Lesson 9 of 9 · about 18 minutes
Who Makes What: Model Families as of October 2026, and a Working Glossary
You'll be able to
- Name the major AI model makers and their main model families as of October 2026
- Tell which families are closed and which release open weights
- Read a model name and a launch announcement with a skeptical eye
- Use the module glossary to decode AI news
Everything in the tables below is as of October 2026. Model names were checked against each company's own documentation that month. A few details come only from news reports, not from the company, and are marked "(secondary source)" with the outlet named. This field moves in weeks. Anthropic alone shipped at least five new models between June and October 2026. Before you rely on any name here, check the company's own model page.
A handful of companies train the most capable models, because training one costs hundreds of millions of dollars in chips and electricity. A much larger group builds products on top of those models or releases smaller open ones. Here's who's who.
The big general-purpose families
| Company | Main family (as of October 2026) | Weights | What to know |
|---|---|---|---|
| OpenAI (US) | GPT-6 line: OpenAI's developer docs list GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna as flagships. Consumer app: ChatGPT. | Closed; also open-weight gpt-oss-120b and gpt-oss-20b (Apache 2.0, August 2025) | Made ChatGPT. GPT-6 Astra was announced September 3, 2026, with few technical details at first (secondary source: AIToolly news report). |
| Anthropic (US) | Claude: Fable 5.1 (most capable, September 1, 2026), Opus 5.5, Sonnet 5.5, Haiku 5.5 | Closed | Fable 5 launched June 9, 2026. Opus 5.5 (September 22), Sonnet 5.5 (September 28) and Haiku 5.5 (October 7) followed. Anthropic lists a 1 million token context window for all four current models. |
| Google DeepMind (US/UK) | Gemini 4 Argon (stable) and Gemini 3 series (3.8 Flash and others stable, 3.1 Pro in preview). Consumer app: Gemini. | Gemini closed; Gemma 4 open-weight | Gemini 3 Pro launched in preview in November 2025. Google released Gemini 4 Argon on October 13, 2026, in the Gemini API and in the Gemini app for Google AI Pro and Ultra subscribers. |
| Meta (US) | Muse Wonder and Muse Wonder Full, Muse Spark (version 1.3 listed), Muse Spark Open, Llama 4 | Muse Wonder and Muse Spark closed; Muse Spark Open, Muse Glimmer and Llama 4 Scout and Maverick open-weight | Muse Spark (April 8, 2026), from Meta Superintelligence Labs, was Meta's first closed frontier model. Meta announced Muse Wonder and Muse Wonder Full on October 12, 2026, and released Muse Spark Open with open weights in October 2026. Meta's developer site lists Muse Spark Open and Muse Glimmer as open-weight. |
| xAI (US) | Grok 4.7 | Closed | Built into X (formerly Twitter). Also offers Grok Imagine image and video models. |
| DeepSeek (China) | DeepSeek V4 (V4-Pro and V4.1-Flash, which replaced V4-Flash) | Open weights on Hugging Face | Its January 2025 R1 reasoning model jolted the industry by matching top US models at far lower reported cost. Its API lists a 1 million token context. |
| Alibaba (China) | Qwen3 series, latest Qwen3.8 | Mostly open; Qwen3.8-2.4T-A95B weights posted on Hugging Face in August 2026; the fuller Qwen3.8-Max is served by Alibaba | A widely used open family. Qwen3.8, previewed July 19, 2026 (secondary source: EqualOcean), has 2.4 trillion total parameters per its model card. |
| Mistral AI (France) | Mistral Large, Medium, Small; Ministral (small); Codestral (code) | Many under Apache 2.0, such as Mistral Small 4, Mistral Large 3 and Ministral 3; others, such as Codestral, offered as "Premier" API models | The best-known European model maker. |
Most companies sell each generation in sizes. Anthropic's naming makes it plain: Haiku is small, fast and cheap; Sonnet is the middle; Opus and Fable are the largest. Google uses Flash and Flash-Lite for fast models and Pro for the big one. A small model is often good enough for sorting, short summaries and quick drafts, and costs a fraction as much.
Image, video and voice
| Company | Media models (as of October 2026) | Notes |
|---|---|---|
| Nano Banana family for images (Nano Banana 2.1, Nano Banana Pro and others); Veo 3.1 for video (listed as preview) | Google says every Gemini-generated image carries a SynthID watermark | |
| OpenAI | GPT-Image-2.5 (Sunburst and Flare versions) for images | Image generation built into ChatGPT |
| Black Forest Labs (Germany) | FLUX 3 (image and video), FLUX.2 | Founded by researchers behind Stable Diffusion; some FLUX models have open weights |
| xAI | Grok Imagine Image 2.0, Grok Imagine Video 1.5 | Built into Grok |
| ElevenLabs (US/UK) | Eleven v4 and v4 Turbo (90+ languages), Eleven v3 (70+), older Multilingual v2 (29) and Flash v2.5 (32, low latency); voice cloning | Powers Dialogue Editor on this site |
| Meta | Muse Image, Muse Voice Transcribe, SAM 3.1 (image segmentation) | Listed on Meta's developer site |
How to read a model name
Names are marketing and they're inconsistent. Some patterns help.
- Bigger number, newer. Gemini 3.8 is newer than Gemini 3.1. Not always better at everything: a newer small model can trail an older big one.
- Size words. Mini, Flash, Lite, Haiku, Small and Ministral mean fast and cheap. Pro, Max, Opus, Large and Fable mean big.
- Preview, beta, experimental. It may change or vanish. Don't build a workflow on it.
- A number with a B (20B, 120B) is a parameter count, mostly used for open models. Closed companies usually don't disclose sizes.
- Dates in names (DeepSeek-V4-Pro-0813) mark a specific snapshot, here August 13.
Reading a launch announcement
Example: decoding a press release
An invented company, Lodestar AI, announces: "Lodestar 4 Ultra is the world's most capable model, beating all rivals on 9 of 12 leading benchmarks, with a 2 million token context window. Available today."
- "9 of 12 benchmarks" chosen by whom? Which three did it lose, and by how much?
- Were the scores run by the company or by independent testers? With the same settings for every model?
- Is "available today" for everyone, or a waitlist, or paid tiers only?
- A 2 million token window: how well does it use information in the middle of that window? Does it cost more for long inputs?
- Weights open or closed? Where is the data processed?
- Is there a model card or system card describing safety testing and known limits?
None of those questions is technical, and each answer can change what the announcement means.
Common mix-ups
- Claim: The chatbot app and the model are the same thing. What the evidence shows: ChatGPT, Claude and Gemini are apps that can switch between several models, add search, memory and tools, and change without notice.
- Claim: Chinese models are far behind. What the evidence shows: DeepSeek and Qwen models have been close to US frontier models on many public benchmarks since early 2025, and they are among the most used open-weight families.
- Claim: Benchmark leaders are best for every job. What the evidence shows: Rankings shift by task. Test two or three models on your actual work before choosing.
Reporting angle
When writing about a model launch, link the company's own model card or documentation, say who ran any benchmarks, and look for at least one independent evaluation before calling a model "best." Note the date in the story: "as of October 2026." For local stories, ask which model a vendor's product actually runs on; many "AI-powered" tools are a thin layer over one of the families above, and the underlying model's known limits carry over.
Research angle
- Claude models overview: model names, context windows and prices.
- OpenAI API models: the vendor's current model list.
- Gemini API models: Google's current model list.
- xAI models: Grok model names and pricing.
- DeepSeek pricing: model names and per-token prices.
- DeepSeek on Hugging Face: the open weights and model cards.
- Mistral models overview: open and commercial model lists.
Module glossary
| Term | Plain meaning | Lesson |
|---|---|---|
| Agent | An AI system that uses tools and takes several steps on its own toward a goal | 2, 8 |
| AGI | Hypothetical AI as capable as people across most work; no agreed definition | 8 |
| API | A way for software to send requests to a model, usually paid per token | 9 |
| Attention | How a transformer lets each token weigh every other token's relevance | 5 |
| Backpropagation | Working backward from the error to find how to nudge each weight | 4 |
| Benchmark | A standard test for scoring and comparing models | 9 |
| Context window | How many tokens a model can work with at once | 5 |
| Deep learning | Machine learning with many-layered neural networks | 1, 4 |
| Diffusion | Generating images or video by removing noise step by step | 7 |
| Expert system | A program built from hand-written if-then rules | 1 |
| Feature | One input a model uses | 3 |
| Fine-tuning | Extra training on a smaller, focused data set after pretraining | 5 |
| Generative AI | AI that produces new text, images, audio, video or code | 2 |
| Gradient descent | Training by small steps downhill on the loss | 4 |
| Grounding / RAG | Giving the model documents or search results to answer from | 6 |
| Hallucination | A confident false or invented statement from a model | 6 |
| Label | The right answer attached to a training example | 3 |
| Large language model (LLM) | A transformer trained on huge amounts of text to predict the next token | 5 |
| Latent space | A compressed numerical version of an image or sound | 7 |
| Loss | A number measuring how wrong a model is | 3 |
| Machine learning | Software that learns patterns from examples instead of following written rules | 1, 3 |
| Multimodal | Handles more than one kind of input or output | 9 |
| Neural network | Layers of simple weighted calculations trained together | 4 |
| Open weights | Downloadable trained parameters | 8 |
| Overfitting | Memorizing training data, noise included, and failing on new data | 3 |
| Parameter | A learned number in a model (weights and biases) | 4 |
| Pretraining | The first, largest training stage: next-token prediction on huge text | 5 |
| Prompt injection | Hidden instructions in content that hijack an AI | 8 |
| Quantization | Storing model numbers with fewer bits to save memory | 8 |
| Reasoning model | A model trained to work through steps before answering | 6 |
| Reinforcement learning | Learning by trial, reward and penalty | 2 |
| RLHF | Tuning a model toward answers human raters prefer | 5 |
| Temperature | Setting for how adventurous the next-token pick is | 5 |
| Test set | Held-back data used only to measure a model honestly | 3 |
| Text encoder | Turns a prompt into numbers that steer a generator | 7 |
| Token | The chunk of text a model reads and writes | 5 |
| Tool use | A model calling outside programs and using their results | 6 |
| Training cutoff | The date after which a model has no training data | 9 |
| Transformer | The 2017 attention-based design behind modern language models | 1, 5 |
| Voice cloning | Making a speech model talk in a specific person's voice | 7 |
| Weight | An adjustable number setting how much an input counts | 4 |
Pictures, Numbers and Infographics
ChatGPT went from its November 2022 launch to use by about a third of American adults in under three years. The figures below track who uses these tools, who builds the models behind them, and how fast the cost of running them fell.
By the numbers
- 34%: share of U.S. adults who say they have ever used ChatGPT, roughly double the share in summer 2023. Pew Research Center, 2025
- 58%: share of U.S. adults under 30 who have used ChatGPT, up from 43% in 2024 and 33% in 2023. Among those 65 and older it was 10%. Pew Research Center, 2025
- 53%: population adoption generative AI reached within three years, faster than the personal computer or the internet. Stanford AI Index, 2026
- Over 90%: share of notable frontier models in 2025 that were produced by industry, meaning companies. Stanford AI Index, 2026
- 40, 15 and 3: notable AI models produced in 2024 by institutions based in the United States, China and Europe. Stanford AI Index, 2025
- Over 280-fold: drop in the cost of running a system that performs at the level of GPT-3.5, from November 2022 to October 2024. Stanford AI Index, 2025
- 50.1%: how often the top model read an analog clock correctly, in a year when one model, Gemini Deep Think, earned a gold medal at the International Mathematical Olympiad. Lesson 6 explains why ability is this uneven. Stanford AI Index, 2026
Check yourself
1. As of October 2026, which of these companies releases open-weight models?
DeepSeek posts its model weights on Hugging Face. Anthropic does not release weights. Google releases open Gemma models while keeping Gemini closed.
2. Google's Gemini app and Gemma are related how?
Gemini is available only through Google's apps and API. Gemma models can be downloaded and run locally.
3. A company says its new model is "the best in the world" and shows a chart of benchmark scores it ran itself. What is the right reaction?
Self-reported benchmarks are a starting point. Companies choose which tests to show. Independent evaluations, and testing on your own tasks, tell you more.
4. Meta's Muse Spark, launched April 2026, differs from Meta's earlier Llama 4 models how?
Muse Spark was Meta's first proprietary frontier model. Llama 4 Scout and Maverick (April 2025) were downloadable.
5. Why does this lesson say to check the date on any list of top AI models?
Anthropic alone released at least five new models between June and October 2026. A list from six months ago is history.
Terms in this lesson
- Frontier model
- One of the most capable models available at a given time, usually from a handful of well-funded labs.
- Model family
- A line of related models from one company, usually in sizes from small and cheap to large and capable.
- Benchmark
- A standard test set used to score and compare models, such as a set of math problems or coding tasks.
- API
- Application programming interface: a way for software to send requests to a model and get answers back, usually paid per token.
- Multimodal
- A model that handles more than one kind of input or output, such as text plus images or audio.
- Training cutoff
- The date after which a model has no training data. It knows nothing later unless given search or documents.
Sources
- Anthropic: Models overview checked 2026-10
- Anthropic: Claude Fable checked 2026-10
- Anthropic: Newsroom checked 2026-10
- OpenAI API docs: Models checked 2026-10
- AIToolly: OpenAI unveils GPT-6 Astra (September 4, 2026) checked 2026-10
- OpenAI: Introducing gpt-oss (August 5, 2025) checked 2026-10
- Google AI for Developers: Gemini models checked 2026-10
- Google AI for Developers: Image generation (Nano Banana models) checked 2026-10
- Google AI for Developers: Gemma docs checked 2026-10
- InfoWorld: Google plans Gemini 4 release before year-end (September 25, 2026) checked 2026-10
- Hugging Face: Qwen/Qwen3.8-2.4T-A95B model card checked 2026-10
- Meta AI: The Llama 4 herd (April 5, 2025) checked 2026-10
- Meta developer site (Muse models) checked 2026-10
- NYU Shanghai RITS: Muse Spark launches as open-weight plans continue checked 2026-10
- xAI docs: Models checked 2026-10
- DeepSeek API docs: Models and pricing checked 2026-10
- DeepSeek on Hugging Face checked 2026-10
- EqualOcean: Alibaba previews 2.4-trillion-parameter Qwen3.8 (July 23, 2026) checked 2026-10
- Mistral AI docs: Models overview checked 2026-10
- Black Forest Labs: Models checked 2026-10
- ElevenLabs docs: Models checked 2026-10