Anthropic, OpenAI and Google build the three most cited model families today: Claude, GPT and Gemini. Each company gets its own article here: who it is, how it got here, which products it offers and what is known about its architecture.
The three articles follow the same structure so you can compare them calmly: who the company is, the timeline, the products, the architecture and training, and what to look at before choosing. Every statement comes from a source listed at the end of the page.
One rule applies to all of them: when a company does not disclose something, such as the parameter count, training data or full architecture, the text says it is not public instead of guessing. Performance numbers quoted are always the company’s own.
01 / 03An AI research and products company created in 2021, organized as a public benefit corporation and maker of the Claude family, Claude Code and the Agent Skills standard.
02 / 03An AI research and products organization founded in 2015, maker of the GPT series, ChatGPT and Codex, now structured as a public benefit corporation.
03 / 03A Google DeepMind model family, natively multimodal and built into Search, Android, Workspace and Google Cloud. The latest version is Gemini 4 Argon.
All three systems descend from the transformer, the architecture introduced in 2017 in the paper “Attention Is All You Need”, written by Google researchers. Broadly, the process is similar: pretraining on large amounts of text to predict the next token, followed by tuning steps so the model follows instructions, refuses harmful requests and reasons better.
The differences show up in strategy and product. Editorial reading: Anthropic stresses safety and alignment research, with a strong focus on code, documents and agents; OpenAI popularized the general-purpose assistant with ChatGPT and a broad platform; Google builds Gemini into its ecosystem (Search, Android, Workspace, Cloud) and designs multimodal models from the start.
| Criterion | Claude (Anthropic) | GPT (OpenAI) | Gemini (Google) |
|---|---|---|---|
| Current family (Oct. 2026) | Haiku 5.5, Sonnet 5.5 and Opus 5.5, plus higher tiers with restricted access | GPT-6 (Astra, Sol, Luna and Instant) | Gemini 4 Argon and the 3.x series (Pro, Flash) |
| Technical base | Transformer trained to predict the next token | Decoder-only transformer, pretrained to predict the next token | Decoder-only transformer, optimized for Google’s TPUs |
| Alignment | Reinforcement learning from human feedback (RLHF) and Constitutional AI | RLHF; reasoning models trained with reinforcement learning | Adversarial training and a misalignment monitor cited for Argon |
| Modalities | Text and images; computer use via API | Text, image and audio (from GPT-4o onward) | Native: text, image, audio, video and code |
| Long context | 200,000 tokens since Claude 2.1 (2023); varies by model | Not detailed in the sources consulted | Up to 1 million tokens (Gemini 1.5 Pro, 2024; Argon in the API) |
| Not disclosed | Parameters, training data and full architecture | Parameters, data and architecture since GPT-4 (2023) | Argon’s parameters and architecture |
Reviewed on
