1. 🧭 Why One Lab's History Is the Clearest Case Study
Our prior posts compared labs against each other at a single point in time — a useful snapshot, but a snapshot hides sequence. Tracing one lineage end to end shows something different: the actual order in which a single lab reached for each of the six levers as the previous one ran into diminishing returns. OpenAI's GPT series is the longest continuous lineage in the industry — eight years, over a dozen major and minor releases — which makes it the cleanest available record of "what did the lab try when the last thing stopped working as well."
2018 → 2026GPT-1 to GPT-6 Astra — eight years, the longest continuous frontier model lineage tracked on this site
4 PivotsDistinct lever shifts we identify across the series: scale → post-training → multimodal scale → test-time compute → scaffolding
85%Human evaluator preference for InstructGPT over raw GPT-3 — despite InstructGPT using far fewer parameters
37-pt GapAstra's harness-dependent ARC-AGI-3 score swing — the clearest sign yet the current pivot has real limits
2. 🧱 Quick Reference: The Stack and the Levers
For readers new to this series: the AGI capability stack has 9 layers (reasoning, generalization, multimodality, memory, tool/computer use, long-horizon planning, reliability, metacognition, safe autonomy), and progress on any layer comes from one of six levers — pre-training scale, compute/efficiency, post-training (RLHF/RLVR/distillation), agentic scaffolding, memory/continual-learning research, and the underlying power/compute constraint. Full definitions in The AGI Threshold and The Six Levers.
Lever 1: Pre-training Scale
Lever 2: Compute/Efficiency
Lever 4: Post-Training
Lever 3: Test-Time Compute
Lever 5: Agentic Scaffolding
Lever 6: Memory/Continual Learning
3. 📅 The Full GPT Timeline
Multimodal/Efficiency Era
Scale Era2018
GPT-1
The founding bet: generative pre-training at scale, then fine-tune
LeverPre-training scale, applied for the first time to a transformer decoder — the entire template for everything that followed.
Lever 1
Scale EraFeb 2019
GPT-2
1.5B parameters — a 10x scale-up of both model size and data over GPT-1
LeverPure Lever 1 — a direct 10x scale-up in parameters and training data, staged for release over safety concerns about coherent long-form generation.
Lever 1
Scale EraMay 2020
GPT-3
175B parameters — few-shot learning as an emergent property of scale alone
LeverLever 1, taken to its most famous conclusion — scale alone produced few-shot learning nobody explicitly trained for, becoming the first commercially viable GPT model via API.
Lever 1
Post-Training EraJan 2022
InstructGPT
Same base model, retrained on human preference — RLHF as its own lever
LeverThe series' first real pivot away from Lever 1 — RLHF fine-tuning on top of an unchanged GPT-3 base produced outputs human evaluators preferred 85% of the time, using far fewer effective parameters worth of "new" capability.
Lever 4
Post-Training EraNov 2022
ChatGPT (GPT-3.5)
RLHF applied to conversation — 100M users in two months
LeverSame lever as InstructGPT, refined for multi-turn dialogue — the product moment that proved post-training could be worth more commercially than the underlying base-model scale-up.
Lever 4
Scale EraMar 2023
GPT-4
Back to scale — now combined with multimodal input (GPT-4V)
LeverA return to Lever 1, this time paired with new data modalities — the multimodal/world-coherence stack layer enters at real strength for the first time in the series.
Lever 1
Efficiency EraNov 2023 – May 2024
GPT-4 Turbo → GPT-4o
128K context, function calling, then 2x faster at half the cost with native audio
LeverA pivot to Lever 2 (compute efficiency) and early Lever 5 (function calling, code interpreter, browsing) — GPT-4o matched Turbo's text/code performance at roughly half the API cost, the clearest efficiency-over-scale release in the series so far.
Lever 2Lever 5
Reasoning EraSep – Dec 2024
o1 → o3
The second major pivot — models that think before they answer
LeverOpenAI's clearest departure from pure pre-training scale — o1 (preview Sep 12, 2024) and o3 (announced Dec 20, 2024) invest extra inference-time computation to reason step by step rather than scaling parameters further.
Lever 3
Agentic EraAug 2025 – Jul 2026
GPT-5 → GPT-5.6
Unifying reasoning and chat lines; tiered variants (Luna/Terra/Sol) for cost/capability tradeoffs
LeverConsolidation of Lever 3 (reasoning) with heavier Lever 5 investment — GPT-5.6 Sol positioned as OpenAI's strongest cybersecurity model yet, with defensive-capability framing (threat modeling, patching, blue-teaming) signaling agentic, tool-using deployment as the target use case.
Lever 3Lever 5
Agentic Era — NowSep 2026
GPT-6 Astra
"Generational leap" — and the harness controversy that questioned what that meant
LeverLever 3 and Lever 5 pushed to their current limits — and the first release in the series where the lab's own safety researchers publicly worried the lever choice (reduced CoT transparency) was outrunning Lever 6 (safe autonomy).
Lever 3Lever 5Current Frontier
4. 🔤 GPT-1 & GPT-2: Pure Pre-Training Scale (2018–2019)
GPT-1 (2018) established the template: generative pre-training on unlabeled text, then task-specific fine-tuning. GPT-2 (February 14, 2019) was a direct 10x scale-up of both parameters (1.5 billion) and training data (8 million web pages) over GPT-1 — released in stages specifically because of how coherent its long-form generation had become. Every bit of capability gain across this era traces to one lever: more data, more parameters. No tool use, no agentic behavior, no reasoning-specific training existed yet — these releases sit almost entirely at the base of the stack (raw language modeling, pre-generalization).
5. 📈 GPT-3: Scale Becomes a Law (2020)
Lever 1: Pre-training Scale
175 billion parameters — two orders of magnitude beyond GPT-2 — produced few-shot learning as an emergent property nobody explicitly trained for. The model could perform new tasks from a handful of examples in its prompt, without any fine-tuning. This became the first commercially viable GPT model via API access (beta launched June 11, 2020), and cemented "scale is the lever" as the field's dominant belief for the next two years.
Stack layers touched: Reasoning (partial), Generalization (first real emergence)
6. 🎯 InstructGPT & ChatGPT: The Post-Training Pivot (2022)
InstructGPT → ChatGPT
Jan 2022 → Nov 30, 2022
Lever 4: Post-Training (RLHF)
The series' first genuine lever shift. InstructGPT was built on the same GPT-3 base but fine-tuned with Reinforcement Learning from Human Feedback (RLHF) — and human evaluators preferred its outputs over raw GPT-3's 85% of the time, despite using far fewer effective parameters worth of new training. ChatGPT applied the same RLHF recipe to multi-turn conversation and reached 100 million users within two months — the fastest consumer app adoption at the time. Nothing about the base model's knowledge changed; what changed was how well that knowledge got extracted and aligned to what users actually wanted.
Stack layers touched: Generalization (more consistent transfer), early Reliability
InstructGPT is the cleanest evidence in the entire GPT lineage that a capability jump and a scale increase are not the same event — the base model didn't get bigger, it got better extracted.
7. 🖼️ GPT-4: Scale + Multimodality (2023)
Lever 1: Pre-training Scale
A return to Lever 1 as the primary driver, but paired with a genuinely new dimension: multimodal input via GPT-4V, letting the model process images alongside text for the first time at real strength. This is the release where the multimodality/world-coherence stack layer stops being a footnote and becomes a real, separately-marketed capability.
Stack layers touched: Reasoning (major jump), Multimodality (new), Generalization
8. 🔧 GPT-4 Turbo → GPT-4o: Efficiency and Early Tools (2023–2024)
GPT-4 Turbo → GPT-4o
Nov 2023 → May 2024
Lever 2: Compute/EfficiencyLever 5: Early Scaffolding
GPT-4 Turbo introduced a 128K-token context window, JSON mode, and the Assistants API — early scaffolding infrastructure. GPT-4o (May 2024) is the clearest efficiency-lever release in the series: it matched Turbo's text and code performance while running roughly twice as fast, at half the API cost, with five times higher rate limits — plus native multimodal audio generation. This is a lever the series hadn't leaned on this explicitly before: making an equivalent-capability model radically cheaper to run, rather than making a more capable one.
Stack layers touched: Tool & Computer Use (function calling, browsing), efficiency across all layers
9. 🧠 o1 → o3: The Test-Time Compute Pivot (2024–2025)
o1 → o3
Sep 12, 2024 → Apr 16, 2025 (o3 GA)
Lever 3: Test-Time Compute
The single clearest pivot in the entire GPT history. o1 (preview Sep 12, 2024; full release Dec 5, 2024) and o3 (announced Dec 20, 2024; GA April 16, 2025 alongside o4-mini) do not simply scale up parameters — they invest additional computation at inference time, reasoning through problems step by step before answering. o3-mini's low/medium/high reasoning-effort variants made this lever explicitly user-configurable for the first time: more thinking time, traded directly against latency and cost, as a dial rather than a fixed model property.
Stack layers touched: Reasoning (major), Reliability (partial, via self-checking)
10. 🤖 GPT-5 → GPT-5.6: Consolidation and Agentic Scaffolding (2025–2026)
GPT-5 → GPT-5.6
Aug 7, 2025 → Jul 9, 2026
Lever 3: Test-Time ComputeLever 5: Agentic Scaffolding
GPT-5 (August 7, 2025) unified the reasoning line (o-series) and the conversational line (GPT-4 series) into one model family. GPT-5.6 (July 9, 2026, after a June 26 limited preview) shipped in three tiers — Luna, Terra, and Sol — explicitly trading capability against cost (Terra: GPT-5.5-competitive performance at half the price). GPT-5.6 Sol was marketed as OpenAI's "strongest cybersecurity model yet," built for defensive agentic work — threat modeling, code review, patching, blue-teaming — a direct signal that Lever 5 (agentic tool use in real security workflows) was now the primary commercial target, not raw benchmark reasoning scores alone.
Stack layers touched: Tool & Computer Use, Long-Horizon Planning, Reasoning (refined, not re-scaled)
11. 🌌 GPT-6 Astra: The Current Frontier — and Its Limits (2026)
Covered in full in our Astra, Claude & Gemini post: GPT-6 Astra (September 3, 2026 limited preview) pushed Levers 3 and 5 further than any prior GPT release — ExploitBench 100%, OSWorld 2.0 at 72.6% in 47% less time, and a solo 24-hour completion of the video game Portal. But it's also the first release in this entire lineage where the lab's own safety researchers went on record worried the lever choice itself (reduced chain-of-thought transparency, in service of faster/cheaper reasoning) was actively regressing Lever 6 — safe autonomy and metacognition — the two stack layers this series has never had a strong lever for at all.
The ARC-AGI-3 harness controversy (62.7% standard harness vs. 99.9%+ OpenAI's own provider-adapter harness) is the most concrete evidence yet that Astra's headline gains are partly a Lever 3/5 story — reasoning-state preservation and evaluation-harness engineering — rather than a Lever 1 story of the base model simply knowing more.
12. 📊 The Full Picture: GPT Releases × Stack Layers
| Release | Reasoning | Generaliz. | Multimodal | Memory | Tool Use | Long-Horizon | Reliability | Metacog. | Safe Auton. |
|---|
| GPT-1/2 | | | | | | | | | |
| GPT-3 | | | | | | | | | |
| InstructGPT/ChatGPT | | | | | | | | | |
| GPT-4 | | | | | | | | | |
| GPT-4 Turbo/4o | | | | | | | | | |
| o1/o3 | | | | | | | | | |
| GPT-5/5.6 | | | | | | | | | |
| GPT-6 Astra | harness-dep. | | | | | | | flagged | |
🟢 Strong 🟡 Good-partial 🟠 Partial 🔴 Early/weak ⚪ Not yet present — editorial synthesis based on sourced release claims above
Read down the Metacognition and Safe Autonomy columns and the pattern is stark: eight years and roughly a dozen major releases, and neither column has ever gone past "partial." Every other column shows real, visible progress release over release. This is the single clearest empirical confirmation, from one lab's own history, of the original AGI Threshold thesis — the top of the stack simply hasn't had a working lever yet, for any release, ever.
13. 🔁 The Pattern: How OpenAI's Lever Choice Evolved
Scale
GPT-1→3
→
Post-Training
InstructGPT
→
Scale + Multimodal
GPT-4
→
Efficiency + Tools
Turbo/4o
→
Test-Time Compute
o1/o3
→
Scaffolding
GPT-5.x/Astra
Every pivot in this sequence happened after the previous lever visibly plateaued for that generation, not before — scale gains between GPT-3 and GPT-4 were still real, but InstructGPT's RLHF pivot happened specifically because a cheaper lever produced comparable perceived quality gains before another expensive scale-up was ready. Test-time compute arrived specifically once further pre-training scale started showing diminishing returns relative to cost. This sequencing is itself evidence for the diminishing-returns pattern our Six Levers post described generically — here it's visible as one company's actual R&D roadmap, not an industry aggregate.
14. 🔮 What This Predicts About What's Next for GPT
🧠
Memory/Continual Learning Is the Obvious Next Reach
Following the pattern, and echoing Ilya Sutskever's (a former OpenAI co-founder) "age of research" thesis, the next lever OpenAI reaches for once scaffolding gains plateau is most likely memory and continual learning — the one lever this lineage has never seriously pulled.
⚡
More Efficient Test-Time Compute, Not More of It
Given industry-wide movement toward uncertainty-aware compute allocation (under 0.6x naive cost per our Six Levers research), expect GPT's reasoning line to optimize thinking-time efficiency next, not simply longer chains.
🛡️
Safe Autonomy Gets Forced, Not Chosen
The Astra CoT-monitorability dissent suggests OpenAI's next release under this pressure isn't optional R&D — it's a response to its own researchers' public concern, making Lever 6 progress here reactive rather than proactively scheduled.
15. 🧭 Verdict
🎯 The Bottom Line
Eight years of one lab's releases show a lever sequence, not a single accelerating curve: scale, then post-training, then scale again, then efficiency and tools, then test-time compute, then scaffolding. Each pivot arrived exactly when the prior lever's returns started thinning — a pattern that is itself the strongest evidence in this entire series that no single lever, including the one currently in use, will carry indefinitely. And across every one of those eight years and every one of those pivots, two stack layers — metacognition and safe autonomy — have never had a working lever pulled for them at all. That's not a criticism unique to OpenAI; it's the same finding this site has reached from every angle so far. The question this post's history actually answers isn't "is GPT close to AGI" — it's "which lever gets tried next," and on the evidence of eight years of pattern, the honest answer is: whichever one hasn't been tried yet, applied to whichever layer is about to become the next bottleneck.