If your ChatGPT Enterprise workspace has GPT-6 switched on, your staff are now using two different GPT-6 Sol models, and the acceptance tests you ran in September cover only one of them. OpenAI's October system card says the Sol and Luna versions released on October 7 go to ChatGPT's Chat tab, while "users accessing GPT-6 Sol and GPT-6 Luna in Codex, and via ChatGPT Work, are still using previously released versions," which it labels September. The October Sol also scores lower than its predecessor on several of OpenAI's own health and safety evaluations. Before you let the new Chat experience reach everyone, re-run your acceptance prompts on the Chat tab, check the admin setting that controls Enterprise access, and confirm your desktop fleet can render the new answers.
The launch itself is a consumer story. OpenAI calls the new response format Intelligent UI and pitches it as visual: tappable buttons, task-specific calculators, interactive charts and editable graphs, shown off with a wing-lift diagram and a savings calculator. The enterprise story sits in the release notes and the system card.
Which Model Is Each ChatGPT Tab Running Now?
As of October 7, 2026, the Chat tab runs the October snapshot of GPT-6 Sol on paid plans, and Work and Codex still run the September snapshot. MacRumors reports that Plus, Pro, Business and Enterprise get GPT-6 Sol, Free and Go get GPT-6 Luna, and that GPT-6 applies only to the Chat tab, with the models behind Work and Codex unchanged. The system card says the October models "will replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT."
The September versions are the ones OpenAI shipped on September 22, when TechCrunch reported that Sol and Luna were available in ChatGPT Work and Codex for most paid accounts and in the API. If your developers moved API workloads after DevDay, they may be on a third model. OrcaRouter notes that GPT-6.1 Sol shipped on September 29 and that OpenAI's current pricing table no longer lists GPT-6 Sol, though the model is still callable. We covered that model's cost and score gap in GPT-6.1 Sol's $5.47 task.
| Surface | Model as of Oct 7, 2026 | Source |
|---|---|---|
| ChatGPT Chat (Plus, Pro, Business, Enterprise) | GPT-6 Sol, October | OpenAI system card |
| ChatGPT Chat (Free, Go, from Oct 8) | GPT-6 Luna, October | MacRumors |
| ChatGPT Work | GPT-6 Sol / Luna, September | OpenAI system card |
| Codex | GPT-6 Sol / Luna, September | OpenAI system card |
| Pro reasoning option | GPT-6 Astra (no Intelligent UI) | OpenAI release notes, via Kingy AI |
| API | GPT-6 Sol still callable; GPT-6.1 Sol current | OrcaRouter |
The practical effect: an analyst who asks the same question in Chat and in Work can get answers from two different snapshots, with different lengths, formats and refusal behaviour. A model snapshot is a frozen, dated version of a model's weights and safety tuning; two snapshots with one marketing name are still two models for testing purposes. Work is the agentic surface OpenAI launched in July to run long tasks across apps and files, so your longer multi-step jobs keep running on the older snapshot while Chat moves to the new one.
What Does the October System Card Show That September's Sign-Off Missed?
It shows regressions on health and safety evaluations, and it measures them mostly against the August GPT-5.6 models, so it tells you little about how the October Sol differs from the September Sol your Work users have. Section 3.1.2 of the system card frames its comparisons "relative to their respective GPT-5.6 (August) counterparts." Several of its tables (the production benchmarks and the under-18 results among them) did not render in the copy we read, so the figures below are the ones OpenAI states in text or that came through in its tables.
On HealthBench Hard, the October GPT-6 Sol scores 27.8 against 31.4 for the August GPT-5.6 Sol, length-adjusted. OpenAI says the slight HealthBench regressions stem mostly "from the length score-adjustment rather than the unadjusted health score," and the unadjusted score actually rose, from 27.1 to 30.9. The more useful number for a workplace is the length: the October Sol's mean HealthBench Hard answer runs 2,396 characters against 1,450, roughly 65% longer. On HealthBench Professional the answers grow from 2,894 to 4,360 characters. If your users paste ChatGPT output into tickets, emails or case notes, expect more of it.
The safety findings are the ones to put in front of whoever signed off your acceptable-use policy. The card says GPT-6 Sol (October) "shows a statistically significant regression on standard self-harm" in production benchmarks, and that both October models regress on "age-restricted content, sexual content, and emotional reliance" in the under-18 evaluations. OpenAI adds that these evaluations "do not account for system-level interventions," that the violations were "borderline but still generally safe," and that it applies "an additional classifier-based block to responses that may contain self-harm, sexual content, and gore." That is the strongest case for OpenAI: the regressions are measured on hard cases, and a separate filter sits in front of the model. It is still a change to a model your staff already use, published the day it shipped.
The card also records gains. Instruction hierarchy robustness is 99.99% for Sol. In the auto-review test, the August GPT-5.6 models exploited gaps in 0.3% of rollouts and the October models in none. On the "respecting warnings" persistence measure, Sol falls from 34% to 28%. OpenAI treats both October models as High capability in cybersecurity and in biological and chemical risk, and puts the October Sol at 68.85% on SEC-Bench Pro against 85.4% for GPT-6 Astra, the model we covered when OpenAI pulled its 6.1 successor.
The October Sol may well be better for your work. You cannot tell from the card, because its published baseline is the August model your Chat users are leaving, while your Work users still have the September one.
What Do Admins Actually Control?
Admins control whether Enterprise users get it at all, and not much after that. OpenAI's October 7 release notes, as quoted by Kingy AI, say "Enterprise access also depends on admin settings," and the model-access guidance says "actual account and workspace permissions still apply." That admin setting is your gate. Business workspaces got access on the same October 7 start date, per MacRumors, and Kingy AI calls the dates "rollout start dates, not a promise that every account switches simultaneously."
The rest is user-side. According to the same help-page excerpts:
- There is "no separate Intelligent UI usage quota; existing model, tool and plan allowances apply," so interactive answers draw on the allowances your plan already has.
- Plus includes Medium and High reasoning effort; Pro, Business and Enterprise also include Extra High.
- The Pro reasoning option "uses Astra and does not support Intelligent UI," so users on that setting will not see the interactive answers.
- A web setting called "Layout and visuals" can reduce visual elements but "does not guarantee a completely text-only experience."
That last point matters for any team that pipes ChatGPT output into a document workflow or a screen reader. You cannot promise those users plain text. Test what an interactive answer looks like when it is copied, shared or exported, including through any shared page or Space your teams rely on, before you tell them it works.
The desktop client is the other gap. OpenAI's guidance, again per Kingy AI, says "the older macOS and Windows desktop apps do not support these capabilities," and gives no version numbers. Desktop Insights, a third-party tracker, lists a macOS app named ChatGPT Classic whose latest tracked release is dated August 6, 2026. If your device management still deploys an older ChatGPT build, users on it will not see what users on the web see, and they will file it as a bug.
How Should You Re-Test a Model You Already Approved?
Treat the October Sol as a new model for acceptance purposes and run your existing test set in the Chat tab itself, since results from the API or Work describe a different snapshot. An acceptance prompt set is the fixed list of real tasks and red-line prompts your organisation ran before approving a model; it only means something when it runs on the surface your staff use. We argued in the September GPT-6 Sol launch piece that you should re-run evals at the effort you deploy, and the same logic applies here to the surface you deploy.
Three things to capture per prompt: whether the answer is correct, how long it is, and whether it rendered as interactive UI. Length is the change the system card makes measurable, and it is the one your users will notice first. If you keep your test set in a tracing tool such as Langfuse, tag each run with the tab and the date so September and October results never get averaged together. Our guide to building a 200-question eval set covers how to write the prompts if you have only a handful.
If you work in healthcare, benefits, HR or any function where staff ask about self-harm, minors or emotional support, add those prompts explicitly. The card's regressions sit in exactly those categories, and your existing set may not test them because September's model gave you no reason to.
This Week:
- Find the admin setting that controls GPT-6 access for your Enterprise workspace and decide, in writing, whether it stays on while you test.
- Run your acceptance prompt set in the Chat tab with GPT-6 selected, and again in Work, and record answer length and format alongside correctness.
- Pull your device-management inventory for ChatGPT desktop builds and schedule updates for anything older than the current release.
This Month:
- Send the system card's self-harm and under-18 findings to your AI risk owner and your HR or wellbeing lead, with the classifier caveat attached.
- Tell staff in one paragraph that Chat, Work and Codex now run different GPT-6 snapshots and may answer the same question differently.
- Add "which snapshot, on which surface" as a required field in your model approval record, so the next silent split shows up in the register.
The Bottom Line
OpenAI now ships model updates per surface inside one product, and ChatGPT Enterprise buyers inherit that split whether they asked for it or not. The cloud providers taught this lesson with managed services: a "same" service that updates on different schedules in different regions forces you to pin and test per region. Our review of model deprecation windows found the same pattern across API vendors. None of the October material describes a way to pin a ChatGPT tab to a snapshot, so the admin setting and your own test set are the only controls you have. Run the Chat tab test before the setting goes on for everyone.
Continue Reading
- GPT-6 Sol Halves the Token Price but Benchmarks It at Top Effort
- GPT-6.1 Sol's $5.47 Task Scores 11 Points Below Astra's $23.80
- Few-Shot Stopped Paying on GPT-4o. Qwen Still Wants It.
- Vendor AI Benchmarks: Only Numbers With Logs Belong in the Deck
- OpenAI's Ultrafast Bills Astra at 6x and Skips EU Residency
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic