says the fix “improves results on visual tasks in the API and Codex, including computer use”.

There are only two possibilities, and OpenAI’s changelog does not say which holds:

1. The launch figures were measured on the shipped, degraded build. In that case Sol’s real computer-use capability is better than OpenAI advertised, and the published numbers understate the product.
2. The launch figures were measured on an internal build that already had the fix. In that case the numbers were true of OpenAI’s model but not of the endpoint developers could actually call between 22 and 25 September.

Neither is scandalous. Both matter to anyone who used those figures to decide something, because in case 1 the benchmark is stale-low and in case 2 the benchmark never described the shipped artefact. And the asymmetry is worth noticing: a vendor bug that makes a vendor’s own launch numbers look worse than reality is the rare error that argues for the vendor’s honesty rather than against it.

Text-only results are not implicated. An image-encoding defect does not touch DeepSWE, SciCode or terminal-task scores, so those can be read at face value. This is also the clearest available argument for why independently measured index scores on private test sets are worth more than launch-day slides: an outside evaluator re-running on a schedule catches a step change like this one. A launch deck, by construction, describes a single moment that may already have passed.

Who was affected without knowing it

The teams most exposed here are the ones that do not think of themselves as doing vision work. Image inputs enter production pipelines through paths nobody labels:

• A document step that rasterises PDF pages before reading them.
• A support workflow that accepts customer screenshots.
• An agent that screenshots a browser to decide its next click — the core loop of computer-use tooling.
• A coding assistant reading a design mock, a failing chart or a screenshot of a stack trace. The changelog names Codex specifically, which means Codex users were in scope whether or not they considered themselves to be sending images.

For developers auditing this, the reliable query is against request logs rather than source code: filter the 22–25 September window for requests carrying image content parts. Searching the codebase for the word “vision” will miss most of them. Anything that comes back needs re-running — and then a second check, on whether a decision has already been made on those numbers. A routing rule, a model selection, a migration sign-off or a published comparison built on pre-fix measurements is now resting on a defect.

There was also independent friction around Sol and Luna’s image handling in this period on third-party surfaces, including reports of image inputs being rejected outright for these models on Bedrock. Those are integration bugs in other people’s adapters rather than OpenAI’s encoding defect, but they compound the same conclusion: the image path on a three-day-old model tier is the least settled part of it.

The thesis, stated plainly

A model ID is not a version number, and this week it stopped even pretending to be one.

OpenAI did the disclosure well and the versioning badly, and only the second half is durable. The bug is fixed; the fix is an improvement; anyone who reruns their evaluations today gets an honest picture of a better model. But the condition that produced the problem is still in place: a frontier tier with image input, sold on computer-use performance, exposing exactly one unversioned string per model. The next silent change behind gpt-6-sol will be just as unpinnable, and there is no guarantee the next one is an improvement or that it comes with a changelog entry.

Capability moves under stable names in both directions. It moved up here. It has moved down elsewhere this month, by routing rather than by bug, and the same month saw image models ship new quality settings behind unchanged token prices. In none of those cases does the rate card or the version string tell you anything happened. The premium-tier procurement arguments that dominate ChatGPT-platform buying decisions are all built on benchmark deltas between named models — and a delta between two numbers is meaningless if one of the names quietly referred to something else when it was measured.

The cheap fix is a timestamp. Write the UTC time next to the model ID in every evaluation artefact you keep, treat that pair as the identity of what you tested, and run a small regression suite on a schedule instead of only at migration time. A dozen fixed prompts with known-good outputs, including image cases, would have caught this on 25 September without anyone reading a changelog at all.

Questions or corrections? Email Pick Right. Want the full list? See all news.