The GPT lineage: from GPT-4 to GPT-5.5

OpenAI's headline models form a clear progression. Each release either widened the context window, added a modality, lowered cost, or moved reasoning from an add-on into the base model. Knowing where a given model sits in this timeline is the fastest way to understand what it can and cannot do.

Where the line is now (mid-2026): The GPT-4 and GPT-4o names are retired from OpenAI's pricing page, superseded by the GPT-5 family. The current lineup is GPT-5.5 ($5 / $30 per million input/output tokens, 1M context), GPT-5.4 ($2.50 / $15), GPT-5.4 mini ($0.75 / $4.50), and GPT-5.3-Codex ($1.75 / $14, 400K) for dedicated coding agents. If you read a 'GPT-4o pricing' table today, it is almost certainly stale.

Architecture: what made GPT-4 different from GPT-3

GPT-3 (2020) was a 175-billion-parameter dense model — every parameter runs on every token. GPT-4 is widely believed to be a Mixture-of-Experts model: instead of one giant network, it routes each token through a small subset of specialized 'expert' sub-networks. Reporting (never confirmed by OpenAI) puts it near 1.8 trillion total parameters with only a fraction active per token, which keeps inference cost closer to a much smaller model while preserving the knowledge capacity of a huge one.

• Mixture-of-Experts routing — only a few experts fire per token, so effective compute is far below the total parameter count
• Far stronger reasoning, coding, and instruction-following than GPT-3.5, with large gains on professional and academic exams
• A 32K context window at launch (8x GPT-3.5), later extended to 128K with GPT-4 Turbo
• Trained with large-scale RLHF (reinforcement learning from human feedback) to make outputs helpful and aligned

The practical lesson is that 'bigger' stopped meaning 'one bigger dense network.' From GPT-4 onward, frontier scale has been achieved largely through MoE — the same idea now used by Gemini, DeepSeek, Mixtral, Qwen, and most other frontier systems.

Is GPT-4o MoE or dense? The honest answer: Unknown — OpenAI has never disclosed GPT-4o's architecture, parameter count, or whether it uses expert routing. No technical report exists. What is known: it was roughly 2× faster and half the price of GPT-4 Turbo at launch, which is consistent with either a smaller dense model or a more efficient MoE — speed and price tell you about deployment economics, not architecture. Every "GPT-4o parameter count" you see online is speculation. The 1.8T MoE reporting applies to the original GPT-4 (via SemiAnalysis), and even that was never confirmed.

GPT-4o: the omnimodal leap

Before GPT-4o, multimodality was stitched together: a vision model described an image, a speech model transcribed audio, and a separate text model reasoned over the results. GPT-4o ('omni') was trained end-to-end across text, images, and audio in a single network, so it perceives and responds across modalities without handoffs. That is why GPT-4o could hold near-real-time spoken conversations, interpret screenshots, and reason over diagrams in one pass.

• Native vision — reads charts, screenshots, handwriting, and documents directly
• Native audio — low-latency voice conversations with tone and interruption handling
• Roughly half the price and twice the speed of GPT-4 Turbo at launch, which is what pushed multimodal AI into everyday consumer use
• Set the template every later frontier model followed: one model, many modalities

The reasoning turn and GPT-5

In late 2024 OpenAI split off a separate o-series (o1, then o3) that 'thinks' before answering by spending extra compute on an internal chain of reasoning. GPT-5 (August 2025) merged that capability back into the main model: a router decides per request whether to answer instantly or to reason at length, so users no longer pick a separate 'reasoning model.' GPT-5.5 continues this — adaptive reasoning is built in, and the model is tuned to use fewer tokens for the same result.

Choosing a GPT model today: Use GPT-5.5 when quality matters most (hard coding, long-horizon agents, high-stakes analysis). Use GPT-5.4 as the cost-efficient frontier default. Route high-volume, simple traffic to GPT-5.4 mini. Use GPT-5.3-Codex for dedicated coding-agent pipelines. Match the model to the task rather than defaulting to the most expensive one.

Benchmarks and key specs

Reading these numbers: Benchmark scores are vendor-reported and shift with model versions and test harnesses. The GPT-4o row reflects its 2024 launch capabilities; the GPT-5.4 and GPT-5.5 rows reflect the current (mid-2026) lineup that replaced it. Treat any single benchmark as directional and test on your own workload.

Frequently asked questions

Is GPT-4o still available? It has been retired from OpenAI's public pricing page and superseded by the GPT-5 family. Some platforms may still expose legacy access for a while, but for new work the direct equivalent is GPT-5.4 (cost-efficient) or GPT-5.5 (flagship).

What is the difference between GPT-4 and GPT-4o? GPT-4 (2023) was the original frontier model; GPT-4o (2024) is a single natively multimodal model that handles text, image, and audio in one network, runs faster, and cost roughly half as much at launch.

Was GPT-4 really 1.8 trillion parameters? That widely cited Mixture-of-Experts figure has never been confirmed by OpenAI. It is informed speculation, not an official spec — treat it as such.

Which GPT model should I use today? Use GPT-5.5 for the hardest agentic coding and reasoning, GPT-5.4 as the cost-efficient default, GPT-5.4 mini for high-volume traffic, and GPT-5.3-Codex for dedicated coding agents.