What’s New in GPT-5 by OpenAI

|
What’s New in GPT-5 by OpenAI

OpenAI’s GPT-5 isn’t just another “next version” drop; it’s the most capable ChatGPT yet, with big jumps in reasoning, coding, and writing. Officially launched in August 2025 after months of hype, it lands at a point when AI tools are sliding into everyday workflows for developers, writers, students, and entire teams.

The launch event wasn’t just slides and numbers. OpenAI’s team showed GPT-5 building full apps from scratch, answering complex research questions, and tackling multi-step agentic tasks without constant babysitting. Now, it’s the engine behind ChatGPT for Pro and Enterprise users, available via the API, and already powering developer favorites like Cursor.

You get three flavors to choose from — GPT-5, GPT-5 mini, and GPT-5 nano — each tuned for a different mix of horsepower and speed. Alongside that, there’s the new Study Mode, tighter Google Apps integration, and a massively expanded context window for conversations and projects that actually need to go long. Smarter? Definitely. But more than that, GPT-5 feels built to work the way people actually work.

GPT-5 Coding Capabilities

Overview & Benchmarks

GPT-5 is OpenAI’s strongest coding model yet, showing up in the numbers. On SWE-bench Verified, it hits 74.9% (o3 scored 69.1%) while using 22% fewer output tokens and 45% fewer tool calls. On the Aider Polyglot test, it clocks in at 88%, cutting the error rate by roughly a third versus o3.

Dev Feedback & Agentic Performance

Over in Cursor, the verdict is pretty loud: “the smartest model we’ve used.” It’s not just about speed—it’s how steerable it feels. You can nudge style, complexity, even “personality,” and it stays on track. Under the hood, GPT-5 is built for agentic coding: it chains tools, plans steps, and drives tasks proactively, thus it shines in setups like GitHub Copilot and Cursor.

Demos & “Software on Demand”

The launch demos were the headline moment. From a single prompt, GPT-5 spun up a fully functional French-learning web app—tracking, flashcards, a little “Snake”-style game, the works—within minutes. That’s the promise of software on demand: describe what you want, and get working code instead of a to-do list.

Vibe Coding & Creative Workflow

Call it vibe coding: you describe the goal in natural language, the model iterates, and you guide the direction rather than micromanaging every function name. Even Sam Altman leaned into this at launch—ask GPT-5 to “vibe code” what you need, and it will.

Multilingual & Full-Stack Strength

It holds up front to back. For frontend engineering (think UI-heavy work), testers preferred GPT -5’s output ~70% of the time over o3, citing better aesthetic sense and ambition. It also digs into deep codebase reasoning, helping you understand architectures, trace bugs, and refactor without losing the plot.

GPT-5 in Cursor

If you hang out in dev circles, you’ve probably heard the buzz: GPT-5 in Cursor isn’t just an upgrade, it’s a whole new vibe. Developers testing it are calling it “the smartest model we’ve used,” and not just because it spits out code fast. It actually gets what you’re asking for, even if your prompt is more “half-baked thought” than fully fleshed-out spec.

Here’s the magic: GPT-5 fills in the blanks, asks for less hand-holding, and keeps the conversation moving even when you’re knee-deep in a tricky debugging session. Cursor’s own team says it feels more “steerable” than anything before — you can nudge its style, complexity, and even its personality without it wandering off into unrelated code.

In practice? It’s like pairing up with that one teammate who always knows where the project’s going, catches your typos before you do, and never loses patience. Whether you’re spinning up a new feature, untangling legacy code, or chasing down a weird bug, GPT-5 in Cursor doesn’t just write — it collaborates. And for many devs, that’s the difference between a tool you use and a partner you trust.

Agentic Tasks

One of GPT -5’s biggest glow-ups is how it handles agentic tasks — those multi-step jobs where the AI has to think ahead, grab the right tools, and adjust on the fly. Instead of you micromanaging every click, GPT-5 breaks your request into bite-sized steps, figures out the smartest order to tackle them, and… gets on with it.

At the launch event, this played out in real time. A presenter asked for an app from scratch, and GPT-5 went to work — backend, frontend, external data pulls, even polishing the interface — all without the user spelling out every detail. The τ²-bench puts it in the mid-90s for reasoning and tool chaining, which is a fancy way of saying it’s really, really good at juggling complex tasks without losing its place.

It feels less like chatting with a bot and more like delegating to a capable junior dev or research assistant — one who keeps the project moving while you focus on the bigger picture. Hook it into tools like Cursor or your go-to productivity apps, and it can bounce between coding, data crunching, and scheduling without dropping the thread.

Writing & Creative Ability

GPT-5 isn’t just flexing its logic muscles — it’s got a serious creative streak too. Compared to GPT-4o, the writing feels more… human. The pacing is smoother, the tone shifts more naturally, and it can slip into whatever voice you need without losing clarity. Want a breezy blog post? A tight grant proposal? A moody scene for your screenplay? It can shift gears effortlessly.

Testers say it handles long-form work like a pro, keeping threads connected over thousands of words without drifting off course. Ask for poetry, and it drops the stiff, fill-in-the-blank style of older models. Request fiction, and you’ll get characters with actual depth. Even dense technical docs read cleaner, with a flow that makes them easier to follow.

The secret sauce is in how it now follows nuanced directions. You can say, “make it playful but still persuasive” or “explain it for beginners without dumbing it down,” and it actually listens. The result isn’t a cold, generic output — it’s more like having a co-writer who understands not just what you’re saying, but the vibe you’re going for.

Model Variants: GPT-5, Mini, and Nano

Not every project needs the same horsepower, so OpenAI split GPT-5 into three flavors—each tuned for a different balance of speed, cost, and capability.

Not every project needs the same horsepower, so OpenAI split GPT-5 into three flavors—each tuned for a different balance of speed, cost, and capability.

  • GPT-5 (standard) – The full-strength model. This is the one you reach for when you need heavy-duty reasoning, complex coding, or long-form creative work. It’s the default for ChatGPT Pro and Enterprise, and it’s available via the API.
  • GPT-5-mini – The faster, lighter, and more budget-friendly sibling. Perfect for well-structured prompts, shorter conversations, and situations where low latency matters more than raw muscle.
  • GPT-5-nano – The featherweight champion. Built for ultra-low latency and tight resource budgets, it’s ideal for mobile apps or massive API call volumes where every millisecond (and cent) counts.

Under the hood, they share the same core architecture. The difference comes down to computational scale and training scope—letting you pick the one that fits your workload without overpaying for performance you don’t need.

Study Mode (New)

Launch & Goals

Released between July 29–31, 2025, GPT -5’s Study Mode shifts ChatGPT from a straight answer engine into something more like a personal tutor. It’s built around step-by-step guidance, gentle hinting, and the Socratic method—nudging you to think through problems instead of just handing over the solution. The aim: deeper learning, not quick fixes.

Features & Usability

You can turn it on from Tools → “Study and learn” in ChatGPT. Once enabled, it personalizes its approach using your memory and conversation context (if memory is on), making lessons feel more tailored. It’s versatile too—capable of running quizzes, flashcards, mind maps, or open-ended reflection prompts to reinforce understanding.

Comparisons & Feedback

In direct tests against Google’s NotebookLM, ChatGPT’s Study Mode scored points for active, learner-specific engagement—think adaptive quizzes and conversational guidance—while NotebookLM stood out for strict document fidelity. Education experts like that Study Mode promotes “productive struggle” and critical thinking, keeping learners engaged rather than passively consuming answers.

Emerging Competition

Google wasn’t far behind. Its Guided Learning feature in Gemini launched soon after, with a similar focus on building conceptual understanding over simple Q&A.

Google Apps Integration

Smart Email & Calendar Assistant

One of GPT -5’s most practical new tricks is its deep integration with Google’s core apps—Gmail and Google Calendar. Soon, ChatGPT (powered by GPT-5) will be able to connect directly to your Gmail, Google Contacts, and Calendar from inside the chat window. The model won’t just respond to prompts—it will know when an email or schedule entry is relevant and pull it in automatically, making conversations feel more connected and context-aware.

Early Beta Features & Testing

During early testing, users could:

  • Summarize long email threads and draft polished responses
  • Turn emails into quick story outlines or actionable to-do lists
  • Create new calendar events without leaving the ChatGPT window

Practical Use Cases

This changes the game for anyone juggling a packed inbox and a busy schedule. You could compose a reply, lock in a meeting time, and get a follow-up reminder all through one conversation. There is no tab-hopping, no manual copy-paste, just smooth, uninterrupted flow.

Security & Permissions

As with other OpenAI connector integrations, access is fully permission-based. You’ll need to explicitly link your Gmail and Calendar, and OpenAI says that your data won’t be used for model training.

By blending AI assistance with real-time access to your communications and schedule, GPT-5 moves ChatGPT from being a helpful standalone tool to a genuine productivity partner—one that can manage your inbox and calendar without breaking the conversational rhythm.

Model Specifications & Pricing

Context Window & Token Limits

GPT-5 comes with a massive context window that changes what’s possible in a single conversation:

  • Input tokens: up to 272,000
  • Output tokens: up to 128,000 (reasoning + generated text)
  • Total context: 400,000 tokens per interaction—enough to handle book-length documents, sprawling codebases, or hours-long discussions without losing the thread.

Pricing (per 1M tokens)

  • GPT-5: Input $1.25 | Output $10
  • GPT-5-mini: Input $0.25 | Output $2
  • GPT-5-nano: Input $0.05 | Output $0.40

This tiered approach lets you choose between raw reasoning power, ultra-low latency, or a cost-friendly balance, depending on your workload.

Performance Benchmarks

Beyond limits and pricing, GPT -5’s performance holds up under testing. On SWE-bench Verified, it scores 74.9%, and on Aider Polyglot, it hits 88%, showing strong results in both English and multilingual coding challenges.

The takeaway: whether you need deep, sustained reasoning or lightweight responsiveness, GPT-5 gives you the flexibility to pick the right model without locking yourself into a single capability set.

GPT-5 vs Gemini 2.5 Pro vs Claude 4.1 vs Grok 4

Analysis

  • GPT-5 leads in official coding benchmarks (SWE-bench, Aider Polyglot) and offers a strong balance of reasoning and creative writing.
  • Gemini 2.5 Pro holds the largest official context window and performs well in coding according to independent tests.
  • Claude 4.1 emphasizes safety and reasoning depth; strong in some third-party SWE-bench runs but lacks official benchmark disclosure.
  • Grok 4 Heavy stands out in third-party academic benchmarks like HLE and ARC-AGI-2 through multi-agent reasoning.

Additional Highlights

While GPT -5’s headline wins are in coding, writing, and integrations, it also ships with a set of quieter but equally meaningful improvements that expand what it can do day-to-day:

  • Multimodal Skills – GPT-5 can now handle text, images, and audio in supported environments. That means coding from diagrams, breaking down complex charts, or even walking you through visual examples in Study Mode.
  • Improved Safety & Alignment—Updated training reduces unnecessary refusals and strengthens factual accuracy, especially for sensitive or nuanced topics. The result is output that’s both more helpful and more trustworthy.
  • Persistent Memory – With memory enabled, GPT-5 remembers past sessions—your preferences, prior work, and relevant context—so it feels less like starting from scratch each time and more like working with someone who knows your style.
  • Expanded Tool Access – Beyond Gmail and Calendar integration, GPT-5 connects to an upgraded toolkit, from deeper file analysis to a sharper web browser, extending its reach into research, project planning, and customer support.
  • Reasoning Benchmarks – Scores on MMLU-Pro and GPQA Diamond show clear jumps over GPT-4o, underscoring that its reasoning upgrades are as strong as its coding and agentic gains.

Together, these changes position GPT-5 as more than a conversational AI—it’s a versatile, multimodal assistant ready to handle everything from complex professional projects to imaginative creative work.

Conclusion

The GPT-5 launch makes one thing clear: AI development isn’t slowing—it’s accelerating into a high-stakes race. With Gemini 2.5 Pro, Claude 4.1, and Grok 4 all pushing their own strengths, competition is driving rapid gains in reasoning, context handling, and integration.

That pace comes with a price. Running these large models demands huge infrastructure investments, and the cost pressures are showing up in plan pricing and usage rates as companies work to keep services both accessible and sustainable.

Even so, the upside is hard to ignore. GPT -5’s sharper coding skills, fluid writing, and self-directed agentic task handling—combined with the flexibility of its three variants—make it a model built for more than conversation. With multimodal abilities, refined safety and alignment, persistent memory, an expanding tool set, and standout reasoning benchmarks, it’s closer to a foundation for building whatever users can imagine.

Paired with APIs, productivity integrations, and the rise of AI agents, GPT-5 has the potential to quietly reshape entire industries—from software development and education to customer support and research. The building blocks are here; the next chapter will be written by what people create with them.

Share This Article