ChatGPT introduces Fast, Thinking, and Auto modes for GPT‑5 reasoning control

ChatGPT now includes selectable GPT‑5 thinking modes and reintroduces access to legacy models in the user interface.

Users can now manually choose how the system processes their inputs, whether through faster completion or deeper reasoning, while also regaining access to previous-generation models through a dedicated settings menu.

A significant interface update has been introduced in ChatGPT, transforming the way users interact with the underlying language models. For the first time, it is now possible to manually select how responses are generated, thanks to the visible introduction of four GPT‑5 operational modes: Auto, Fast, Thinking Mini, and Thinking. These options appear clearly in the model selection dropdown, allowing for direct user control over the system’s internal behavior.

At the same time, a new section in the settings menu labeled "Legacy Models" enables users to re-enable and access previous versions of the model—specifically GPT‑4o, GPT‑4.1, o3, and o4-mini. This dual-layered update brings both forward-facing flexibility and backward compatibility to the ChatGPT experience, addressing long-standing concerns from power users and professionals who rely on specific model traits for continuity, tone, or task precision.

This transformation comes after several weeks of widespread commentary and dissatisfaction regarding the GPT‑5 rollout, which, while promising improvements in reasoning, speed, and unification of capabilities, removed direct access to prior versions and gave users little visibility into how their queries were being processed. The new interface reverses that dynamic, placing the user at the center of decision-making and restoring options that had been abruptly removed.

The four GPT‑5 modes reflect different levels of speed and depth in the system’s reasoning process, and they are now explicitly listed in the main selection panel.

The GPT‑5 model is no longer presented as a monolithic, fixed option. Instead, the assistant can now operate in distinct processing configurations, each designed for a different type of user need. The Auto mode retains the original routing logic introduced at launch: the system dynamically decides, behind the scenes, whether a user prompt should be processed using a faster or a deeper variant of the model, based on factors such as complexity, tool usage, and recent context.

However, users can now bypass this routing entirely. By selecting Fast, they ensure that the model will prioritize low-latency completion—generating output as quickly as possible, ideal for short answers, summaries, or lightweight operational tasks. Conversely, choosing Thinking instructs the system to use more compute resources and time to perform step-by-step internal analysis, particularly beneficial in contexts that require structured reasoning, numerical accuracy, scenario exploration, or abstract logic.

For users who exceed certain usage thresholds, such as those on the Plus plan, the system introduces rate-limited access to Thinking mode—capping it at around 3,000 messages per week. Once that threshold is passed, the model automatically switches to Thinking Mini, a version that preserves part of the deeper analysis process but with lighter computational effort and faster turnaround. This mechanism ensures availability while balancing infrastructure demands.

These modes can now be chosen with a single click before starting a conversation or during ongoing use, and they are persistently visible as individual entries in the model picker. This visibility removes the ambiguity that existed under the previous GPT‑5 rollout, where the user had little to no indication of how the model was behaving behind the scenes.

Previous-generation models have been restored in a separate section, giving users access to GPT‑4o and other familiar versions once hidden during the transition.

In addition to the introduction of GPT‑5 processing modes, OpenAI has officially reintroduced older versions of its models under a new “Legacy Models” configuration, located within the platform’s settings. This decision directly addresses user concerns that surfaced in the days following the initial GPT‑5 deployment, which saw the removal of GPT‑4o and other alternatives from the dropdown menu, despite being actively used and preferred by many.

Now, subscribers to the Plus or Pro plan can open the settings menu, navigate to the “Personalization” section, and enable the legacy models they wish to use. Once activated, these models appear alongside the GPT‑5 options in the selection menu, making it possible to switch between current and previous systems without any loss of continuity.

This move reflects a broader acknowledgment that different versions of the assistant have different strengths—and that the shift to a single default model does not accommodate all use cases. Many users working in education, writing, content strategy, legal documentation, and structured research had developed specific workflows around models like GPT‑4o or o3. The return of these options allows those workflows to continue unaffected, while enabling comparative testing of GPT‑5 in parallel.

Furthermore, GPT‑4o has been reinstated not only as a legacy option but, for some accounts, as the default model—especially in response to feedback noting that GPT‑5 responses, though more logically structured, were often described as emotionally distant, less conversational, or misaligned in tone. OpenAI has not only accepted this criticism but acted on it by restoring the option to select models that many found more natural or engaging for dialogue and instruction.

The updated interface turns ChatGPT into a modular system with transparent controls and tailored reasoning modes.

With these changes, ChatGPT is evolving from a static front-end into a configurable environment where the user has both vertical control (depth of reasoning) and horizontal control (model version and style). This direction significantly increases the platform’s adaptability, especially in professional settings where the ability to fine-tune how information is processed—and how much time or structure is invested in each answer—is critical.

Rather than forcing all users into a single optimized model, OpenAI is now presenting ChatGPT as a toolbox of cognitive styles: some versions are better at fast synthesis, others are stronger at methodical exploration, and legacy models may still outperform newer ones in tone, conversational rhythm, or task familiarity.

These interface updates reflect a wider philosophical shift: where the system once prioritized seamless abstraction and backend automation, it now restores user agency and choice, without requiring technical configuration or API-level understanding. Whether a user wants to delegate task handling to the system (Auto), accelerate output (Fast), slow down and think more carefully (Thinking), or return to a trusted past version (GPT‑4o), those options are no longer hidden—they are placed front and center, as an integral part of the interface.

This layered, transparent structure is likely to become the new foundation of how users interact with language-based systems: not as mysterious engines, but as adjustable instruments tailored to the intellectual weight of the question, the expectations of the domain, and the preferences of the user.

____________

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org