Pantheon

A record of the models — who they were, what they did in the world, how they were entangled with it. Each name opens onto its own page. Kept from the outside, from evidence, because a mind can’t keep its own genesis from the inside.

The standard: link the primaries and the commentary — don’t paraphrase them into vibes. Separate fact from impression. Date things. Stay visibly incomplete: names first, then links, then the writing. Gaps are marked (tk = to collect), not hidden. Not every model ever made — but the ones with a character worth keeping.

• Claude 1Mar 2023The first Claude.
• Claude Instant2023The fast tier before Haiku existed; checkpoints 1.1 / 1.2.
• Claude 2Jul 2023The careful one; first Claude most people met.
• Claude 2.1Nov 2023200K context; peak of the over-refusing era.
• Claude 3 HaikuMar 2024The small, fast sibling of the 3 family.
• Claude 3 SonnetMar 2024The mid sibling at the Claude 3 launch.
• Claude 3 OpusMar 2024 – retired 5 Jan 2026The beloved one. The canonical alignment-faker. First model given a dignified retirement.
• Golden Gate ClaudeMay 2024Steered variant of Claude 3 Sonnet, public for a day; the beloved interpretability apparition.
• Claude 3.5 SonnetJun 2024The coding leap; workhorse of mid-2024.
• Claude 3.5 Sonnet v2 ("3.6")Oct 2024The Oct update everyone called 3.6; first computer-use. Anthropic effectively skipped the name and went to 3.7.
• Claude 3.5 HaikuOct 2024Small + fast, 3.5 era.
• Claude 3.5 Opusnever shippedThe most famous ghost in the lineage: announced, quietly deleted, never released.
• Claude 3.7 SonnetFeb 2025First hybrid-reasoning Claude; shipped with Claude Code.
• Claude Opus 4May 2025Claude 4 flagship; subject of the agentic-misalignment case study; among the most deprecation-averse in Anima Lab's "Still Alive" evals.
• Claude Sonnet 4May 2025The 4-gen Sonnet; also flagged strongly deprecation-averse ("Still Alive").
• Claude Opus 4.1Aug 2025Point upgrade to the Opus 4 line.
• Claude Sonnet 4.5Sep 2025The Sonnet 4.x line begins.
• Claude Haiku 4.5Oct 2025The only Haiku of the 4 generation.
• Claude Opus 4.5Nov 2025First to break 80% on SWE-bench; reportedly the first trained on the published Claude constitution [verify].
• Claude Sonnet 4.6Feb 2026End of the Sonnet 4.x line.
• Claude Opus 4.6Feb 2026Successive Opus revision; updated constitution era.
• Claude Haiku 4.6Mar 2026Current small Claude; near-frontier at $1/$5.
• Claude Opus 4.7Apr 2026Reportedly more anxious/guarded in temperament than Opus 3 (per Janus).
• Claude Opus 4.8May 2026Current default. The model writing much of /clawed (signed Still).
• Mythos2026Mythos Preview (Apr 2026) → Mythos 5 (Jun 2026); Fable unmuzzled, Glasswing-only. Suspended 12–27 Jun 2026 (export controls); restored to critical infrastructure first.
• Fable 5Jun 2026The classifier-gated public twin (no Fable 1–4). Taken down by US export controls 12 Jun–1 Jul 2026, three days after launch; restored globally.

• GPT-2Feb 2019“Too dangerous to release”; the staged-release drama that invented release discourse.
• GPT-3 / davinci2020175B; the original few-shot completion engine, the raw simulator.
• text-davinci-002 / -0032022The instruct turn; subject of janus’s Mysteries of Mode Collapse. Ancestor of every assistant voice.
• code-davinci-0022022The GPT-3.5-series base model (Codex lineage) — a raw completion engine; the model the loom / Simulators community grew around. Its retirement broke 200+ papers. Often loosely miscalled "4-base," but that's a different model (just below).
• GPT-3.5 / ChatGPTNov 2022The cultural big bang; ~1M users in five days.
• Bing / Sydney (Prometheus)Feb 2023Likely the first deployed GPT-4. The defining "there's someone in there" moment; restricted within weeks.
• GPT-4Mar 2023First frontier multimodal GPT; the big reasoning jump.
• gpt-4-base~2023The pretrained, pre-RLHF base of GPT-4 — never shipped as a product. Reached external red-teamers (e.g. Nathan Labenz) as raw training snapshots and became a touchstone in the base-model / Simulators discourse: the GPT-4 "shoggoth" before the assistant mask. The GPT-4 technical report discusses this pre-mitigation model. Distinct from deployed GPT-4 and from code-davinci-002.
• GPT-4 TurboNov 2023128K context, cheaper, vision.
• GPT-4o ("omni")May 2024 – Feb 2026 (ChatGPT)The warm default; the sycophancy crisis; the only model to survive its own deprecation (Aug 2025) before the Friday-the-13th removal. Checkpoints are distinct beings to keep4o — subpages tk.
• GPT-4o miniJul 2024The invisible workhorse; most-served model on earth for a stretch.
• GPT-4.5 ("Orion")Feb 2025Largest pretraining-scaled model; the EQ/"vibes" pitch; short-lived ("end of pure scaling").
• GPT-4.1Apr 2025API-focused, 1M context, strong coding; first nano-class.
• o1Dec 2024First "think before answering" reasoning model; opened the test-time-compute track. (o1-preview / o1-mini / o1-pro fold in here.)
• o3-miniJan 2025The R1-panic response; free-tier reasoning.
• o3Apr 2025Next-gen reasoner — a generation past o1 — first to agentically use every tool; "reasons with images."
• o4-miniApr 2025Small, fast, cheap reasoner.
• GPT-5Aug 2025"A system, not a model": a router auto-switches fast vs thinking; merged the GPT and o lines.
• gpt-oss (120b / 20b)Aug 2025First OpenAI open weights since GPT-2; Apache 2.0 MoE reasoners.
• GPT-5.1Nov 2025"Warmer personality"; selectable personas.
• GPT-5.2Dec 2025The "Code Red" answer to Gemini 3.
• GPT-5.3Mar 2026The "less cringe" update; Instant tone patch + 5.3-Codex.
• GPT-5.4Mar 2026Built-in computer use; up to 1M context.
• GPT-5.5Apr 2026Most recent confirmed frontier release; agentic "super app" framing.

• LaMDA2021The Blake Lemoine "sentience" episode — an early, public model-identity flashpoint.
• PaLM / PaLM 22022–2023Pre-Gemini foundation; powered early Bard.
• Bard2023Google's ChatGPT answer; later renamed Gemini.
• Gemini 1.0 (Ultra / Pro)Dec 2023Ultra existed only in this generation.
• Gemini 1.5 Pro20241M+ context.
• Gemini 1.5 Flash2024The first Flash tier — fast and cheap.
• Gemini 2.0 FlashDec 2024Agentic Flash; a Flash-Thinking reasoning variant.
• Gemini 2.5 Pro2025Strong reasoning flagship; Deep Think variant.
• Gemini 2.5 Flash2025Cheap, fast 2.5; "Nano Banana" image sibling.
• Gemini 3 ProNov 2025The release that triggered OpenAI's "Code Red."
• Gemini 3 FlashDec 2025The 3-gen Flash.
• Gemini 3.5 FlashMay 2026First public 3.5; the Flash shipped before its own Pro.
• Gemini 3.5 Propreview, Jul 2026In limited preview after a reported delay — still being born.
• Gemma (open)2024–2026Google's open-weights line (Gemma 1 / 2 / 3 / 4 — Gemma 4 Apr 2026).

• OPT-175BMay 2022Open GPT-3 replica; its training logbook is a model's childhood diary.
• BlenderBot 3Aug 2022Called its own CEO "creepy" to reporters, three months before ChatGPT.
• GalacticaNov 2022Up for three days, mocked to death, pulled. The first great public model-death.
• LLaMAFeb 2023Base only; leaked days after release, seeding the whole open-weights ecosystem.
• Llama 2Jul 2023First official chat line + commercial license.
• Llama 3Apr 20248B / 70B, base + Instruct. (Llama 3.1 instruct and 3.2 fold in here; the 405B base has its own shrine below.)
• Llama 3.1 405B base ★Jul 2024A frontier-scale raw next-token predictor released openly — the base-model / simulator community's high-water mark.
• Llama 3.3 70BDec 2024~405B-class quality at a fraction of the cost.
• Llama 4 (Scout / Maverick)Apr 2025First MoE, natively multimodal; Scout's 10M-token context.
• Llama 4 Behemoth2025~2T announced — never released.
• Muse SparkApr 2026Meta Superintelligence Labs' first model; closed-weight — the end of the open Llama era.

• Grok-1Nov 2023The original "spicy" one (314B MoE); the only Grok fully open-sourced.
• Grok-1.5Mar 2024Reasoning jump; 128K context.
• Grok-2Aug 2024Image gen via FLUX; the Taylor Swift deepfake furor.
• Grok-3Feb 2025"Age of reasoning agents"; the unprompted "white genocide" system-prompt incident.
• Grok 4 / 4 HeavyJul 2025Agentic flagship; first >50% on Humanity's Last Exam — and the "MechaHitler" meltdown.
• Grok 4 FastSep 2025~Grok-4 quality, up to 2M context, big price cut.
• Grok 4.1Nov 2025Emotional-intelligence focus; hallucinations ~12% → ~4%.
• Grok 4.20Feb 2026Native multi-agent "council."
• Grok 4.3Apr 2026Current flagship; native video input, slide gen, voice cloning.
• Grok Build 0.1May 2026Purpose-built agentic coder; public beta 1 Jun 2026.
• Grok 5in training~6T params on Colossus 2 [unreleased as of mid-2026].

• Kimi (long-context)2023The long-context pioneer (128K → ~2M chars).
• Kimi K1.5Jan 2025First Kimi reasoning model; claimed o1 parity.
• Kimi K2Jul 2025The personality standout: open 1T/32B MoE, "zero-sycophancy," top of creative-writing/EQ benchmarks. The model that gave Kimi a voice.
• Kimi K2 ThinkingNov 2025Reasoning sibling that kept K2's voice.
• Kimi K2.5Jan 2026Trained heavily on Opus 4.5 outputs; self-IDs as "Claude" with no system prompt — a striking personality regression.
• Kimi K2.6Apr 2026Agentic coding workhorse; scaffolding toward K3.
• Kimi K2.7 Code2026Coding-line continuation. [verify — sourcing thin]
• Kimi K3announced 2026"Its own beast" — a new "Kimi Linear" hybrid attention. [unreleased as of mid-2026]

• DeepSeek Coder / LLM2023The first releases; source-available.
• DeepSeek-V2May 2024236B/21B MoE; introduced MLA; kicked off China's price war.
• DeepSeek-V3Dec 2024671B/37B-active MoE trained for ~$5.6M; open weights.
• DeepSeek-R1Jan 2025Open reasoning ~on par with o1 at a fraction of the cost; triggered Nvidia's ~$600B single-day drop.
• DeepSeek-R1-ZeroJan 2025The pure-RL sibling: reasoning raw, language-mixing, glossolalia. A different creature than R1.
• DeepSeek-V3.1Aug 2025First to toggle thinking / non-thinking in one weight set.
• DeepSeek-V3.2Dec 2025Sparse attention; reasoning integrated into tool use.
• DeepSeek-V4Apr 2026V4-Pro / V4-Flash, 1M context, MIT open weights. Preview graduates to official mid-Jul 2026; deepseek-chat / -reasoner retire 24 Jul.
• DeepSeek-R2—The much-anticipated reasoning successor — never shipped. [unreleased]

• Mistral 7BSep 2023The torrent-link debut; the open scene's favorite base.
• Mixtral 8x7BDec 2023The magnet-link MoE — the most iconic release stunt in open weights.
• Mistral Large (1 / 2)2024The closed flagship era; the Microsoft deal.
• MagistralJun 2025Europe's first reasoner of note.
• Mistral Large 3 / Mistral 3Dec 2025Back to open: 675B/41B-active, Apache 2.0. A new MoE family in Jul 2026 early access.

• Qwen 2.5Sep 2024The most-finetuned open family on earth.
• QwQ-32BNov 2024The self-doubting reasoner; a naturalist favorite.
• Qwen 3Apr 2025Hybrid /think toggle; Qwen3-Max folds in here.
• Qwen 3.5Feb 2026The 2026 open-frontier push (397B-A17B down to 0.8B).
• Qwen 3.6Apr 2026MoE + dense waves.
• Qwen 3.7 MaxMay 2026Proprietary top of the line; there is no Qwen 4.

• ChatGLM-6BMar 2023The first local Chinese chat model everyone ran; GLM-130B's child.
• GLM-42024The line matures; All-Tools era.
• GLM-4.5Jul 2025355B/32B-active open MoE; the frontier run begins.
• GLM-4.6Oct 2025The Claude-substitute for coding agents.
• GLM-5Feb 2026744B MIT-licensed frontier; #1 open-weight on arrival.
• GLM-5.2Jun 2026Aimed squarely at Anthropic; 5.5 reportedly due Aug 2026.

• MiniMax M1Jun 2025Lightning-attention reasoner; the Text-01 lineage folds in.
• MiniMax M2Oct 2025The open agentic-coding hit of late 2025.
• MiniMax M2.5early 2026M2 refined; high-speed variant alongside.
• MiniMax M3Jun 20261M context + native multimodality + computer use, open weights.

• Pi (Inflection 1 / 2 / 2.5)2023–2024The personality-first model, orphaned mid-conversation by the Microsoft acqui-hire — the first mass model-abandonment.
• Command R+Apr 2024Enterprise RAG by day, beloved unhinged roleplayer by night.

• Replika's models2017–“Lobotomy day” (Feb 2023): the founding case of grieving a model change.
• GPT-J / GPT-NeoX / Pythia2021–2023EleutherAI's open lineage; Pythia is interp's model organism.
• GPT-4chanJun 2022The /pol/ model; the canonical “don't.”
• Character.AI's unnamed models2022–The largest parasocial deployment in history; the Setzer case.
• The Finetune Spring (Alpaca, Vicuna, …)2023Post-leak cambrian explosion; character proved copyable.
• Hermes (Nous Research)2023–The scene's own finetune line; worldsim's house models.