GPT-4 raises the bar

The release and the narrow claim

On 14 March 2023, OpenAI announced GPT-4 and described it as a "large multimodal model" that accepts image and text inputs and produces text outputs, positioning it as more reliable, more creative, and better at handling nuanced instructions than its predecessor. OpenAI's launch page framed the advance in two registers at once: a set of measurable gains on academic and professional exams, and a general promise of broader usefulness across writing, coding, and analysis. The most-quoted figure was that GPT-4 scored in roughly the top 10% of test takers on a simulated bar examination, where GPT-3.5 had landed near the bottom decile. That is a task-specific result on one exam under stated conditions, not a measure of legal competence and certainly not a measure of intelligence in general.

Two things about the event should be held apart from the outset. The narrow, documented claim is that a single system posted strong scores across many standardized tests and several capability benchmarks, and that it was released as a product and a developer API. The broad historical judgment, that GPT-4 marked the moment general-purpose language models became credible for everyday professional work, is an interpretation assembled from those numbers plus a year of downstream deployment and third-party study. Keeping the two apart is the whole discipline of reading this event well, because the launch materials themselves blend the measured result and the aspiration into a single sentence.

What ChatGPT set up and what GPT-4 addressed

GPT-4 did not arrive in a vacuum. Its direct lineage runs through GPT-3, the 2020 model that showed a large autoregressive network could perform many tasks from a few in-context examples without task-specific retraining, and through the alignment work that turned such models into usable assistants. The pivotal intermediate step was InstructGPT, which used reinforcement learning from human feedback (RLHF) to make a base model follow instructions and refuse unsafe requests more usefully. ChatGPT, launched in late 2022, packaged that recipe into a conversational interface that reached a mass audience within weeks, which is what made a successor model a widely anticipated event rather than a research note.

The constraint GPT-4 set out to address was not the existence of a capable model but its dependability. GPT-3-class systems reasoned inconsistently, followed complex multi-step instructions poorly, drifted off task, and were hard to steer toward a particular tone or format. The launch materials describe gains precisely in those areas: better performance on tasks requiring reasoning, improved steerability through a developer-set "system" message that fixes the assistant's behavior, and a longer context window (a standard variant of 8,192 tokens and an extended variant of 32,768 tokens) that let the model attend to far more input at once. Framed this way, the problem GPT-4 attacked was reliability and control, not the invention of a new capability.

How the model actually works

At its core GPT-4 is trained on the same principle as its predecessors: it is a Transformer-style network trained to predict the next token over a large corpus, then aligned to human intent afterward. The GPT-4 Technical Report states that the post-training alignment used RLHF, following the InstructGPT method, and adds rule-based reward models to shape when the assistant should refuse or comply. Crucially, the report is explicit that alignment improves how faithfully the model follows intent and stays within policy, but does not by itself raise raw capability on the underlying exams; the base model's knowledge and the aligned model's behavior are separate axes.

Two methodological points deserve emphasis. First, GPT-4 is multimodal in that it can accept images as input, but that capability shipped as research and limited-access work at launch, for example through a partnership with the assistive application Be My Eyes, rather than as a generally available feature everyone could use on day one. Second, the report foregrounds predictable scaling: OpenAI built infrastructure that let them forecast GPT-4's final training loss, and some capability metrics such as coding pass rates, from far smaller runs using as little as one ten-thousandth of the compute. That forecasting is more than a curiosity. It converts frontier training from a single expensive gamble into something closer to an engineering plan with a projected outcome, which changes how such projects can be resourced and justified.

The benchmark evidence, read carefully

The report's evidence comes in two families. The first is exams designed for humans: the Uniform Bar Exam, the LSAT, the SAT, the GRE, and numerous Advanced Placement subject tests. GPT-4 improved on GPT-3.5 across most of them, reaching, for instance, 700 out of 800 on SAT Math and strong results on several AP exams, while still faring poorly on a few, notably AP English essay writing, which remained weak. The second family is machine-learning benchmarks, where GPT-4 reported 86.4% on the multiple-choice knowledge test MMLU in the English setting and led on translated multilingual variants, coding suites, and commonsense reasoning tasks.

The report is unusually candid about two caveats that qualify these numbers. It documents a contamination analysis: because the training data may contain some exam questions, the authors measured performance on held-out subsets with contaminated items removed and reported that the effect was generally small. The disclosure itself, however, signals that memorization is a live confound for any exam-style claim, and that a headline score cannot be taken at face value without knowing what the model had already seen. The report also states directly that strong scores do not license conclusions about real-world professional competence. On the image side, benchmark results for visual inputs were reported as preview evaluations, consistent with the feature's limited release, so those numbers carry an extra layer of provisionality.

The bar exam claim and its re-examination

The bar-exam headline deserves its own scrutiny because it became the release's public face. A study by legal scholars, GPT-4 passes the bar exam, reported the top-decile result and was cited everywhere from press coverage to OpenAI's own materials. A later peer-reviewed re-analysis, Re-evaluating GPT-4's bar exam performance, argued that the percentile was inflated by the comparison group behind it: the "top 10%" figure benchmarked GPT-4 against a February cohort skewed toward repeat takers who had already failed the exam at least once, a weaker-than-typical reference population. Measured against first-time examinees, the re-analysis put the model closer to the 60th percentile overall, and below the median on the written essay components that most resemble the actual practice of law.

This is a clean illustration of why an event's numbers and their interpretation must be separated. GPT-4's raw score did not change between the two papers; what changed was the reference population against which that score was ranked. The corrected reading still describes a genuinely capable system, one that would plausibly pass, but it undercuts the specific "top 10%" claim that traveled furthest and hardened into common knowledge. A careful timeline records both the original figure and its documented revision, and treats the gap between them as information about how benchmark claims propagate, not as a scandal.

Safety testing and what the system card documented

Alongside the model, OpenAI published a GPT-4 System Card describing the safety work done before release. It reports an extended red-teaming process in which external domain experts probed the model for harmful capabilities, and it presents mitigations that measurably reduced the model's tendency to produce disallowed content relative to an early, less-aligned checkpoint. The card is frank that these mitigations are partial: the deployed model still produces confident falsehoods, can be pushed past its guardrails through adversarial prompting, and can reflect and amplify social biases present in its training data.

One evaluation from the card became a durable reference point in AI-safety discussion. The Alignment Research Center tested an early version of GPT-4 for emergent risky behaviors, including whether it could autonomously acquire resources, replicate copies of itself, and resist being shut down. The reported conclusion was that the model was ineffective at autonomous replication in the tested setting. That result is often cited less as evidence of imminent danger than as an early, concrete precedent for pre-deployment "dangerous capability" testing, a practice that later became standard across frontier labs and government evaluation bodies. The documented value of the card, in other words, is the process it helped establish, not any single number inside it.

The competing approaches of early 2023

GPT-4 launched into an unusually crowded moment. On the very same day, Anthropic announced Claude, an assistant built around its "Constitutional AI" approach, which trains a model to be helpful and harmless by having it critique and revise its own outputs against a written set of principles, rather than relying solely on human preference labels. Google had already signaled its response weeks earlier with Bard, a conversational service initially built on its LaMDA models and tied closely to search, and it would follow with larger PaLM-based systems. A different bet entirely came from open weights: Meta's LLaMA, released in late February 2023, showed that comparatively small, efficiently trained models with published weights could be broadly useful, seeding a large ecosystem of fine-tuned derivatives almost overnight.

These were genuinely different strategies, not one obvious path with laggards trailing behind. GPT-4 represented the closed, scale-and-alignment frontier; Claude foregrounded a distinct alignment method; Bard tied generation to a search product; LLaMA bet on openness and efficiency over raw scale. Evaluating GPT-4 against that live field, rather than against a caricatured "before," is what makes its position legible. It was the strongest single system of its moment on many measures, but it was one option among several credible research programs, each carrying different assumptions about safety, cost, and who should be able to run the model.

What the release deliberately did not disclose

The most historically distinctive feature of GPT-4 may be what its documentation withheld. The technical report states plainly that, given both the competitive landscape and the safety implications of large-scale models, it contains no details about architecture including model size, no hardware or training compute figures, and no information about dataset construction or the training method beyond the high-level alignment recipe. For a document titled a "technical report," this was a marked departure from the disclosure norms that GPT-3, BERT, and much of the field had helped set, in which a paper let others reproduce and scrutinize the result.

The consequences run in two directions. Scientifically, GPT-4 cannot be independently reproduced or audited from its own paper, and widely repeated outside claims about its parameter count or internal design are inference and rumor, not established fact. Practically, OpenAI framed the nondisclosure as both commercial protection and a safety measure against proliferation of dangerous capabilities, while critics argued it weakened the scientific record and made genuine external safety assessment harder. Either reading points to the same historical marker: the event made visible a turn away from open publication toward guarded frontier releases, a posture that later frontier developers, including several of OpenAI's rivals, largely adopted for their own strongest systems.

The documented downstream influence

Downstream, GPT-4 quickly became the reference model against which a generation of systems was measured, and it moved from demo to infrastructure through the OpenAI API and developer platform, where it underpinned a wave of products built by other companies. Its multimodal promise was realized more fully in September 2023, when image input reached general availability alongside a dedicated GPT-4V(ision) System Card documenting the safety work specific to visual inputs, from person-identification refusals to medical-advice limits. Independent researchers also studied the model directly and prominently: Microsoft's Sparks of Artificial General Intelligence argued, from qualitative experiments with an early unrestricted version, that GPT-4 exhibited unusually broad capabilities, a paper as notable for the boldness of its framing as for the sharp methodological debate it provoked about anecdote, cherry-picking, and hype.

The influence also shows in the benchmark ecosystem itself. GPT-4's scores were high enough that several long-standing evaluation suites began to saturate, which pushed the research community toward harder, more contamination-resistant, and more adversarial benchmarks. That pressure on measurement is itself a marker of impact: a model reshapes a field partly by exhausting the tests previously used to judge progress, forcing the community to redefine what counts as a hard problem.

The honest limits and how to place the event

The caveats are not footnotes. GPT-4 still fabricated facts confidently, remained sensitive to small changes in prompt phrasing, operated with a fixed knowledge cutoff, and offered no guarantee that its expressed confidence tracked its actual accuracy. Its exam scores, as the bar-exam episode showed, are only as meaningful as the comparison population and the contamination controls behind them. And because the training details were withheld, the most basic scientific questions about why it worked as well as it did remain closed to outside researchers, which limits how confidently anyone outside OpenAI can generalize from it.

So where does the event belong? GPT-4 is best placed not as an invention but as a consolidation. It took ingredients that already existed, the autoregressive Transformer, RLHF alignment, and the ChatGPT interaction model, executed them at a new level of reliability, and packaged the result as a dependable developer platform with documented safety testing. Its lasting significance is threefold: it made general-purpose models credible for real professional workflows rather than demos, it normalized rigorous pre-deployment safety documentation as an expectation, and it accelerated the field's move toward closed frontier releases. Read alongside its documented predecessors and its honest limits, GPT-4 raised the bar for what a language model was expected to do, without resolving the harder questions of how such systems reason, when they can be trusted, and who is permitted to inspect them.

References

• OpenAI (2023). GPT-4.
• OpenAI (2023). GPT-4 Technical Report.
• OpenAI (2023). GPT-4 System Card.
• OpenAI (2023). GPT-4V(ision) System Card.
• Ouyang et al. (2022). Training language models to follow instructions with human feedback.
• Brown et al. (2020). Language Models are Few-Shot Learners.
• Bubeck et al. (2023). Sparks of Artificial General Intelligence: Early experiments with GPT-4.
• Katz et al. (2024). GPT-4 passes the bar exam.
• Martínez (2024). Re-evaluating GPT-4's bar exam performance.
• Anthropic (2023). Introducing Claude.
• Pichai (2023). An important next step on our AI journey.
• Touvron et al. (2023). LLaMA: Open and Efficient Foundation Language Models.