OpenAI just released GPT5, their newest language model, and while the company is celebrating it as a major advancement, I’m left with mixed feelings. After watching their hour-and-a-half launch stream, I can’t help but wonder if we’ve reached a point where incremental improvements are becoming the norm rather than revolutionary leaps.
The timing of this release is particularly interesting. Just days ago, OpenAI released GPT4 OSS as an open-weight model available to everyone for free. Now we know why – they were clearing the runway for their new flagship product.
What’s New in GPT5?
OpenAI describes GPT5 as having “PhD-level intelligence” across virtually every domain – a significant step up from GPT4’s “college student” level smarts. The model has crushed every benchmark it’s been tested against, including:
- 100% score on competition math benchmarks (with Python access)
- 96.7% without tools
- 100% on the Harvard-MIT Mathematics Tournament (with Python)
- 89% on PhD-level science questions
- 74.9% on software engineering benchmarks (slightly beating Anthropic’s Claude Opus 4.1)
Beyond raw intelligence, GPT5 brings several practical improvements. The model now automatically determines how long it needs to think about a problem, eliminating the need to choose between different versions. It’s also reportedly much faster and better at picking up subtle details in complex prompts.
Perhaps most importantly for everyday users, GPT5 is available to everyone – even those on free plans. Plus subscribers get more usage, while Pro subscribers get access to GPT5 Pro with extended reasoning capabilities.
New Features Worth Noting
Several new features caught my attention during the presentation. OpenAI is integrating Gmail and Google Calendar access, allowing the model to function more like a true assistant by managing your schedule and emails. This feature is rolling out next week, starting with Pro users.
The voice feature is now available to all users, even those on free plans, with paid subscribers getting nearly unlimited access. The voice experience is also more customizable – you can instruct it to give one-word answers or help you learn a new language.
Personality customization is another interesting addition. You’ll be able to choose between different personalities like “concise and professional,” “thoughtful and supportive,” or even “a bit sarcastic.” The initial options include cynic, robot, listener, and nerd.
The Demos Were Good, Not Great
The demos shown during the launch were impressive but not mind-blowing. They created an SVG animation explaining the Bernoulli effect, coded a French mouse game (basically Snake with cheese), built a financial dashboard, and developed a 3D castle game with interactive elements.
What struck me most wasn’t that these tasks were impossible before, but that they could now be accomplished with a single prompt. The model seems to require less back-and-forth to achieve complex results.
When I tested GPT5 myself with a prompt to create a Vampire Survivors clone, it produced a functional game with about 565 lines of code in just three minutes. The game featured different enemy types, level-ups, and weapon upgrades – all from one prompt. This is genuinely impressive and shows real improvement in coding capabilities.
Safety Improvements
OpenAI emphasized safety improvements in GPT5, claiming significant reductions in hallucinations (0.7% vs. 4.5% in previous models) and deception. The model is also better at understanding intent, distinguishing between malicious requests and legitimate ones.
Rather than simply refusing to help with potentially problematic requests, GPT5 will more often point users to ethical, legal resources – a more helpful approach than the frustrating “I’m sorry, I can’t help with that” responses we’ve grown accustomed to.
Why I’m Underwhelmed
Despite all these improvements, I found myself underwhelmed by the launch. The leap from GPT3.5 to GPT4 felt revolutionary – we saw capabilities we didn’t even know were possible, like drawing websites on paper and having them built with code.
This release feels more like the incremental improvement we saw from GPT4 to GPT4o. I was expecting something truly mind-blowing – perhaps Sora integration for video generation, or dramatic improvements to the image generator.
I also wish they had addressed some persistent limitations. The model still struggles to truly mimic writing styles when asked to create content that sounds like a specific person. The memory features, while improved with Gmail and Calendar integration, don’t address the need for project-specific memory or longer-term recall.
What’s becoming clear is that the major AI companies have identified coding as the killer use case for large language models. Anthropic leaned into this with Claude, and now OpenAI is following suit. If a model can write code well, it can theoretically solve almost any problem – even if it can’t solve it directly, it can write code that does.
GPT5 is undoubtedly impressive and will be useful for many people, especially developers. But for the average person looking for help with everyday tasks, this update may not be the game-changer we were hoping for. The revolution continues, but perhaps at a more measured pace than we’ve come to expect.
Frequently Asked Questions
Q: Who can access GPT5 and when is it available?
GPT5 is available to all OpenAI users, including those on free plans. Plus subscribers get more usage allowance, while Pro subscribers get access to GPT5 Pro with extended reasoning capabilities. The model was released on August 7th, 2023, and is available immediately.
Q: How does GPT5 compare to previous models in terms of performance?
OpenAI describes GPT5 as having “PhD-level intelligence” compared to GPT4’s “college student” level. It scored 100% on several math benchmarks when using Python, achieved 96.7% without tools, and outperformed all previous OpenAI models across various tests. It even slightly outperformed Anthropic’s newest Claude Opus 4.1 on coding benchmarks.
Q: What new features come with GPT5?
GPT5 introduces several new features including Gmail and Google Calendar integration (rolling out next week), voice capabilities for all users, personality customization options, and a 400,000 token context window. The model also automatically determines how long it needs to think about a problem, eliminating the need to choose between different versions.
Q: How has GPT5 improved in terms of safety?
OpenAI claims GPT5 has significantly reduced hallucinations (0.7% vs. 4.5% in previous models) and deception. Instead of refusing to help with potentially problematic requests, it more often points users to ethical, legal resources. The model is also better at understanding user intent, distinguishing between malicious requests and legitimate ones.
Q: Is GPT5 available through the API for developers?
Yes, GPT5 is available immediately through OpenAI’s API in three versions: GPT5 (highest-end), GPT5 Mini (middle ground), and GPT5 Nano (smaller, faster). The API includes new parameters like “reasoning effort” to control how much the model thinks before responding and “verbosity” to control output length. It’s already integrated with coding tools like Cursor, Windsurf, JetBrains, and GitHub Copilot.
