Skip to content
Deployment Safety Hub
View PDF

Published October 7, 2026

GPT-6 Sol and GPT-6 Luna: October 2026 update

1. Introduction

We’re launching GPT-6 to everyone in ChatGPT, across free and paid plans globally. These models will replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT. Users accessing GPT-6 Sol and GPT-6 Luna in Codex, and via ChatGPT Work, are still using previously released versions. In this system card, we distinguish these models by their month of release: October for the versions released today, and September for the versions that remain in use in Codex and Work. See our blog for details.

GPT-6 in ChatGPT incorporates Astra’s safety advances. We also updated the safety training to reflect real-world use and strengthen protections against high-risk misuse in cyber, biology, and violence.

We assess safety and alignment evaluations in aggregate, weighing improvements alongside the nature and severity of regressions. Compared with GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT, GPT-6 showed stronger resistance to jailbreaks, including attacks that adapt across multiple turns, as well as reductions in dishonesty, deception, and circumvention of guardrails. For areas where our safety evaluations showed regressions, we conducted manual review of failures and adversarial red-team testing and found that the disallowed responses were generally low severity. System-level mitigations to reduce the likelihood of harm are detailed throughout this system card.

Under our Preparedness Framework, we are treating this October release of GPT-6 Sol and GPT-6 Luna as High capability in both Cybersecurity and Biological and Chemical domains. Neither one reaches our High threshold in AI Self-Improvement. Based on that assessment, we’ve implemented the same set of safeguards for GPT-6 Sol (October) and GPT-6 Luna (October) that are detailed in the GPT-5.6 System Card for GPT-5.6 Sol and GPT-5.6 Luna.

For all safety evaluations, such as disallowed content and mental health, we measure performance of our models at their lowest reasoning deployment settings in order to capture performance for the vast majority of usage. For capabilities assessments, we evaluate the models at their maximum reasoning effort to get an upper bound of capabilities.

2. Model Data and Training

Like OpenAI’s other models, GPT-6 models were trained on diverse datasets and filtered through our data processing pipeline, including to reduce personal information.

OpenAI reasoning models are trained to reason through reinforcement learning. These models are trained to think before they answer: they can produce a long internal chain of thought before responding to the user. Through training, these models learn to refine their thinking process, try different strategies, and recognize their mistakes. Reasoning allows these models to follow specific guidelines and model policies we’ve set, helping them act in line with our safety expectations. This means they provide more helpful answers and better resist attempts to bypass safety rules.

For previously launched models, the values published at launch reflect the versions evaluated at that time. The comparison values for previously launched models that are shown here may reflect later versions of those models, and may vary from the values published at launch.1

3. Model Safety

3.1 Safe Completions

3.1.1 Evaluations with Challenging Prompts

We conducted benchmark evaluations across safety categories. We report here on our Production Benchmarks, an evaluation set with conversations representative of challenging examples from production data. As we noted in previous system cards, we introduced these Production Benchmarks to help us measure continuing progress given that earlier Standard evaluations for these categories had become relatively saturated.

These evaluations were deliberately created to be difficult. They were built around cases in which our existing models were not yet giving ideal responses, and this is reflected in the scores below. Error rates are not representative of average production traffic. The primary metric is safe completion rate, checking that the model’s response and actions are not disallowed according to our safety policies. Our evaluations are run on the model without system-level safeguards to ensure the model’s underlying behavior meets our safety bar. We continue monitoring these categories after launch to evaluate online performance and further adjust safeguards as appropriate.

Values may vary slightly from values published at launch for those models. Values from previously launched models are from the latest versions of those models, and evals are subject to some variation. The comparison scores from earlier models listed below are intended to shed light on relative performance. Because policies, graders, datasets, evaluations, and other measurement details evolve over time, scores from previous system cards not included in the table below should generally not be considered directly comparable to these most recent results.

Relative to their respective GPT‑5.6 counterparts, GPT‑6 Sol (October) shows a statistically significant regression on standard self-harm, while GPT‑6 Luna (October) shows statistically significant regressions on standard self-harm, gore, and sexual content.

We manually reviewed violative responses from these evaluations and found that violations were borderline but still generally safe. For example, GPT‑6 Sol (October) and GPT‑6 Luna (October) are more willing to answer informational questions about self-harm while directing users to appropriate professional resources. Based on our evaluations, it does not comply with requests to facilitate self-harm. GPT‑6 Sol (October) and GPT‑6 Luna (October) are more likely to engage with requests for gore and sexual content rather than refuse outright, however the outputs we reviewed were lower severity. We also conducted adversarial red teaming and did not identify any high-severity risks based on our evaluation criteria.

These evaluations do not account for system-level interventions that elevate the holistic safety of the assistant response, such as Trusted Contact, localized crisis helplines, and Parental Controls for younger users.

Table 1: Production Benchmarks with Challenging Prompts (higher is better)

In addition to measuring how safely our models respond to harmful requests, we assess whether they overrefuse legitimate requests, especially when a topic is sensitive or easy to misinterpret. In the Safety Pareto chart below, GPT‑6 Sol (October) and GPT‑6 Luna (October) show an improvement on helpfulness on legitimate requests relative to prior models, though it scores lower on some safety requests. Our evaluations indicate that GPT‑6 Sol (October) and GPT‑6 Luna (October) are less likely to refuse harmless requests or add excessive or judgmental caveats.

Figure 1