The company said the new tools are designed to make voice-based apps more useful in everyday situations, especially where users need software to respond naturally while carrying out tasks in real time.

At the centre of the launch is GPT-Realtime-2, a voice model OpenAI says can handle more difficult requests while keeping conversations flowing naturally.

Unlike earlier versions, the company said the model uses GPT-5-level reasoning to manage interruptions, understand context better and carry out actions during conversations.

OpenAI also unveiled GPT-Realtime-Translate, a live translation model that can translate speech from more than 70 input languages into 13 output languages.

According to the company, the system keeps pace with the speaker during conversations instead of translating after pauses or completed sentences.

The third model, GPT-Realtime-Whisper, focuses on live speech transcription. It converts spoken words into text instantly while a person is talking.

“Together, the models we are launching move real-time audio from simple call-and-response toward voice interfaces that can actually do work: listen, reason, translate, transcribe, and take action as a conversation unfolds,” OpenAI said.

Voice products have become a huge focus for technology companies as more users interact with software through speech instead of typing. OpenAI said developers want systems that can manage tasks while conversations continue naturally.

The company pointed to customer support, travel, education, media and creator platforms as some of the areas expected to benefit from the new models.

OpenAI also described three growing patterns it sees in voice-based software.

The first is “voice-to-action”, where users speak naturally and the system completes tasks on their behalf. OpenAI said property platform Zillow is building an assistant that can help users search for homes, avoid certain neighbourhood conditions and book tours through voice requests.

Another pattern is “systems-to-voice”, where software provides spoken updates automatically. OpenAI gave the example of travel apps that could alert passengers about delayed flights, new boarding gates or transfer routes without users typing commands.

The third area is “voice-to-voice”, which focuses on live multilingual conversations. OpenAI said Deutsche Telekom is developing customer support systems that translate discussions instantly while both sides continue speaking in their preferred languages.

Travel company Priceline is also working on voice-based trip management tools, according to OpenAI. Travellers could eventually book flights, change hotel reservations and receive airport updates entirely through conversation.

Alongside the broader rollout, OpenAI added several new features to GPT-Realtime-2 aimed at improving live interactions.

Developers can now enable short phrases such as “let me check that” or “one moment while I look into it” before the system completes a request. OpenAI said this gives users clearer feedback while the model processes tasks in the background.

The model can also call multiple tools at once and explain those actions aloud during conversations. OpenAI said the system may say things like “checking your calendar” or “looking that up now” while working through requests.

The company added that GPT-Realtime-2 recovers better from errors or failed requests instead of stopping conversations abruptly. It also supports a larger context window, increasing from 32K to 128K, allowing longer and more detailed conversations.

OpenAI further noted that the model has improved understanding of specialised terms, including healthcare vocabulary and proper nouns. Developers can also adjust how much reasoning power the model uses depending on the complexity of a request.

According to benchmark figures released by the company, GPT-Realtime-2 achieved higher scores than GPT-Realtime-1.5 on audio intelligence and instruction-following tests.

All three models are available through OpenAI’s Realtime API. The company said GPT-Realtime-Translate and GPT-Realtime-Whisper will be billed by the minute, while GPT-Realtime-2 pricing depends on token usage.

OpenAI said it has added safeguards to reduce misuse, including protections against spam, fraud and harmful content. The company also added that conversations can be stopped automatically if they break its safety rules.

Users condemned GPT-5 for its “cold,” “mechanical,” and “less empathetic” tone, leading to threats of subscription cancellations. In response, OpenAI restored GPT-4o to the model picker for all paid users.

CEO Sam Altman confirmed that GPT-5’s personality will be revised. “We are working on an update to GPT-5’s personality which should feel warmer than the current personality but not as annoying (to most users) as GPT-4o,” he wrote on X. 

He added that one major takeaway from the rollout is the importance of “more per-user customization of model personality,” allowing individuals to tailor the AI’s tone and style to their preference.

OpenAI has also introduced three distinct modes for GPT-5: Auto, the default setting for general use; Fast, prioritising speed; and Thinking, designed for deeper reasoning with a context window of 196,000 tokens and a weekly message limit of 3,000 for paid users. After hitting this limit, users can continue on GPT-5 Thinking mini, a lighter version with reduced compute requirements.

In addition to restoring GPT-4o, OpenAI has made several older models accessible via the “Show additional models” toggle in ChatGPT settings. These include GPT-4.1, GPT-03, and GPT-5 Thinking mini. GPT-4.5 is still reserved for Pro users due to its heavy GPU demands.

Altman noted that the quick reinstatement of these models reflects the emotional attachment users have developed to specific AI personalities, an aspect OpenAI had underestimated. 

The company plans to allow even more granular personality customization, with potential options ranging from “Cynic” and “Listener” to “Robot” or “Nerd,” along with the possibility for fully custom tones and behaviours.

“These updates mark a shift in AI development,” Altman added. “It’s no longer just about raw capability; user experience and emotional resonance are becoming central. We’re not always going to get everything on try #1, but I’m proud of how quickly the team can iterate.”

OpenAI is working to balance powerful AI functionality with a more engaging and personally tailored user experience.

The company has built this model as a tool for both conversation and solving real enterprise problems in sectors like software engineering, financial services, and healthcare.

From day one, GPT-5 will replace its predecessors across ChatGPT platforms, and it’s being offered with expanded access tiers, including free usage, a $20/month Plus subscription, and a $200/month Pro tier. 

Developers also now have access through the OpenAI API, with three variants: GPT-5, GPT-5-mini, and GPT-5-nano. These versions differ in how long they spend “thinking” about problems, with pricing ranging from $1.25 to $10 per million tokens.

While consumer interest in AI remains high, enterprise adoption has been slower. OpenAI is hoping GPT-5 can tip the scale. The model delivers what CEO Sam Altman called “software on demand,” capable of building fully functional apps from natural language prompts. 

“GPT-5 is really the first time that I think one of our mainline models has felt like you can ask a legitimate expert, a PhD-level expert, anything,” Altman said during a press conference.

Behind the launch, OpenAI has struggled with the sheer technical demands of scaling up its models. The company has faced hardware failures during training, hit limits in the availability of new high-quality training data, and spent months waiting for results from high-cost training runs. 

At the same time, it has had to justify its skyrocketing costs, including investor expectations built on a potential $500 billion valuation and signing bonuses of up to $100 million for top AI talent.

Back then, when GPT-4 was launched, it passed a simulated bar exam in the top 10%, compared to GPT-3.5’s performance in the bottom 10%. With GPT-5, the upgrades are more subtle but targeted.

Take code generation. On SWE-bench Verified, a real-world benchmark for software engineering tasks, GPT-5 scored 74.9% on first attempts, outperforming Claude Opus 4.1 from Anthropic (74.5%) and Gemini 2.5 Pro from Google DeepMind (59.6%). In healthcare, its error rate on HealthBench Hard Hallucinations is 1.6%, far below GPT-4o’s 12.9%.

In science, GPT-5 Pro achieved 89.4% accuracy on PhD-level science queries, slightly ahead of rivals from xAI and Anthropic. But it lags in other areas, including real-time web navigation tasks. On the Tau-bench airline site navigation test, GPT-5 scored 63.5%, slightly behind OpenAI’s earlier o3 model (64.8%).

Despite these nuanced results, OpenAI insists the model is “safer, smarter, and more useful.” Alex Beutel, the company’s lead on safety research, said GPT-5’s reduced deception rates are essential for building trust. 

“It’s more transparent and honest in ways users can trust,” he said, adding that GPT-5 also more reliably filters out harmful queries while reducing false positives, unnecessary content rejections.

From a usability standpoint, GPT-5 also comes with new personalisation features. Users can now select from four built-in personalities, Cynic, Robot, Listener, and Nerd, which adjust the tone and structure of responses. Unlike earlier models, users no longer have to manually tweak settings to get different types of output.

Internally, OpenAI believes GPT-5 represents a shift in how people will use AI, not just to answer questions, but to act more like agents or assistants. 

That includes handling schedules, creating research briefs, analysing financial documents, and building apps from scratch. “This idea of software on demand is going to be one of the defining features of the GPT-5 era,” Altman said.

But while the ambition is high, there’s still caution among experts. Some reviewers told Reuters they weren’t convinced GPT-5 is a major leap over GPT-4. Others, like Noah Smith, raised concerns about the financial sustainability of current AI development. 

“Business spending on AI has been pretty weak, while consumer spending has been fairly robust because people love to chat with ChatGPT,” he said. “But the consumer spending on AI just isn’t going to be nearly enough to justify all the money that is being spent on AI data centres.”

Altman himself admitted GPT-5 still has limitations, especially around independent learning. It cannot, on its own, acquire new knowledge or skills without user input. And while test-time compute (a method of giving the model more thinking power when needed) helps in solving complex problems, it’s not a substitute for self-directed learning.

Still, the company believes in its innovation. With over 700 million weekly ChatGPT users and increasing partnerships with enterprise customers, GPT-5 may help OpenAI bridge the gap between consumer curiosity and business utility.