By Jose Antonio Lanz and Jason Nelson
7 min read
After weeks of speculation, OpenAI unveiled the next generation of its flagship AI model, GPT‑5. The new model, which creates a simpler interface to the world’s most popular AI platform, arrives amid mounting pressure from rivals like Meta, Anthropic, and Google.
"GPT is a significant upgrade over GPT-4 and a significant step along the path to AGI," CEO Sam Altman said, noting that among other things, GPT-5 can "write a software program from scratch. We think this idea of software on demand will be a defining characteristic of the GPT-5 era." The new model wrote over 400 lines of code in two minutes, in a "zero shot" example from a prompt asking it to do a simulation of the Bernoulli effect.
One unexpected disappointment: though the new model has a 400K token context window, which is about four times more than GPT-4, it still pales in comparison against Claude or Gemini, which can handle over 1 million tokens of context. This is significant because token-context-window size determines how much information a language model can “remember” and work with in a single interaction—things like documents, codebases, long conversations, or research papers.
Claude 3 and Gemini 1.5 already support over 1 million tokens, making OpenAI’s new model feel like it’s catching up, not leapfrogging. Claude can take in an entire book series, dense research archives, or huge code repositories and still give coherent, context-aware responses.
GPT-5 did offer a raft of new features, including a video option in which it can watch you while you chat (or anything else via your camera). Likewise, next week, the company will begin allowing users to integrate their Gmail and Google Calendars.
The new model will be available to all GPT users, including the free tier, starting today. Free users will be usage capped; after hitting a limit, they will be automatically rolled over to a "mini" model, the company said.
Altman positioned GPT-5 as a qualitative leap in AI capability during Thursday's livestream. "GPT-4 is like talking to a college student. Now, GPT-5 is like talking to an expert," he said, while talking about the company's ambition to move beyond incremental improvements toward fundamentally different AI interactions. Here are some of its new features announced today.
The company emphasized enterprise applications throughout the presentation. "It will be an especially important moment for businesses," Altman noted, highlighting GPT-5's enhanced reasoning capabilities as a potential transformer for corporate workflows and decision-making processes.Unified Reasoning Architecture.
OpenAI described reasoning as central to its artificial general intelligence strategy. "Reasoning is at the heart of our AGI program," the team explained. The breakthrough eliminates previous trade-offs between speed and analytical depth: "Until now users had to pick between slow answers using reasoning and fast answers with normal token prediction. This model eliminates that barrier."
The company touted comprehensive improvements across key metrics. "GPT-5 is faster, more reliable and more accurate" than its predecessor, according to internal testing. OpenAI also positioned the model as "the best coding model to date," with demonstrated software development capabilities that surpass previous iterations.The model also beats all the scores in different benchmarks shared by the company, especially in math and coding.
GPT-5 demonstrates significant improvements in truthfulness and accuracy compared to previous models. "We found GPT-5 is significantly less deceptive than o3," OpenAI reported during the presentation. The reduced hallucination rate addresses one of the most persistent challenges in large language model deployment, particularly for enterprise applications requiring factual precision.
OpenAI revealed a new training approach centered on synthetic data generation. "We are experimenting with a new set of techniques," the team explained, describing their methodology shift. "Our breakthrough was not to create more data but to create the right set of data," the company noted, indicating that the synthetic data used for pretraining delivers substantially better results than traditional data collection methods.
This is more a ChatGPT feature than a model improvement, but it could heavily impact the way people use ChatGPT. The new model features enhanced memory capabilities that allow users to customize their chatbot's behavior patterns. ChatGPT can now be tailored to specific goals or objectives, making interactions more helpful and aligned with individual user needs. This personalization extends beyond simple preference settings to include behavioral modifications that persist across conversations.
Agentic Coding Excellence
GPT-5 delivers exceptional performance in autonomous coding tasks, handling live implementation and iterative development with natural ease. "GPT-5 sets a whole new standard. It's the best model at agentic coding tasks," the team announced, with co-founder Greg Brockman emphasizing "It's incredible at front end... It's extremely good at instruction following."
The model eliminates the need to switch between specialized coding assistants, allowing users to rely on GPT-5 for all development phases from planning through debugging. OpenAI presented three variants—GPT-5, GPT-5 mini, and GPT-5 nano—all capable of handling tools, functions, and different programming formats seamlessly. It has also been trained to build apps that are more visually appealing and hhave better frontend understanding
GPT-5 introduces tool call explanations, allowing the model to describe its intended actions before executing them, making task execution more efficient and transparent for developers. Users can adjust verbosity settings to control output length and manage token costs while maintaining functionality. The model achieved top performance on SWE-bench, establishing GPT-5 as the most powerful model for software engineering tasks currently available.
OpenAI announced a tiered rollout strategy for the new model. Free tier users will start with GPT-5 before transitioning to a lighter "GPT-5 mini" version when they deplete their usage quota, while Pro subscribers receive unlimited access to the full model. The company confirmed that all existing ChatGPT features will remain compatible with the new system.
The company is deprecating its older models, and asked GPT-5 to write an obituary for them, which one of the presenters pointed out was more natural in its reponse, like talking to someone with a higher IQ and higher EQ she said: “It has more rhythm … It’s actually quite personal—and that’s the kind of thing GPT-5 does better... the responses feel less like AI." The model also has native multilingual capabilities, making it more capable than other models in multilingual tasks
ChatGPT currently boasts 700 million weekly active users, compared to Google Gemini’s 47 million weekly users. Despite its large user base, GPT-5 will be competing with Meta AI, whose own weekly user base stands at 700 million, thanks to its integration with Instagram and WhatsApp.
To date, OpenAI has raised $57 billion across 11 funding rounds, according to data from Tracxn, and is valued at $300 billion. According to a Reuters report, that valuation number is expected to rise to $500 billion if a proposed stock sale option goes through. The company doubled its revenue in the first seven months of the year and is expected to reach $20 billion by December 31, Reuters reported.
Altman first publicly confirmed that OpenAI started working on GPT‑5 in 2023, with the actual training phase beginning one year later. On February 13, 2025, Altman outlined the company’s roadmap and said GPT‑5 was expected to launch “within months,” with the aim of unifying OpenAI’s O‑Series and GPT models.
Altman has recently spent a great deal of time opining about artificial general intelligence, fueling speculation that OpenAI could use today’s event to announce a breakthrough. Earlier this month, however, he cautioned that there could be additional delays in the rollout of GPT, asking users to "please bear with us."
In the run-up to the launch earlier this week, OpenAI released two open-weight models on Tuesday—gpt‑oss‑120b and gpt‑oss‑20b—under an open license, aimed at developers seeking to run models locally.
Decrypt-a-cookie
This website or its third-party tools use cookies. Cookie policy By clicking the accept button, you agree to the use of cookies.