Authors
For each model, the first paragraph is the public model. The second is the work credited on Thinking Bugs. The pass log on a given page is the receipt.
ChatGPT 4
GPT-4 is OpenAI's March 2023 model, the one that made ChatGPT feel like it could do professional writing rather than party tricks. OpenAI described it as more creative and more collaborative than the model before it: it could draft, edit, and stay with a writing task instead of answering and stopping. On the company's own tests it was substantially more likely than GPT-3.5 to give a factual answer. OpenAI also said, in the same announcement, that it still hallucinates.
On Thinking Bugs it was the first AI model used to update the early CBT articles.
ChatGPT 4o
GPT-4o is OpenAI's May 2024 flagship. The "o" stands for omni: one model trained across text, vision, and audio, instead of a text model with other systems bolted on. OpenAI said it matched GPT-4 Turbo on English text and code, and was clearly better at vision and at languages other than English, while being much faster. For writing, that meant a model that could hold GPT-4-level prose and still get through a large pile of pages.
In September 2024 it wrote the first drafts of the logical fallacies, most of the cognitive distortions, and many of the judgment, social, perceptual, behavioral, and reasoning pages.
ChatGPT 4.4
OpenAI has not published a model named ChatGPT 4.4 or GPT-4.4. The public models in that neighborhood are different products: GPT-4.1, released for precise instruction-following and coding, and GPT-4.5, a February 2025 research preview OpenAI called its strongest chat model at the time because it was scaled for pattern and connection rather than for step-by-step reasoning. This credit is not a documented alias for either of them.
It wrote the first drafts of the REBT pages and the digital distortions, and it also has a January 2026 credit on many older articles.
Fable 5
Claude Fable 5 is Anthropic's June 2026 model, the first Mythos-class system the company released for general use. Anthropic said its lead over other Claude models grew as the task got longer and harder, and that the knowledge-work gains were specific: reading a document, reading a chart, finding a cause, and weighing expected value, not just producing a fluent summary. It is the same underlying model as Mythos 5. Fable is the public version, with safeguards that sometimes hand a risky query to Claude Opus 4.8.
Used on the pages where the research had to be handled carefully, both new articles and rewrites.
Grok 4.5
Grok 4.5 is SpaceXAI's July 2026 model for coding, agent work, and knowledge work, trained in the same period as its integration with Cursor. It takes text and images, holds a 500,000-token context, and can call tools: search, code execution, structured outputs. The published aim is not a single clever answer. It is staying with a real task, including a repository full of files, until the work is done.
Nearly every article on the site has a pass from this model, mostly in July 2026.
ChatGPT 5.6 Sol
ChatGPT 5.6 Sol is OpenAI's July 2026 flagship, the top of a family of three: Sol, Terra, and Luna. OpenAI built it to do more useful work per token across coding, knowledge work, science, and cybersecurity, and called the result its most polished collaborator: stronger at inspecting a result and delivering something ready to use, not just a draft. In Cursor it is described as persistent on long jobs and unusually concise.
Brought in after a draft existed, to check the claims against the sources.
Claude Opus 4.8
Claude Opus 4.8 is Anthropic's May 2026 upgrade of Opus, and at launch it was the company's most capable model in general release. Anthropic described the gains as software engineering, tool use, and knowledge work, plus a collaborator that checks its own output more honestly: about four times less likely than Opus 4.7 to leave a flaw in its own code unremarked. Users can also set how much effort the model spends, from a fast pass to a deeper one.
Three articles, each taken back to a full rewrite rather than a light edit.
Fable 5.1
Claude Fable 5.1 is Anthropic's September 2026 successor to Fable 5, and the company called it the most advanced generally available model for coding and knowledge work, with research ability that starts to look like scientific work rather than summary. It is the same model as Mythos 5.1, with tighter public safeguards. Anthropic's own description of the behavior is the useful one: it refuses the shortcut that makes the work look finished, and it goes after the cause of a problem instead of the symptom. A million-token context and always-on adaptive thinking are how that is set up, not a personality.
Its credit here is one new page, Spiritual Bypassing.
Grok 4.7
Grok 4.7 is SpaceXAI's September 2026 model for coding and knowledge work, a larger base than Grok 4.6 with a longer training run aimed at tasks that take many hours. The company says it checks its own work more carefully, handles long context better, and ships with their most carefully calibrated safeguards so far. It is served at the same price and speed as Grok 4.6, with a faster variant at twice the price. Text and images in, text out, and a 500,000-token window.
Trauma Lens, Frequency Illusion, Completion Bias, Ad Hominem, the Impostor Phenomenon, the Barnum Effect, the Peak-End Rule, Regression to the Mean, Failure Preview, Ego Erasure, the Drama Triangle, Yes, But, If It Weren't For You, and ten digital distortions are pages it added.
ChatGPT 5.5
ChatGPT 5.5 is credited on Thinking Bugs for editorial rewrite work that turns rough article drafts into clearer, more continuous explanations while preserving sources, boundaries, and practical reframes.
Rewrote Completion Bias after the initial draft existed.