GPT-5: How the Fifth GPT Was Actually Built

One Product, Several Models Underneath

By 2025, OpenAI's ChatGPT interface had accumulated a confusing lineup of separate model choices — fast general-purpose models alongside slower, more deliberate "reasoning" models users had to pick between manually. GPT-5's headline change was structural rather than purely architectural: it replaced that manual picker with a single named system that decides, per request, whether to answer quickly or to think longer, without the user needing to know which internal model actually handled it. This is a genuinely different kind of "how was it built" question than any prior entry in this series — the interesting engineering isn't only inside one model, it's in the system that decides which model to use at all.

Too Many Models, Too Much User Judgment Required

By the GPT-4 era, OpenAI's product lineup had split into fast conversational models and separate, slower "reasoning" models better suited to hard multi-step problems — leaving users to guess, before every conversation, which one they actually needed. This is the same problem this series' Compute Economics article frames from the infrastructure side: routing a request to more compute than it needs wastes resources, while routing a hard problem to a fast, shallow model wastes the user's time on a wrong or shallow answer. GPT-5's problem statement was explicitly to remove that manual decision from the user entirely.

A Real-Time Router, Not Just a Bigger Model

GPT-5's core idea, as OpenAI publicly described it: a unified system with an internal router that decides, per message, whether a fast model can answer well or whether the request should be handed to a deeper, more deliberate reasoning process — all presented to the user as a single, consistent "GPT-5" experience. This is a different kind of scaling bet than any prior entry in this series: rather than making one model bigger (GPT-3's approach) or keeping one model's size undisclosed (GPT-4's approach), GPT-5 scales by composing multiple specialized models behind one interface and an automated decision layer.

Continuing the Uncountable-Team Pattern

Consistent with GPT-4, OpenAI's public GPT-5 materials do not provide a clean, citable author or team-size count. This isn't a gap specific to this case study — it's the continuation of the pattern this series' GPT-4 article first documented: past a certain organizational scale, a simple headcount stops being how a frontier lab describes who built something, and this series' Other Teams' Code article's Conway's Law framing applies more forcefully with each release, not less. A system with a router deciding between multiple internal models plausibly implies at least separate specialized teams for the fast model, the reasoning model, and the routing layer itself — a reasonable inference from the architecture, not a confirmed organizational fact.

What's Confirmed, What Isn't

OpenAI's public GPT-5 announcement describes the system's routing behavior and its improved coding and instruction-following performance relative to prior models, alongside multiple product-facing variants (including smaller, faster variants for cost- and latency-sensitive use, and an extended-reasoning tier for the hardest problems). As with GPT-4, no parameter count, layer architecture, or training-compute figure has been publicly disclosed for any component of the system. This continues, rather than reverses, the opacity trend this series' GPT-4 article identified as a dividing line in the field.

Beyond a Binary Refuse-or-Comply

Alongside GPT-5, OpenAI described a safety-training shift it calls "safe completions": rather than a model either fully answering a request or flatly refusing it, the model is trained to give the most helpful response possible within safety constraints — for example, explaining general safety information about a sensitive topic while withholding specific operational detail that would make the response genuinely dangerous. This is a direct evolution of the refusal-behavior mitigations this series' GPT-4 article describes as part of its System Card, and connects to this series' Alignment article's broader point that alignment techniques keep having to get more nuanced as models get more capable and more widely used — a flat refuse/comply binary becomes less adequate, not more, as capability grows.

Coding, Reasoning, and Reduced Hallucination — Reported, Not Independently Re-Verified Here

OpenAI's public materials report improvements in coding benchmarks, agentic task completion, and reduced hallucination rates relative to prior GPT models, alongside a companion system card describing safety evaluation work. This article does not restate specific benchmark scores as settled fact, in keeping with this series' Self-Assessment Deep Dive article's discipline: a lab's own reported benchmark numbers are a real data point worth citing, but they are not the same as independent, reproduced verification, and readers should treat any specific figure as something to check against the primary system card rather than take from a secondary summary like this one.

A Product Launch, With a System Card Attached

GPT-5 launched simultaneously across ChatGPT (replacing the prior manual model picker for most users) and the API, continuing the tiered product-access pattern this series' GPT-4 article identifies as the post-2023 norm, alongside a system card following the template GPT-4 established. The router itself — deciding fast-vs-deep per request — is also a release-time safety lever in its own right: routing genuinely hard or sensitive queries toward the more deliberate model is, functionally, another layer of the same proactive-mitigation discipline this series' On-Call Reality article describes as better than reacting after the fact.

Still Being Written

Unlike GPT-1 through GPT-4, GPT-5's legacy is not yet settled history — it is the newest entry in this series, released just over a year before this article. What can be said with confidence is that the routed, multi-model-behind-one-product pattern it introduced is a structurally different kind of scaling bet than any prior GPT, and it previews a plausible future where "which specific model answered your question" becomes a genuinely uninteresting detail compared to whether the routing and safety layers around it work well.

Readiness Checklist

⚠️ What's Missing or Uncertain

Where This Case Study Goes Next

This series continued sooner than expected: GPT-6 "Astra" launched just over a year later, on September 3, 2026, forced into an unusually cautious release process by a real safety incident earlier that summer — and became the first OpenAI model to cross into "Critical" territory on the company's own cybersecurity risk framework.

🔗 Reference Links

🎥 Recommended Videos

🧭 Closing — The Unit of Scaling Changed Again