OpenAI drops GPT-4: Smarter, multimodal, and passes the bar

Just when everyone got used to ChatGPT confidently inventing fake citations, OpenAI rolled out a model that casually aces human professional licensing exams.

OpenAI has officially launched GPT-4, its latest multimodal flagship. It accepts both text and image inputs to produce text outputs, and exhibits human-level performance across a wide spread of academic and professional benchmarks.

Crushing human exams and low-resource languages

The benchmark leap over GPT-3.5 is stark. On a simulated Uniform Bar Examination, GPT-3.5 scored around the bottom 10 percent of test-takers. GPT-4 scored in the top 10 percent. Without any test-specific training, it posted similar top-tier numbers on AP exams and Olympiad challenges.

While multimodal capability is the headline feature, it comes with a catch. Users can input diagrams, photos, or mixed text and graphics, but image input is currently in a closed research preview with a single partner. Text input is live now for ChatGPT Plus subscribers and developers on the API waitlist.

Language coverage also saw a massive upgrade. OpenAI translated the MMLU benchmark across 57 subjects into dozens of languages. In 24 of the 26 languages tested, GPT-4 outperformed the English-language performance of GPT-3.5, showing remarkable competence even in low-resource languages like Welsh, Latvian, and Swahili.

Behind the curtain, OpenAI rebuilt its entire deep learning stack with Microsoft Azure. After treating GPT-3.5 as a test run, GPT-4 became OpenAI's first large model whose training run was completely stable and whose final performance was accurately predictable before training finished. Alongside the model, OpenAI also open-sourced its OpenAI Evals framework to let the community test and report model weaknesses.

Developers also gain real control via system messages. Instead of fighting the fixed default persona of ChatGPT, builders can now steer the tone, task, and boundaries directly within API prompts.

Six months in alignment jail

OpenAI spent six full months iteratively aligning GPT-4 after its pretraining. They partnered with more than 50 external domain experts in cybersecurity, biorisks, and trust and safety to conduct adversarial testing.

Those guardrails paid off on internal metrics. GPT-4 is 82 percent less likely to respond to requests for disallowed content and scored 40 percent higher on factual accuracy tests compared to GPT-3.5.

Yet, familiar flaws remain. GPT-4 still hallucinates facts, reasoning errors persist, and its knowledge remains cut off at September 2021. OpenAI pointed out an intriguing drawback: while the raw base model had great confidence calibration, the RLHF post-training process actually degraded it. GPT-4 can now be confidently wrong, making it risky to trust blindly in high-stakes settings.

Why it matters

GPT-4 turns language models from a fun conversational demo into genuine enterprise infrastructure.

For everyday users and builders, the difference shows up when tasks get hard. While standard chat feels marginally better, complex logic, code generation, and multi-step reasoning are in an entirely different class. The monthly 20 dollar Plus subscription is practically a no-brainer for professionals, and developers finally have steerable system messages to build real agents.

For the tech industry, OpenAI has just raised the bar to a punishing height. The fact that OpenAI can now predictably scale models on Azure infrastructure means brute-force compute plus disciplined alignment is winning the race, forcing Google and open-source rivals to play catch-up.

OpenAI's approach is both shrewd and cautious. Withholding public image input avoids immediate PR firestorms, while a six-month red-teaming sprint buys immense goodwill with regulators. But the underlying hazard is real: making a model 40 percent more factual also makes its remaining hallucinations far harder to spot. In high-stakes coding or legal review, unearned confidence is the most dangerous bug of all.