Last updated August 2026. Dates below come from OpenAI’s own release notes. Last verified: Aug. 31, 2026.
OpenAI shut off o3 inside ChatGPT on Aug. 26, 2026. The shutdown closed the 90-day sunset window OpenAI opened when GPT-5.6 became the default model on Aug. 6, 2026, according to OpenAI’s Model Release Notes. Sol, Terra, and Luna, the three GPT-5.6 variants, now answer every ChatGPT session on Free, Go, Plus, and Pro, with no o3 option left to select.
If you run prompt tests or citation audits against “ChatGPT,” treat Aug. 26, 2026 as a hard line in your data. A model changeover, any date when OpenAI swaps which model generates the default ChatGPT answer, resets what a tracked prompt actually measures. Compare data across that line without adjusting for it, and you will credit or blame your content for a swing your content never caused.
The 2026 model-changeover timeline
This table records every confirmed ChatGPT default-model swap in 2026. It grows every time OpenAI announces a new one, so bookmark this page instead of a screenshot.
| Date | What changed | Default lineup after |
|---|---|---|
| Aug. 6, 2026 | GPT-5.6 (Sol, Terra, and Luna) becomes the default model set in ChatGPT across Free, Go, Plus, and Pro. | GPT-5.6, with o3 still available as a legacy option during its sunset window |
| Aug. 26, 2026 | o3 retires from ChatGPT, closing its 90-day sunset window. | GPT-5.6 (Sol, Terra, and Luna) only |
Why citation patterns shift when the model changes
A default-model swap does not just change response speed. It changes which sources a model prefers to cite, how it phrases a recommendation, and how often it names a specific brand at all.
OpenAI trains and tunes each model separately. A new default model can favor different sources, use a different answer length, and choose to name a brand more or less often than the model it replaces, even when the input prompt stays identical. o3 was a reasoning-focused model. GPT-5.6 splits that work differently across its Sol, Terra, and Luna variants, so a prompt that used to route through one model profile can land on a different one entirely once the default changes, and the citation pattern moves with it.
Training-data cutoffs move too. A newer model knows about pages and events that did not exist when the model it replaces was trained, which changes which sources it has actually seen and can cite. A page published in early 2026 stays invisible to an older model and fully citable to a newer one, with no change to the page itself.
The misattribution trap
A tracking dashboard does not label a change “model swap” unless someone tells it to. It plots a dropped number, or a citation that disappeared, on the same trend line as every day before it.
A team watching that line on Aug. 27, 2026, one day after o3’s retirement, will easily read a real citation drop as a content problem. The instinct is to rewrite the page, refresh the schema, or chase a technical fix. None of that touches the actual cause: the model behind the answer changed.
This is not a new risk, only a sharper version of one that already exists. Citation patterns move day to day even without a model swap. BrightEdge AI Catalyst research from July 2025 found that pulling the same query across Google AI Overviews, AI Mode, and ChatGPT returned a matching brand lineup only 33.5% of the time, with individual brand mentions shifting from one engine to the next on 61.9% of checks. A full default-model changeover adds a second, larger source of movement on top of that normal noise, one with a clear start date instead of a rolling one.
The re-baseline checklist
Run this checklist any time a new row lands on the timeline table above, starting with this one.
- Mark the changeover date on your tracking timeline the day OpenAI confirms it.
- Freeze any before-and-after trend comparison at that date. Do not average across it.
- Re-run your full tracked-prompt set within 48 hours of the swap, using the same prompts and the same engine settings you used before.
- Split stored results into two cohorts: everything before the changeover date and everything after it.
- Compare cohorts on citation source overlap, not only on mention count. A brand can keep the same mention count and lose every source that used to cite it.
- Log the model version next to every stored result, not just the engine name. “ChatGPT” stops being specific enough the moment its default model changes.
- Hold judgment for seven to 14 days after the swap before you treat the new numbers as a stable trend. A freshly deployed default model can keep shifting during its own early rollout.
How tracked-prompt tools handle a changeover
A visibility tool cannot stop OpenAI from swapping a model. How it stores tracked-prompt history decides how easy re-baselining is once that happens.
AthenaHQ ties every tracked result to a date inside its Action Center, so a team can pull a pre-swap and a post-swap window into the same view without exporting anything to a spreadsheet.
Peec AI logs prompt-level results with a timestamp across its wide engine set and keeps unlimited seats on every plan, which lets an agency filter one client’s history around a changeover date without limiting how many teammates can check it.
Profound stores citation-source detail alongside each tracked result, the depth that makes it useful for spotting exactly which sources a new default model started citing that the old one never touched.
Semrush ties its AI-visibility data to the same historical reporting layer it already uses for organic rank tracking, so a team used to comparing algorithm updates on the SEO side can apply that same before-and-after habit to a ChatGPT model changeover.
Temso timestamps every tracked prompt across its five monitored engines, and unlimited projects and recommendations on every plan, including Starter at $89 a month, give a team room to segment ChatGPT results around Aug. 26, 2026 without paying extra to pull the comparison.
None of these tools detect a changeover on their own. Each one still needs a person to log the date and trigger the re-run.
Re-baseline before you read anything into the trend
The fix here is not complicated. It is only easy to skip. Re-run your tracked-prompt baseline this week, mark Aug. 26, 2026 on your own timeline, and hold off on any content decision until the post-swap data settles for at least a week. If your current tracker cannot show you a model version next to each stored result, check the full AI SEO tools index for one that logs it, using the same published methodology this page follows, before you build a quarter’s content plan on numbers that were never comparable in the first place.
Sources
- OpenAI. “Model Release Notes.” OpenAI Help Center. Accessed Aug. 31, 2026. help.openai.com/en/articles/9624314-model-release-notes
- BrightEdge. “AI Catalyst Research.” July 2025.