GPT-6 Sol October Snapshot: Safety Regressions Split ChatGPT

What happened

OpenAI's October 7 update to GPT-6 did not land evenly across ChatGPT. According to OpenAI's October system card, the new Sol and Luna versions released that day go only to ChatGPT's Chat tab. People using GPT-6 Sol and Luna through Codex or ChatGPT Work are still on the earlier September releases. 1 As of October 7, 2026, paid plans therefore run the October snapshot of GPT-6 Sol in Chat, while Work and Codex keep the September snapshot. 1

The complication is that the October system card itself reports regressions in health and safety compared with the earlier build. 1 For organizations running ChatGPT Enterprise with GPT-6 enabled, staff are now effectively using two different GPT-6 Sol models. Any acceptance testing done in September covers only one of them. 1

On the public release timeline, the update appears as "GPT-6 & Intelligent UI in ChatGPT (All Users)," dated October 7, 2026. That makes it OpenAI's most recent entry in at least one industry release tracker. 2

Why the split matters

Version fragmentation inside one product is easy to overlook. Most users see one brand name, "GPT-6 Sol," and assume one model sits behind it. In practice, which model answers depends on which tab they open. 1 Many enterprises build their governance around a single question: which model did we approve? That question no longer has one answer.

The practical advice is to re-run existing acceptance prompts against the Chat tab before turning on the admin setting for everyone. 1 This is sensible, and it points to a wider problem. A model name is not a stable identifier when the vendor can swap the underlying snapshot in one surface and leave it alone in others.

The safety detail is what makes this more than a bookkeeping issue. Newer snapshots are usually treated as improvements. Here, OpenAI's own documentation says the newer build moved backward on health and safety measures. 1 Teams that approved GPT-6 for sensitive uses, such as anything near medical or wellbeing questions, have particular reason to check the Chat tab separately rather than assume the September results still apply.

There is a fair reading in OpenAI's favor. Disclosing regressions in a system card is the kind of transparency critics have asked for. Keeping Work and Codex on the older build may also be a deliberate choice to protect more workflow-critical surfaces while the new snapshot is tested in Chat. The available material does not say why OpenAI split the rollout, so that motive remains speculation.

The crowded release calendar

The GPT-6 update arrived during an unusually busy stretch. One tracker counts 140 model releases and updates from six providers as of October 7, 2026. 2 On the same day as OpenAI's update, Anthropic released Claude Haiku 5.5, which the tracker lists as the most recent release overall. 2

Haiku 5.5 is described as a small, fast model and the final member of the 5.5 family. It targets high-volume tasks such as summarization, classification and live customer support, and it costs considerably less than Haiku 4.5. 2 It is also the first Haiku with an adjustable effort setting. It is positioned as a low-cost helper agent that can work alongside Opus 5.5 and Sonnet 5.5. 2

Google's latest tracked release is Gemini Nano Banana 2.1, dated October 6, 2026. Meta's most recent is Muse Spark 1.3, from September 2. 2

This pace is part of why the ChatGPT version split matters. When major vendors ship updates days apart, buyers rarely have time to re-evaluate each one carefully. A vendor quietly running different snapshots in different parts of the same product adds another layer that evaluation teams have to track.

The takeaway

The October GPT-6 rollout is a reminder that "which model" now means name, snapshot and surface together. OpenAI did document the change and its safety regressions. 1 The burden still falls on administrators to notice that Chat, Work and Codex have diverged, and to test accordingly.

For enterprise teams, a reasonable approach is to treat each ChatGPT surface as its own deployment. Keep a record of which snapshot each one runs, and require fresh acceptance tests whenever a system card announces a new version. This matters most when that system card admits the new version is worse in some areas. As release calendars fill up across OpenAI, Anthropic, Google and Meta 2, that discipline is likely to become a routine requirement.