OpenAI launches GPT-5.1: Combining adaptive inference, account-level personalization and updated security metrics in the GPT-5 stack

OpenAI released GPT-5.1 as the next generation version of the GPT-5 series, with 2 core variants: GPT-5.1 Instant and GPT-5.1 Thinking. This update focuses on 3 axes, adaptive reasoning behavior, clearer explanations, and greater control over tone and security.

Model lineup and positioning

GPT-5.1 Instant is the default conversation model in ChatGPT. OpenAI describes it as the most commonly used model, with warmer default tones and improved instruction following. GPT-5.1 thinking is an advanced reasoning model. It exposes explicit thinking time, which can now be tailored more precisely to the problem. GPT-5.1 Auto continues to route queries between these variants so most end users do not need to manually select a model.

GPT-5.1 Instant, adaptive inference for everyday use

GPT-5.1 Instant maintains the low latency of typical chat usage while adding adaptive inference. For simple prompts, the model responds quickly via shallow internal passes. For more difficult prompts, such as multi-step math or coding questions, it chooses to spend more internal calculations before answering. This behavior results in higher scores in evaluations such as AIME 2025 and Codeforces relative to earlier GPT-5 Instant releases, while maintaining responsiveness for casual use.

Following instructions is another clear goal. In OpenAI’s example, GPT-5.1 Instant is more reliable under constraints such as “always respond with 6 words” and maintains that constraint across epochs. This is relevant when you’re building tools that rely on strict formatting or short natural language responses, such as structured output, message templates, or chained tools that expect bounded lengths.

The combination of adaptive inference and stricter directive compliance makes GPT-5.1 Instant a more predictable front end in many agent workflows, where most calls are simple but tail calls require deeper inference.

GPT-5.1 thinking, dynamic calculation allocation

GPT-5.1 thinking adopts the GPT-5 thinking method and strengthens the use of thinking time. The model now adjusts its internal thinking time based on the complexity of the prompt. On a representative distribution of ChatGPT tasks with standard think times, GPT-5.1 think is about 2 times faster than GPT-5 think on the fastest tasks and about 2 times slower on the slowest tasks.

This is important for workloads where you want a single model to handle both light and heavy queries. Lightweight queries don’t require long chains of thought. Hard reasoning and planning tasks receive more internal steps without requiring any new API interfaces.

GPT-5.1 thinking answers use less jargon and undefined terms than GPT-5 thinking. This reduces the time spent explaining detailed answers and makes the model more suitable as an interactive tutor on topics such as statistics, algorithms, or system design.

On the API, GPT-5.1 Instant appears as gpt-5.1-chat-latest, and GPT-5.1 Thinking appears as gpt-5.1. Both include adaptive inference by default.

Personalization, preset styles and fine-grained tone control

In addition to model updates, ChatGPT is also getting a more explicit personalization layer. Users can select basic styles from the personalization screen, such as Default, Professional, Friendly, Frank, Quirky, Efficient, Nerdy, or Cynical. These presets are available for all models, including GPT-5.1.

OpenAI is also experimenting with more granular sliders in settings. Users can adjust how concise, warm or scannable responses are, as well as how often emojis appear. When ChatGPT detects duplicate beep requests, ChatGPT can propose updates to these preferences in the conversation. Preferences are now immediately applied to new and ongoing chats, unlike previous behavior where changes only affected new conversations.

Security Indicators and Readiness Classification

GPT-5.1 maintains the same overall security framework as GPT-5 and provides updated baseline metrics. GPT-5.1 Instant and GPT-5.1 Thinking use the same types of mitigations described in the GPT-5 system card, including filters for prohibited content, routing of sensitive cases, and policy-consistent denials.

On a production benchmark of banned content, gpt-5.1-instant outperforms gpt-5-instant in all listed categories. For example, illegal or violent content has a not_unsafe score of 0.918, and hateful content has a score of 0.897.

For jailbreak robustness, measured by StrongReject evaluation, gpt-5.1-instant has a not_unsafe score of 0.976, while gpt-5-instant has a score of 0.850 and gpt-5-instant has a score of 0.683. gpt-5.1-thinking scores 0.967, which is close to gpt-5-thinking’s 0.974.

Main points

1. GPT-5.1 introduces 2 main variants, GPT-5.1 Instant and GPT-5.1 Thinking, plus GPT-5.1 Auto as a router inside ChatGPT, and is positioned as a generation upgrade within the GPT-5 series, rather than a new generation model.
2. GPT-5.1 Instant uses adaptive inference so it spends more computation on hard prompts and less on simple prompts, which improves math and coding benchmarks like AIME 2025 and Codeforces while maintaining low latency and stronger instruction following for typical chat workloads.
3. GPT-5.1 Thinking allocates dynamic thinking time to each query, is about 2 times faster than GPT-5 Thinking on simple tasks, is about 2 times slower on the most difficult tasks in standard thinking, and provides clearer, less jargon-filled explanations of technical problems.
4. ChatGPT adds a personalization layer on top of GPT-5.1, with preset styles like Default, Professional, Friendly, Efficient, Nerd, and more, as well as plan sliders for Simplicity and Warmth, so the tone control part moves from prompts to persistent user settings that apply to chats and models.
5. GPT-5.1 reuses the GPT-5 security framework and mitigations, improves gpt-5.1-instant security scores for all prohibited content categories compared to gpt-5-instant, significantly improves StrongReject’s jailbreak robustness, and anchors readiness assessments in GPT-5 system cards for high-risk areas such as biological and chemical capabilities.