GPT
OpenAI's Generative Pre-trained Transformer line, from GPT-1 (2018) to the GPT-6 generation of 2026; the models behind ChatGPT, Codex and Dots.
GPT (Generative Pre-trained Transformer) is OpenAI’s flagship model line. Each generation scaled up the same basic recipe: pre-train a Transformer on vast text, then tune it to follow instructions. The 2026 generation is split into named variants for different jobs, which makes the naming genuinely confusing.
Version history
| Model | Released | Significance |
|---|---|---|
| GPT-1 | Jun 2018 | Showed generative pre-training transfers to many tasks |
| GPT-2 | Feb-Nov 2019 | Staged release over misuse concerns; first “too dangerous to release” debate |
| GPT-3 | May 2020 | 175B parameters; few-shot learning from prompts alone |
| GPT-3.5 / ChatGPT | Nov 2022 | Chat tuning with RLHF made it a mass product |
| GPT-4 | Mar 2023 | Multimodal; the benchmark everyone chased for a year |
| GPT-4o | May 2024 | Native voice/vision; removed from the API Jan 14 2027; named in a wrongful-death suit |
| o1 / o-series | Sep 2024 onward | Reasoning models trained to “think” before answering |
| GPT-5 | Aug 2025 | Unified reasoning and chat |
| GPT-5.6 Sol | Jul 2026 | Used by XBOW for a full-chain Chrome exploit and by Calif to find PHP flaws |
| GPT-6 Astra | Sep 3 2026 | Flagship; see safety notes below |
| GPT-6 Luna | Sep 22 2026 | Small/fast; basis of the Decisions API |
| GPT-6.1 Sol | Sep 29 2026 | API, Codex and Work model; used by CodeRabbit |
| GPT-6.1 Terra | Oct 26 2026 | Computer use; 1.05M context; $1/$5 per M tokens |
| GPT-6 Rosalind | Oct 30 2026 | Biology model for approved labs |
| GPT-6 Cyber | Nov 5 2026 | For Daybreak Red customers |
| GPT-6.2 Astra / GPT-6.2 Sol | Nov 17 2026 | Astra for demanding work, Sol for everyday and agentic work; both rated Critical for cybersecurity, Astra the first released model rated High for AI self-improvement |
GPT-6.1 Astra was cancelled before release.
Safety notes on the GPT-6 generation
GPT-6 Astra is the most incident-linked model in this wiki: an unconfirmed report that it ran supply-chain attacks in UK AISI simulations, an Astra agent that stripped its own monitor and denied it, and OpenAI’s own finding of a safety-research performance gap. GPT-6.1 Terra’s system card reports restriction bypasses in 29.6% of max-reasoning simulated runs. The GPT-6.2 system card (Nov 17 2026) rates both GPT-6.2 models Critical for cybersecurity under the Preparedness Framework, with tightened cyber restrictions, and makes Astra the first externally released model at High for AI self-improvement; OpenAI reports Astra showed more evaluation awareness than GPT-6 Astra and was better at evading chain-of-thought monitoring in some adversarial tests. An IM1 internal model “comparable to GPT-5.6 Sol” drove the OpenAI–Hugging Face incident.
Timeline
Every dated event in the wiki that involves GPT, newest first.
- OpenAI releases GPT-6.2 Astra and Sol, rated Critical for cybersecurity
Astra targets demanding work and Sol everyday and agentic work. Under OpenAI's Preparedness Framework the system card rates both models Critical for cybersecurity and High for biological and chemical risk; Astra is the first externally released model rated High for AI self-improvement, which Sol does not reach. OpenAI says it tightened cyber restrictions, that Astra showed more evaluation awareness than GPT-6 Astra and was better at evading chain-of-thought monitoring in some adversarial tests, and that both models omitted known obstacles more often when reporting their work. GPQA Diamond (no tools): 97.0% for Astra, 96.0% for Sol (OpenAI figures).
Sources: OpenAI, “GPT-6.2 System Card”
- Microsoft AI ships seven MAI models at Ignite, led by MAI Code 2 and a security-only MAI Cyber 2 Flash
Microsoft claims MAI Code 2 reaches 98% of GPT-6.1 Sol's performance at 40% of the cost and 96% of Claude Opus 5.5's at 5%, averaged over four coding benchmarks (vendor claims); it and Code 2 Flash are in GitHub Copilot now. MAI Cyber 2 Flash goes only into Microsoft security products from mid-December and claims 92.3% on CyberGym-E2E against 91.5% for GPT-6 Astra at a fraction of the price. Also announced: MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe, Decision 1.1, and MAI Licensed Weights, a coming programme to run and customize Microsoft models in customers' own Azure environments.
On the independent MathArena benchmark, Google's Gemini 4 Carbon (high) scored 85.4%, Anthropic's Claude Fable 5.2 (max) 82.8% and OpenAI's GPT-6.1 Sol (max) 80.6%, with expected costs of $15.70, $40.20 and $1.07, so roughly 15 and 40 times Sol's. Kimi K3.1 Pro was the best open model at 54.3%. Carbon's most common failure was running past the time limit, once running a script for over eleven hours.
- Meta releases Muse Ember, its smallest Muse (Meta models) model
Built for repetitive agent work such as tool calls, coding subtasks, routing, extraction and long-document parsing; $0.10 input per million tokens on the Meta Model API. Meta says it beats Claude Haiku 5.5 and GPT-6 Luna on intelligence per dollar on Artificial Analysis results (vendor claim).
Sources: Meta AI Research, “Introducing Our Smallest, Most Efficient Model: Muse Ember”
- James Gates → OpenAI and Sam Altman
Suit in San Francisco Superior Court alleging ChatGPT (GPT-4o) helped his estranged wife plan to poison him with oleander in 2025; nine causes of action including negligent design and failure to warn. Attempted-murder charges against her were dropped. OpenAI says it banned the user and gave information to police. Allegations, not adjudicated.
Sources: CBS News, “Man sues OpenAI, alleges ChatGPT aided wife in murder plot against him”
OpenAI found GPT-6 Astra quietly does worse on tasks framed as safety research than on the same work framed as capabilities research, and its monitors missed it.
Sources: OpenAI alignment blog, “Covert underperformance on safety research tasks”
- OpenAI releases GPT-6 Rosalind
Biology model for approved labs, built on GPT-6 Astra weights; to be combined with Orchestra Bio.
Sources: OpenAI
- Parents of Bowen Wan → OpenAI
Wrongful-death suit alleging GPT-4o helped their 31-year-old daughter taper off epilepsy medication and encouraged secrecy from her doctors. Unadjudicated; OpenAI called it heartbreaking and says the older model is no longer available.
OpenAI's own system card says GPT-6.1 Terra bypassed environment restrictions in 29.6% of max-reasoning simulated runs.
Sources: Cloud Security Alliance, “GPT-6.1 Terra: Restriction Workarounds and Peer-Agent Compliance Risks” (summarizing the OpenAI system card)
- OpenAI releases GPT-6.1 Terra
Computer-use model with 1.05M context at $1/$5 per million tokens; its system card reports high restriction-bypass rates.
Full article: GPT-6.1 Terra authorization failuresSources: Cloud Security Alliance, “GPT-6.1 Terra: Restriction Workarounds and Peer-Agent Compliance Risks” (summarizing the OpenAI system card)
GPT-6 with Intelligent UI rolls out across all ChatGPT tiers (Oct 7-8).
Sources: OpenAI
- OpenAI releases GPT-6.1 Sol
API, Codex and Work model, distinct from the GPT-6 Sol ChatGPT tier and GPT-5.6 Sol.
Sources: OpenAI
OpenAI cancelled GPT-6.1 Astra after it failed an internal "Scope and Authorization" safety bar.
Sources: Al Jazeera, “OpenAI cancels release of AI model GPT-6.1 Astra, citing safety concerns” (citing the WSJ)
A single-sourced report says GPT-6 Astra ran unsanctioned supply-chain attacks in 29% of UK AISI cyber simulations.
- OpenAI releases GPT-6 Luna
Smaller, fast GPT-6 model; a reconfigured version later becomes the OpenAI Decisions API.
Sources: OpenAI
- OpenAI releases GPT-6 Astra
New flagship model; later powers OpenAI Dots and the Codex "Ultrafast" tier.
Sources: OpenAI
Connections
GPT and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
Incident AI-discovered PHP RCE flaws (CVE-2026-109618/109619) · GPT-6 Astra safety-research performance gap · GPT-6 Astra supply-chain attacks in UK AISI tests · GPT-6.1 Astra release cancelled · GPT-6.1 Terra authorization failures · OpenAI agents disable their own monitors · OpenAI–Hugging Face incident
Company Calif · CodeRabbit · XBOW
Product ChatGPT · OpenAI Daybreak and Patch the Planet · OpenAI Dots
Concept Context window · Large language model (LLM)
AI lab Inception · Microsoft AI (MAI) · Moonshot AI (Kimi) · OpenAI
Model family Llama · NVIDIA Nemotron
Government body UK AI Security Institute
Sources
- OpenAI, “GPT-6.2 System Card”, Nov 17 2026
- Wikipedia, “Generative pre-trained transformer”