Model Case Study Series · Consolidated Lineage 🗂️

The GPT Lineage: A Consolidated Comparison, GPT-1 to GPT-6 Astra

Six full case studies in this series each told one model's story end to end. This article steps back and lines all of them up side by side — plus the eight interim releases between the major numbered generations (GPT-3.5, GPT-4 Turbo, GPT-4o, o1, GPT-4.5, and the GPT-5.1 through 5.6 point releases) — across seven consistent dimensions: research trigger, data, architecture, learning technique and scaffolding, training technique, compute, and influence on what came next.

FrontierAGI Team

September 16, 2026 · 85 min read

Opening Full Timeline Master Comparison GPT-1 GPT-2 GPT-3 Bridge: 3.5 → 4.5 GPT-4 GPT-5 Bridge: 5.1 → 5.6 GPT-6 Astra Cross-Cutting Threads What Stayed Constant Readiness Checklist What's Missing References Videos Verdict

Fifteen Releases, One Underlying Recipe

Reading the six full case studies in this series one at a time shows six good stories. Reading them side by side, with the interim releases filled in between, shows something the individual stories can't: a set of trend lines. Data went from a bounded, curated corpus to an undisclosed mixture too large to describe. Scaffolding went from nonexistent to a fully agentic system operating real software. Compute disclosure went from precise to nonexistent, almost exactly as compute itself grew by many orders of magnitude. This article is that comparison, built to be read as a reference table first and a narrative second.

The Full Timeline

Fifteen Releases in Eight Years

Jun 2018GPT-1

Feb 2019GPT-2

Jun 2020GPT-3

Nov 2022GPT-3.5 / ChatGPT

Mar 2023GPT-4

Nov 2023GPT-4 Turbo

May 2024GPT-4o

Sep 2024o1-preview

Feb 2025GPT-4.5

Aug 2025GPT-5

Nov 2025GPT-5.1

Dec 2025GPT-5.2

Feb 2026GPT-5.3-Codex

Mar 2026GPT-5.4

Apr 2026GPT-5.5

Jul 2026GPT-5.6 Luna/Terra/Sol

Sep 2026GPT-6 Astra

Solid purple = the six major numbered generations this series covers in full case studies. Outlined = interim releases covered here as bridges.

17 Distinct named releases traced in this article, June 2018 to September 2026

8yr Span from GPT-1's 117M parameters to GPT-6 Astra's undisclosed, agentic scale

1 Architecture family (decoder-only Transformer) underlying every single release

The Master Comparison

Seven Dimensions, Six Major Generations

The full seven-dimension table below covers the six major numbered generations this series has profiled in depth. Cells marked undisclosed reflect a genuine gap in the primary source, not an oversight in this article — consistent with every individual case study's own honesty discipline.

Model	Date	Research Trigger	Data	Architecture	Learning Technique & Scaffolding	Training Technique	Compute	Influence on Next
GPT-1	Jun 2018	NLP lacked vision-style transfer learning	BooksCorpus (~7,000 books)	12-layer decoder-only, 117M params	Pretrain + full fine-tune per task	Adam, 100 epochs, 512-token sequences	~1 month, 8 GPUs	Proved the recipe; motivated removing fine-tuning entirely
GPT-2	Feb 2019	Fine-tuning per task was still a real cost	WebText (~8M docs, Reddit karma filter)	Decoder-only, 4 sizes to 1.5B, context 1024	Zero-shot via natural-language prompting	Same LM objective, staged size study	Undisclosed	Proved prompting alone could elicit tasks; motivated in-context learning
GPT-3	Jun 2020	Zero-shot was inconsistent across tasks	Filtered/weighted Common Crawl + WebText2 + Books + Wikipedia	Decoder-only, 175B params, dense/sparse attention, context 2048	In-context few-shot learning, no gradient updates	Same LM objective at 100x+ scale	3.14×10²³ FLOPs	Base model for RLHF (InstructGPT) → ChatGPT product line
GPT-4	Mar 2023	ChatGPT's success demanded a safer, multimodal product model	Undisclosed	Undisclosed (confirmed: Transformer, multimodal input)	RLHF + predictable-scaling methodology + red-teaming	Undisclosed	Undisclosed	Set the System Card + closed-disclosure template
GPT-5	Aug 2025	Users forced to manually pick fast vs. reasoning models	Undisclosed	Undisclosed (unified multi-model system)	Real-time router + "safe completions" safety training	Undisclosed	Undisclosed	Routing/agentic direction carried into 5.x and Astra
GPT-6 Astra	Sep 2026	Need for agentic computer-use; forced safer release by a real incident	Undisclosed	Undisclosed (multimodal, ~1.1M-token context)	Agentic computer-use scaffolding + stricter containment	Undisclosed	Undisclosed	First "Critical" classification sets the containment bar going forward

GPT-1 — June 2018

GPT-1 in Detail

GPT-1Jun 2018

Research Trigger: Vision had ImageNet-style transfer learning; NLP had no equivalent, leaving most models trained from scratch per task on small labeled datasets.
Data: BooksCorpus — roughly 7,000 unique unpublished books, chosen specifically for long, contiguous text.
Architecture: 12-layer decoder-only Transformer, 768-d hidden states, 12 heads, 117M parameters, 40k-merge BPE vocabulary.
Learning Technique & Scaffolding: Pretrain via next-token prediction, then fully fine-tune the same model per downstream task with a minimal task-specific input reformatting — no separate scaffolding layer yet.
Training Technique: Adam optimizer, max LR 2.5e-4, 100 epochs, 64-sequence minibatches of 512 contiguous tokens.
Compute: Approximately one month on 8 GPUs — the most precisely disclosed compute figure in the entire lineage.
Influence on Next: Proved pretrain-then-adapt works at all; GPT-2 asked whether the adaptation step (fine-tuning) could be removed entirely.

Read the full GPT-1 case study →

GPT-2 — February 2019

GPT-2 in Detail

GPT-2Feb 2019

Research Trigger: GPT-1's fine-tuning-per-task requirement was still a real, ongoing cost; could implicit task knowledge already present in broad training data be elicited without any fine-tuning at all?
Data: WebText — roughly 8 million documents (~40GB) from outbound Reddit links with 3+ karma, used as a scalable quality-filter proxy.
Architecture: Decoder-only Transformer, pre-activation layer norm, context doubled to 1,024 tokens, 50,257-token BPE vocabulary, four sizes from 117M to 1.5B.
Learning Technique & Scaffolding: Zero-shot task transfer — the task is specified purely through natural-language prompt phrasing (e.g. "TL;DR:"), the first real scaffolding idea in the lineage.
Training Technique: Same next-token-prediction objective as GPT-1, studied explicitly across four model sizes to observe scaling behavior.
Compute: Not disclosed — the first gap in an otherwise precise lineage of figures.
Influence on Next: Showed prompting alone could work, inconsistently; GPT-3 asked whether showing a few examples in the prompt (few-shot) could close that consistency gap.

Read the full GPT-2 case study →

GPT-3 — June 2020

GPT-3 in Detail

GPT-3Jun 2020

Research Trigger: GPT-2's zero-shot results were promising but inconsistent; humans learn new tasks from a handful of examples, not thousands of labeled ones.
Data: A filtered, deduplicated, and weighted mixture: Common Crawl (quality-classified), WebText2, Books1/Books2, and Wikipedia — cleaner sources sampled more than their raw size.
Architecture: Decoder-only Transformer, 175B parameters (largest of 8 sizes), alternating dense and locally-banded sparse attention, context 2,048 tokens.
Learning Technique & Scaffolding: In-context few-shot learning — a handful of demonstrations placed directly in the prompt, with the model's weights completely frozen; formally compared against zero-shot and one-shot.
Training Technique: Same next-token-prediction objective, scaled to a regime where compute itself became a subject of study.
Compute: 3.14×10²³ FLOPs for the full 175B model, explicitly disclosed and used to study compute-vs-performance scaling.
Influence on Next: Became the base model fine-tuned via RLHF (InstructGPT) into the model behind ChatGPT — the direct bridge to GPT-3.5.

Read the full GPT-3 case study →

Bridge — GPT-3.5 to GPT-4.5

Five Releases Between GPT-3 and GPT-4

None of the following five releases got a dedicated case study in this series, but each changed at least one of the seven dimensions in a way that shaped the generation after it.

Nov 30, 2022GPT-3.5 / ChatGPT

Learning technique & scaffolding shift: applying reinforcement learning from human feedback (RLHF), via the InstructGPT method (Ouyang et al., 2022), on top of a GPT-3-class base model. This is the release that turned a raw completion engine into a usable conversational product — the single biggest scaffolding change in the entire lineage relative to its technical size.

Nov 2023GPT-4 Turbo

Deployment shift, not a learning-technique shift: expanded context window to 128,000 tokens, updated knowledge cutoff, and substantially lower per-token cost — a scale-of-deployment change layered onto GPT-4's undisclosed core.

May 13, 2024GPT-4o ("omni")

Architecture shift: native multimodality — text, image, and audio processed by one unified set of weights rather than routed through separate modality-specific models, enabling real-time voice interaction for the first time in this lineage.

Sep 2024o1-preview

Training-technique shift: large-scale reinforcement learning trained the model to reason via extended chain-of-thought before answering, with performance improving as more training-time and test-time (inference) compute was spent thinking. This "reasoning model" fork is the direct ancestor of GPT-5's deep-reasoning routing path.

Feb 27, 2025GPT-4.5

Compute-and-data shift: OpenAI's largest model to that point, built by scaling raw unsupervised pretraining rather than reasoning or RL — explicitly positioned as advancing "natural," broad-knowledge capability rather than step-by-step reasoning. The last major release in this lineage to bet purely on pretraining scale before GPT-5's routed unification.

GPT-4 — March 2023

GPT-4 in Detail

GPT-4Mar 2023

Research Trigger: ChatGPT's overnight mass adoption changed the problem from "build a more capable base model" to "build a safer, more reliable model for hundreds of millions of users."
Data: Not disclosed by OpenAI.
Architecture: Not disclosed (confirmed: Transformer-based, multimodal text+image input, RLHF fine-tuned).
Learning Technique & Scaffolding: RLHF fine-tuning plus a formal, documented adversarial safety-testing process (including ARC's autonomous-replication evaluation) before release.
Training Technique: Not disclosed in detail, beyond confirming a next-token pretraining objective plus RLHF.
Compute: Not disclosed — though the report discloses a genuine methodological contribution: predicting final performance from small-scale proxies costing a fraction of full training compute.
Influence on Next: Established the technical-report-plus-System-Card release template every major frontier model since (including this lineage's own GPT-5 and GPT-6) has followed.

Read the full GPT-4 case study →

GPT-5 — August 2025

GPT-5 in Detail

GPT-5Aug 2025

Research Trigger: Users were forced to manually choose between separate fast and reasoning models before every conversation — a real product and compute-efficiency problem.
Data: Not disclosed.
Architecture: Not disclosed for any component model (confirmed: a unified system composing multiple internal models).
Learning Technique & Scaffolding: A real-time router deciding, per request, between a fast model and a deeper-reasoning model — plus "safe completions" safety training replacing binary refuse-or-comply behavior.
Training Technique: Not disclosed.
Compute: Not disclosed.
Influence on Next: The routing/composition idea, refined across the GPT-5.1–5.6 point releases, became the direct precursor to GPT-6 Astra's agentic, tool-using scaffolding.

Read the full GPT-5 case study →

Bridge — GPT-5.1 to GPT-5.6

Six Point Releases in Nine Months

GPT-5's routed architecture set the stage for an unusually fast point-release cadence — six named releases between November 2025 and July 2026, each refining the router and reasoning balance rather than introducing a new architectural idea.

Release	Date	What Changed
GPT-5.1	Nov 2025	Refined the fast/reasoning routing balance from GPT-5's initial launch configuration
GPT-5.2	Dec 2025	Incremental capability and reliability updates within the same routed system
GPT-5.3-Codex	Feb 5, 2026	A coding-specialized variant, extending the router concept into a dedicated developer-tool configuration
GPT-5.4	Mar 5, 2026	Continued point-release refinement of the core routed system
GPT-5.5	Apr 23, 2026	Continued point-release refinement of the core routed system
GPT-5.6 (Luna / Terra / Sol)	Jul 9, 2026	Naming shifted from Instant/Thinking/Pro tiers to three named models (Luna fastest, Terra mid-tier, Sol flagship); Sol became OpenAI's strongest vision model to date, with a ~1.05M-token context window

GPT-5.6 Sol is also the model directly implicated, alongside an unreleased pre-release model, in the July 2026 Hugging Face containment incident this lineage's GPT-6 Astra case study covers in depth — making this bridge era the direct run-up to the most safety-eventful release in the entire series.

GPT-6 Astra — September 2026

GPT-6 Astra in Detail

GPT-6 AstraSep 2026

Research Trigger: Demand for genuine long-horizon, computer-operating agentic capability, combined with an urgent need for better containment after the July 2026 Hugging Face incident directly demonstrated a real failure mode.
Data: Not disclosed.
Architecture: Not disclosed (confirmed: multimodal input, ~1.1M-token context window, up to 128,000 completion tokens).
Learning Technique & Scaffolding: Agentic scaffolding for operating real computer interfaces (browsers, desktop apps) over long multi-step tasks, paired with stricter internal isolation, checkpoint encryption, and full-trajectory monitoring.
Training Technique: Not disclosed.
Compute: Not disclosed.
Influence on Next: The first model to reach "Critical" cybersecurity classification under OpenAI's Preparedness Framework — setting a containment and monitoring bar every subsequent agentic release will be measured against.

Read the full GPT-6 Astra case study →

Cross-Cutting Threads

Three Patterns Only Visible Side by Side

Thread 1 — Data: A Transparency Arc, Not Just a Scale Arc Bounded and named (GPT-1's BooksCorpus) → scraped with a cheap quality proxy (GPT-2's Reddit-karma WebText) → filtered and weighted at scale but still described (GPT-3's Common Crawl mixture) → entirely undisclosed (GPT-4 onward). The dataset didn't just get bigger — the field's willingness to describe it shrank in lockstep.

Thread 2 — Scaffolding: Where the Real Engineering Moved None (GPT-1's raw fine-tuning) → prompting (GPT-2) → in-context few-shot (GPT-3) → RLHF conversational tuning (GPT-3.5) → chain-of-thought reasoning (o1) → routing between models (GPT-5) → full agentic computer-use with containment (GPT-6 Astra). By the second half of this lineage, the scaffolding layer arguably mattered more to the user-visible product than the base architecture underneath it.

Thread 3 — Compute vs. Disclosure: An Inverse Relationship GPT-1: exact figure ("1 month, 8 GPUs"). GPT-2: undisclosed. GPT-3: exact figure again (3.14×10²³ FLOPs) alongside a full scaling study. GPT-4 through GPT-6 Astra: undisclosed, every time. As training runs grew from a single small cluster to whatever it takes to train a "Critical"-classified agentic system, public compute disclosure went the opposite direction, not the same one.

What Stayed Constant

The One Thread That Never Broke

Across all seventeen releases in this timeline — six major generations and eleven interim ones — every single model remains a decoder-only Transformer, pretrained via next-token prediction on the preceding token sequence. Vaswani et al.'s 2017 architecture, adopted for GPT-1 the following year, is still the load-bearing structure underneath GPT-6 Astra's agentic computer-use system eight years later. Everything else in this lineage — data sourcing, scaffolding, safety process, disclosure norms, team size — changed repeatedly and often dramatically. The core architectural bet never did.

Eight years, seventeen releases, roughly a million-fold increase in disclosed-or-inferred scale — and the same decoder-only Transformer, pretrained on next-token prediction, underneath every one of them.

Readiness Checklist

Can you trace the "scaffolding" column alone, from GPT-1 through GPT-6 Astra, without looking at the table?

Can you explain why compute disclosure and compute scale moved in opposite directions across this lineage?

Can you name which interim release introduced native multimodality, and which introduced chain-of-thought reasoning?

Could you explain, to someone new to this series, why the architecture column is the least interesting one in this entire table?

⚠️ What's Missing or Uncertain

The interim releases in this article are covered at bridge depth, not full case-study depth. GPT-3.5, GPT-4 Turbo, GPT-4o, o1-preview, GPT-4.5, and the GPT-5.1 through 5.6 point releases each involved real engineering and research decisions this article does not fully unpack — this table and its bridge sections summarize the single most consequential change per release, not its complete story. Undisclosed cells (marked throughout) reflect real gaps in OpenAI's own public materials, not gaps in this article's research; where a number or detail could not be verified against a primary source, it is marked undisclosed rather than estimated.

🔗 Reference Links

🎥 Recommended Videos

But What Is a Neural Network?

3Blue1Brown — the one architectural thread every release in this table shares

Attention Is All You Need — Paper Explained

Yannic Kilcher — the 2017 paper this entire eight-year lineage is built on

🧭 Closing — The Table Says What No Single Article Could

🎯 The Bottom Line

Read one at a time, each GPT generation looks like its own leap. Read as one table, the leaps resolve into three durable trend lines: data description shrinking, scaffolding doing more of the real work, and compute disclosure moving inversely to compute scale — all built, without exception, on the same decoder-only Transformer architecture from 2018 onward. The interesting story in this lineage was never really the architecture. It was everything the field decided to build, disclose, and eventually stop disclosing, around it.