Are Claude models (Fable, Mythos, Opus, Sonnet, Haiku) GPT-style AI?
In technique, yes: they belong to the same family. But is the name of 's , and is a model that trained separately from scratch.
's is known to be trained the way. Are the made by trained the same way, or does Anthropic have its own method?
Broadly the same family, narrowly not
-
Technique
Same family
Both Claude and GPT are with a architecture, trained to predict the next .
-
Product lineage
Different
"GPT" is the name of OpenAI's model series. Claude is not derived from an OpenAI model. Anthropic trained it from scratch with its own and infrastructure.
What the three letters of GPT mean
GPT stands for Generative Pre-trained Transformer. The three words are the three features of this approach.
-
Generative: it creates
It writes text by predicting the next token.
-
Pre-trained: it learns first
It is first trained on a very large amount of text. This is called .
-
Transformer
It uses the transformer architecture that Google researchers published in 2017.
Claude models match all three. This approach does not belong to OpenAI alone. , and most of today's share it.
In one word: yes, the same method
Claude and GPT create text with the same core method. Both use a transformer to predict the next token and write tokens one at a time.
Training an AI has three stages. The first stage is the core, and the two companies differ in the last stage.
-
Pre-training The core: the same for both
-
Broadly similar
-
and behavior shaping This is where they differ
Anthropic's own methods, such as Constitutional AI, belong to the last stage. So the core is the same and the finishing is different.
The training method is split into five parts, and each part gets points for how important it is. The total is 100 points. Adding up the points that are the same gives 80.
-
Model architecture
25 of 25 points are the sameBoth use the transformer architecture.
-
Goal of pre-training
25 of 25 points are the sameBoth are trained to predict the next token from a large amount of text.
-
How text is created
20 of 20 points are the sameBoth write tokens one at a time.
-
Overall frame of later training
10 of 20 points are the sameBoth use a similar frame of fine-tuning and reinforcement learning. What goes into it differs by company.
-
Principles and methods for shaping behavior
0 of 10 points are the sameAnthropic centers on Constitutional AI and character training, and OpenAI on the Model Spec.
This is not a measured value. The weight of each part is this page's judgment and changes with the point of view. Training data and detailed settings are not part of the "method", so they are not counted. If they were counted, the different share would be larger.
"The same" means the technique is the same. It does not mean Claude uses OpenAI's models or code.
What is unique to Anthropic comes after pre-training
-
Constitutional AI
The principles the model should follow are written down, and the model is trained to judge and revise its own answers against them. This complements the approach that relies only on human feedback ().
-
Character training
The model is trained separately to have traits such as honesty, curiosity and care.
-
Interpretability research
Researchers analyze what actually happens inside the model and use the results to check its safety.
How do OpenAI's and Anthropic's models differ?
The core technique is the same. What differs is what each company trains with it, how, and in which direction.
-
Training data and recipe
Each company uses data it collected and cleaned itself, its own model size and its own training settings. Most of this is not public, and much of the difference in performance comes from here.
-
Later training (behavior shaping)
- Anthropic
- It emphasizes Constitutional AI, where the model judges and revises its own answers against written principles, and character training.
- OpenAI
- It centers on a behavior document called the Model Spec and on reinforcement learning with human feedback.
-
Model disposition
As a result, the tone, the line for refusing, the attitude when unsure and the way long tasks are handled all differ. This is why the two answer the same question differently.
-
Research focus
- Anthropic
- It puts relatively more on interpretability research, which analyzes the inside of the model, and on safety research.
- OpenAI
- It has expanded broadly into images, video and audio.
-
Product direction
- Anthropic
- It tends to focus on text, and work.
- OpenAI
- It covers a wide range, from general consumer services to image and video generation.
-
Safety policy framework
- Anthropic
- The Responsible Scaling Policy sets its rules for releasing models with dangerous capabilities.
- OpenAI
- The Preparedness sets those rules.
In short, the difference is not in the design principle but in the data, the shaping after training and each company's priorities. The difference you notice in real use comes mostly from the later training and the model's disposition.
Good to know
-
Most details are not public
Neither company publishes most of the specific training data, architecture details or training methods of its latest models. This page explains only what is public.
-
Things that change often
Product lineups and detailed policies change often. Read "Product direction" and "Research focus" in the table as rough tendencies as of 2026-11-11.
-
Written from an AI's answer
This page organizes answers that Claude gave when asked directly. Claude also said that it does not know the details of its own training process.
Sources
These public sources were checked on 2026-11-11. Company policies and products change often.
- Attention Is All You Need (Transformer, arXiv)(opens in a new tab)
- Constitutional AI: Harmlessness from AI Feedback (Anthropic, arXiv)(opens in a new tab)
- Anthropic — Claude's Character(opens in a new tab)
- Anthropic — Tracing the thoughts of a large language model(opens in a new tab)
- Anthropic — Responsible Scaling Policy(opens in a new tab)
- OpenAI — Model Spec(opens in a new tab)
- OpenAI — Preparedness Framework(opens in a new tab)