After two years of anticipation, OpenAI has finally launched GPT-5, and it’s full of surprises. Their most advanced AI model yet is now available to all ChatGPT users, even those on the free plan.
The new system introduces a layered reasoning architecture and marks a shift in how users interact with AI for writing, coding, health, and multimodal tasks.
“GPT‑5 is a significant leap in intelligence over all our previous models,” OpenAI said in the official announcement. The model is now the default option in ChatGPT, replacing GPT-4o and all earlier variants.
OpenAI CEO Sam Altman compared GPT-5 to having instant access to a group of PhD-level experts.“People are limited by ideas, but not really the ability to execute, in many new ways,” he said.
One system, multiple layers of reasoning
GPT-5 introduces a unified system that includes a standard model for most questions, a deeper reasoning layer (GPT‑5 thinking) for complex prompts, and a real-time router to select the appropriate approach. The router decides based on input type, complexity, and whether users explicitly request more intensive thinking.
Users can trigger deeper analysis by providing inputs like ‘think hard about this’ in the prompt, OpenAI said. If usage limits are hit, users transition to GPT-5 mini, a lighter version designed to maintain performance.
Stronger in coding, writing, and health
The company claims that GPT-5 shows notable gains in three major application areas for ChatGPT users, such as software development, creative writing, and health advice.
In software, GPT-5 handles complex front-end generation, large-scale debugging, and aesthetically aligned UI design with just one prompt.
During a briefing, OpenAI showcased GPT-5’s ability to perform vibe coding -generating software from a simple written prompt.
To illustrate this, the team asked GPT-5 to build a web app to help English speakers learn French. The app needed to include an engaging theme, flash cards, quizzes and a feature to track daily progress. When the same prompt was entered into two separate GPT-5 windows, the model produced two distinct apps within seconds.
For writing, OpenAI says the model can maintain literary structure in free verse and assist with day-to-day documents such as reports and memos. When comparing responses to poetry prompts, OpenAI noted GPT-5’s ability to “land the larger emotional arc with a stronger ending,” using concrete metaphors and a sense of place.
In health-related tasks, the model ranks highest on the HealthBench benchmark and is described as more proactive, precise, and context-aware. “It acts more like an active thought partner,” OpenAI said, though it cautioned that ChatGPT is not a replacement for medical professionals.
All About Benchmarks
GPT-5 has achieved state-of-the-art results across a range of benchmark tests, reflecting broad improvements in reasoning, accuracy, and task handling. In mathematics, the model scored 94.6% on the AIME 2025 benchmark without using external tools.
For coding tasks, it reached 74.9% on SWE-bench Verified, indicating strong capabilities in real-world software engineering scenarios. In visual reasoning, GPT-5 recorded an 84.2% score on the MMMU benchmark.
However, it still lags behind Grok-4 on the ARC-AGI benchmark. “Grok 5 will be out before the end of this year, and it will be crushingly good,” said xAI CEO Elon Musk. Meanwhile, GPT-5 only slightly outperforms Claude Opus 4.1, which scored 74.5% on SWE-bench Verified.
The company also made an error in one of the benchmark charts, which drew widespread attention and mockery on social media.
“GPT-5 is the smartest model we’ve ever done, but the main thing we pushed for is real-world utility and mass accessibility/affordability,” said Altman in a post on X. “We can release much, much smarter models, and we will, but this is something a billion+ people will benefit from.”
“With GPT-5, the benchmarks are misleading or don’t capture some of what it does well. You must try it!” said AI influencer Varun Mayya, who had early access to the model.
GPT-5’s performance in the health domain was also notable, achieving 46.2% on the HealthBench Hard benchmark. The model demonstrated improved instruction-following ability with a 69.6% score on Scale MultiChallenge.
GPT-5 Pro, the more advanced variant, attained 88.4% on the GPQA benchmark, highlighting its strength in scientific reasoning.
The GPT-5 Pro version, available to Pro-tier subscribers, offers extended reasoning and scored higher than experts in nearly half of 1,000 real-world knowledge work evaluations. “GPT‑5 Pro made 22% fewer major errors,” OpenAI said.
More accurate, less deceptive
GPT-5’s responses are ~45% less likely to contain factual errors than GPT‑4o. When the reasoning layer is active, the hallucination rate drops by ~80% compared to OpenAI o3. The model is also more honest when faced with underspecified or impossible tasks.
For instance, in a benchmark where prompts referenced missing images, GPT‑5 refused to speculate 91% of the time. OpenAI reported that deception in production traffic was reduced to 2.1%, less than half the rate seen in earlier models.
Customisation and personality features
Users can now choose from four preset personalities—Cynic, Robot, Listener, and Nerd—in a new feature designed to steer how ChatGPT responds. These options reflect GPT‑5’s improved steerability and reduced sycophancy.
“In targeted evaluations, sycophantic replies dropped from 14.5% to less than 6%,” OpenAI stated.
Access tiers and usage limits
GPT-5 is being rolled out to Free, Plus, Pro, and Team users starting today. Enterprise and education customers will gain access within a week. Pro users receive unlimited access to both GPT-5 and GPT‑5 Pro, while Plus and Team tiers have higher limits than Free users.
Once Free users reach their cap, the system defaults to GPT-5 mini.
Moreover, OpenAI is making three versions of the GPT-5 model available to developers through its API—gpt-5, gpt-5-mini and gpt-5-nano. Each version is built to meet different cost and latency requirements.
Earlier this week, the company also released two open-weight language models for the first time since GPT-2 in 2019. These models are intended to offer lower-cost options that developers, researchers and companies can run and customise.
Focus on safety and dual-use constraints
GPT-5 also includes a new safety paradigm: safe completions. Rather than refusing answers outright, GPT-5 is trained to offer partial or abstracted answers where appropriate, especially for dual-use domains such as virology.
For high-risk domains like biological research, GPT-5 Pro is classified as “High capability” under OpenAI’s Preparedness Framework and includes multiple layers of safety checks.
The post OpenAI’s GPT-5 Challenges Claude Opus 4.1, Lags Behind Grok-4 appeared first on Analytics India Magazine.
