GPT-4

GPT-4, released March 14, 2023, was OpenAI's frontier model post-ChatGPT and the first widely-deployed model that crossed several "human-expert" thresholds: passing bar exams in the top decile, scoring ~5 on AP Biology, solving novel coding interview questions reliably. It was also the first frontier model whose technical report withheld nearly all training details — model size, architecture, dataset, compute. GPT-4 is the inflection point at which frontier-model research became a closed-door commercial activity.

What GPT-4 demonstrated

The GPT-4 technical report (OpenAI, 2023) emphasised capabilities rather than methodology:

• Bar Exam (Uniform) — 90th percentile on a simulated test (vs GPT-3.5's 10th).
• AP Biology, AP Chemistry, AP Calculus BC, AMC 10 — high or top scores.
• LeetCode Hard, MMLU, HumanEval — substantial gains over GPT-3.5.
• Theory-of-Mind, multi-step reasoning — qualitative improvements.

The headline framing: GPT-4 reached expert human performance on several professional and academic exams. Whether that constitutes "intelligence" or just sophisticated pattern matching against memorised content was (and remains) contested — but the practical capability gap to GPT-3.5 was real and large.

Multimodal input

GPT-4's headline architectural feature was vision input: the model accepted images interleaved with text and could answer questions about them. The vision capability was rolled out gradually — text-only at March release, image input via API months later, and integrated into ChatGPT Plus through 2023.

This was the first widely-available frontier multimodal LLM. The architecture — almost certainly some variant of Flamingo's frozen-LM-with-vision-bridge or a fully-jointly-trained equivalent — was not disclosed.

What was withheld

The GPT-4 technical report explicitly stated:

Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.

This was a methodological discontinuity. GPT-1, GPT-2, GPT-3, and InstructGPT had been documented in detail. GPT-4 was not. Subsequent frontier-model releases — Claude 2/3, Gemini, GPT-4 Turbo — followed the same closed pattern.

External estimates put GPT-4 at around 1–1.8T parameters in a Mixture-of-Experts arrangement, with ~13T training tokens. None of these are confirmed. Pretraining cutoff was September 2021 (later updated for GPT-4 Turbo).

Alignment

GPT-4's alignment was reported in detail (relative to other aspects). Key elements:

• RLHF following the InstructGPT recipe.
• Rule-based reward models trained on safety-policy violations to supplement preference RLHF.
• Red-teaming by ~50 external experts in domains like chemistry, biosecurity, cybersecurity.
• Pre-deployment evaluation by ARC (now METR) on autonomous-replication and similar capability-acquisition tasks.

The alignment work was substantial enough that the model was less harmful per output than GPT-3.5 by most external metrics. Internal jailbreaks remained possible but harder.

Hallucinations and limits

GPT-4 was the first frontier LLM where the gap between capability and reliability became the dominant practical concern. The model could often solve a problem at expert level but occasionally produce confident hallucinations on basic factual queries. Failure modes:

• Citations — invented papers, plausible-but-fake URLs, fabricated quotes.
• Calibration — equally fluent on true and false statements.
• Multi-step arithmetic — improved over GPT-3.5 but still inconsistent on multi-digit problems.
• Adversarial prompts — system-prompt extraction, jailbreaks via roleplay or low-resource languages.

The "wild capability + occasional brittleness" combination is the GPT-4 phenomenology, and the framing every later frontier model has had to address.

What GPT-4 enabled commercially

GPT-4 opened the API in late March 2023. The 12 months that followed produced the first sustained wave of LLM-powered products:

• Coding assistants (Cursor, Cody, Copilot Chat).
• Customer-service automation (Intercom Fin, etc.).
• Search integration (Bing Chat, Perplexity, Phind).
• Knowledge-work tools (Notion AI, Microsoft Copilot, Google Duet).
• Agentic frameworks (early AutoGPT, BabyAGI, LangChain agents).

By 2024, "wraps GPT-4" was a common-enough product description to be a meme.

What GPT-4 was, methodologically

GPT-4 was the model where the public research-paper recipe ended at the frontier. Everything since — Claude 2/3, Gemini, GPT-4 Turbo, GPT-4o, o1, GPT-5 — has been reported with even less detail than GPT-4. Open-source efforts (LLaMA, Mistral, Qwen) followed, partly as a reaction to this closure.

For curriculum purposes: GPT-4 is the last frontier model whose existence is a single named release. After 2023, frontier models become continuously-updated systems where "GPT-4" or "Claude 3" refers to a moving target rather than a paper.

What to read next

• Claude — the contemporary alignment-focused frontier model.
• LLaMA — the open-weights reaction.
• Frontier Models — the continuing frontier story.