On August 7, 2025, OpenAI took the wraps off GPT-5 — calling it the most capable model they’ve ever built and a major step toward expert-level AI. In Sam Altman’s words, GPT-3 felt like talking to a high school student, GPT-4 like a college student, but GPT-5 is “a legitimate PhD-level expert in anything, on demand.” It’s faster, more reliable, and more intuitive, and it’s now the default in ChatGPT for free and paid users alike, with an expanded API offering for developers.
The “think just enough” reasoning model
At the heart of GPT-5 is a shift in how it reasons. Previous models forced you to choose between quick, surface-level answers and slower, chain-of-thought “reasoning” modes. GPT-5 decides for itself when to pause and think, striking the balance between speed and depth. You can still force more deliberate reasoning with a “think hard about this” prompt or, for paid users, by selecting the dedicated GPT-5 Thinking model. This paradigm makes it feel less like you’re prompting a machine and more like you’re collaborating with an intelligent partner that knows when to slow down.
From vibe coding to serious engineering
The launch demos made it clear that GPT-5 is a leap forward for software development. In one segment, the model built a fully interactive French-learning web app — complete with flashcards, quizzes, a progress tracker, and a playful “mouse and cheese” mini-game — in minutes. In another, it constructed a moving SVG to illustrate the Bernoulli effect for a student’s physics homework. It scaffolds projects, writes modular, well-documented code, runs builds, fixes errors, and iterates — all without hand-holding. This is more than “vibe coding”; it’s autonomous, high-quality engineering that can ship production-ready features.
Multimodal, voice, and learning upgrades
GPT-5’s multimodal reasoning is its strongest yet, outperforming both previous OpenAI models and most human experts on visual reasoning benchmarks. Its voice mode sounds natural, can see through your device’s camera for context, and supports smooth language translation. A new “study and learn” mode guides you step-by-step through topics — in one demo, it practiced Korean café orders at slow, normal, and even hyper-fast native speed.
Reliability, safety, and “safe completion”
OpenAI has re-engineered its safety training. Instead of bluntly refusing or fully complying with sensitive prompts, GPT-5 uses “safe completion” to maximize helpfulness within safety constraints. That might mean giving partial, high-level answers and directing you to safer resources. This change reduces frustrating over-refusals while avoiding risky detail dumps. It also makes GPT-5 the most factual and reliable model OpenAI has ever shipped.
Personalization and memory
Personalization is deeper than ever. Paid users can set chat colors, adjust personality tone (supportive, concise, even lightly sarcastic), and — starting with Pro accounts next week — connect GPT-5 to Gmail and Google Calendar. The model can then help plan your days, find time for personal goals (like marathon training), and pull in relevant context from your email.
Health and humanity
One of the most compelling launch moments came from a couple sharing their experience using ChatGPT during a cancer diagnosis. GPT-5, they said, was faster, more thorough, and more context-aware than GPT-4 — not just explaining biopsy results but anticipating what wasn’t on the page, suggesting pending tests, and framing questions for their doctor. On HealthBench, an evaluation built with 250 physicians, GPT-5 scored higher than any previous model.
Industry and enterprise use cases
From life sciences to finance to government, GPT-5 is already in production. Amgen is using it to analyze complex clinical data in drug design. BBVA reports GPT-5 outperforms every other model they’ve tested for financial analysis, cutting work from three weeks to a few hours. Oscar Health calls it the single best model for mapping complex policy to patient conditions. And starting this month, two million U.S. federal employees will gain ChatGPT access, with GPT-5 as their default model.
Benchmarks in context
On SWE-bench, GPT-5 hits 74.9% (vs. 69.1% for O3). On Aider Polyglot, it scores 88%, showing strong multi-language coding ability. Its 97% on the tau² tool-calling benchmark more than doubles the previous best, signaling a major jump in agentic, multi-step workflows. Instruction following hits 99% on Kali, and HealthBench results lead the field. Combined with a 400K-token API context window, these aren’t just numbers — they’re enablers for real-world utility.
Developer and API details
GPT-5 comes in three API sizes — GPT-5, GPT-5 Mini, and GPT-5 Nano — priced to fit latency and budget needs. Base GPT-5 runs at $1.25 per million input tokens and $10 per million output tokens, with Mini and Nano offering lower-cost, faster options. Developers get new controls: reasoning effort (including a new “minimum” mode), verbosity (low/medium/high), regex/grammar-constrained outputs, and a free-form custom tools interface beyond strict JSON.
The first reactions
The launch sparked a wave of commentary online. Satya Nadella celebrated the moment on X, saying “it’s incredible to see how far we’ve come since” GPT‑4. Tom’s Guide noted smarter reasoning, reduced hallucinations, and upgraded features across writing, coding, and safety. TechAU highlighted developer access and pricing tiers—$1.25 per million input tokens. WIRED called GPT‑5 a “major milestone” toward AGI, emphasizing expert‑level capabilities and efficiency. Tom Warren reported on a leaked GitHub post describing enhanced agentic and code‑task strengths. Across Threads and other platforms, #GPT5 was filled with developer experiments and head-to-head comparisons (anecdotally observed).
Looking ahead
As Greg Brockman noted, GPT-5’s current abilities are “early ideas that will go much further.” The combination of reasoning, tool use, multimodality, and long-context handling points toward AI that can work alongside humans for hours, days, or even weeks on complex projects. For now, GPT-5 is here, it’s live, and it’s already changing how people build, learn, and make decisions.