GPT-6 Sol and Luna's image fix: what to retest

If GPT-6 Sol or Luna recently struggled with screenshots or other image inputs, revisit those failures before ruling either model out. OpenAI's API changelog records a 25 September 2026 fix for an image-encoding bug that degraded image understanding. The vendor says the update improves visual tasks in the API and Codex, including computer use, and recommends repeating affected evaluations. OpenAI API changelog

That is a useful maintenance update for teams already trialling these models. It is not evidence that every visual workflow is now dependable.

What changed, and when

Both models were released on 22 September. The fix is recorded under 25 September; Mimir first identified the entry during its 27 September scan. The note does not provide an exact deployment time, a list of affected requests or a measured before-and-after improvement. We have not run independent tests of the fix.

The announcement concerns image understanding. A screenshot-reading failure deserves another look; the note does not establish that unrelated text-only failures have changed. Keep those investigations separate.

The model documentation lists image input and text output for both GPT-6 Sol and GPT-6 Luna. This distinction matters when describing the update to colleagues: analysing a supplied image and generating new artwork are different jobs.

Start with the work that actually failed

Our recommendation is to collect a small set of affected examples from your own trial: an unreadable invoice field, a misinterpreted dashboard or a screenshot that led an assistant towards the wrong control. These are illustrative cases, not reports of Mimir customer incidents.

Use approved test data and preserve the original image. Record the model identifier, prompt, image dimensions, detail setting, reasoning setting and run date. Keep the previous response alongside the new one.

First repeat the workflow with its original configuration. If you also change the crop, prompt and model settings, you will not know which change helped. Make those adjustments in a separate round and label them.

For browser or desktop actions, use a test environment. Replaying an old task against a live account could repeat an action that already succeeded, even if the explanation was poor.

Check the decision, not just the description

A more convincing explanation is not enough. Define the observable result you needed: the correct invoice reference, the correct chart series or the intended control selected without an unwanted action.

Have a reviewer inspect the source image and the resulting output. Record unresolved errors and whether a person had to correct the answer. Repeat important cases rather than treating one successful attempt as proof of reliability.

OpenAI's vision guidance still describes limitations involving small text, rotated images, visually complex graphs and precise spatial localisation. A fix to one encoding problem does not remove those documented limitations. Images and vision guidance

For a workflow that reads a screen and then changes a business system, retain the same approval boundaries. Our guide to AI agents for small businesses explains the wider operating context.

Budget for the rerun

The current standard text-token rates, for prompts up to 272,000 input tokens, are US$2 input and US$10 output per million tokens for Sol; Luna is US$0.10 input and US$0.50 output. These are usage rates, not a fixed price per document or screenshot. Cached input, longer prompts and other processing tiers have separate rates. Sol pricing, Luna pricing

Image inputs consume billable tokens too. Use actual usage records for your rerun, including retries and tool charges where applicable. A cheaper model only helps if the completed task meets your requirements without excessive review work.

What we would do next

Reopen the affected visual cases, preserve the old evidence and document the new outcome. Keep a fallback for anything that still fails. Broader changes to production automation should follow a scoped review, such as the work covered by our automation and systems hardening service.

The useful result is a clear decision about your workflow: ready to proceed, needs a change, or still unsuitable. OpenAI's fix gives you a reason to reassess; your evidence determines the decision.