Claude Opus 5.5
Anthropic
First model of the Claude 5.5 family. Anthropic says it matches Fable 5.1 on most work at lower cost than Opus 5, and positions it for complex work that needs careful judgment.
The models worth knowing, from the milestones that built the field to today's frontier. As of 2026-10-20. Rows marked vendor-confirmed were checked against the lab's own announcement or model page on the date above. Rows marked tracker came only from third-party release trackers. Names, prices and availability move fast, so confirm on the vendor page before relying on one.
The families to know in October 2026. Each lab ships a large, a mid-size and a small model.
Anthropic
First model of the Claude 5.5 family. Anthropic says it matches Fable 5.1 on most work at lower cost than Opus 5, and positions it for complex work that needs careful judgment.
Anthropic
The faster, lower-cost partner to Opus 5.5. Anthropic positions it for well-scoped everyday tasks, bug fixes and documents.
Anthropic
Anthropic's cheapest and fastest small model, for high-volume work like classification and summaries, and as a subagent next to the larger models.
Anthropic
A higher tier in the Claude 5 family, generally available on the API and the major clouds. Claude Mythos 5.1 is the same model with looser safeguards, limited to trusted access programs for security and life sciences work. Anthropic says Opus 5.5 performs at its level on most work.
OpenAI
Astra (Sep 3) is OpenAI's flagship. Sol and Luna (Sep 22) are cheaper, faster tiers, and GPT-6.1 Sol (Sep 29) is described as near-Astra intelligence at about a fifth of the price.
Google's frontier model. Announced Sep 30 to trusted testers, then generally available on Oct 13 through the API, the Gemini app and Search. Gemini 3.8 Flash is the cheaper fast option.
xAI
xAI's current model, which it calls its most powerful for coding and knowledge work. Available through the xAI API, Cursor and Grok Build. The announcement does not say whether the weights are open.
Meta
Meta's most capable model so far, built on Muse Spark and aimed at agent work. Muse Wonder is in the Meta Model API and Muse Code, and Muse Wonder Full is for Muse Code subscribers. Meta's announcement does not say whether the weights are open. Muse Spark Open is in the next group.
Models whose weights you can download and run yourself. Licenses differ, so read them.
Meta
Meta's open-weight model: 816B parameters in total, 42B active per token, 1M-token context, and text, image, audio and video input. Released under the Muse Model License, which has limited commercial terms.
DeepSeek
A 552B-parameter multimodal mixture of experts with a 1M-token context, MIT-licensed. A minor update, V4.1-Flash-1017, followed on Oct 17.
Alibaba
Open weights for a 2.4T-parameter Max-class model (95B active, Aug 12), a 27B model (Aug 14) and the efficient Flash-Next preview (Aug 26). The first Max-scale open release from Qwen.
Mistral
A 1.05T-parameter multimodal mixture of experts (52B active). Public preview on Oct 6. Mistral says the open weights arrive by the end of October, so check whether they are out yet.
Google's open models for local use, in four sizes (E2B, E4B, 26B mixture of experts, 31B dense) under the Apache 2.0 license.
Moonshot AI
A 2.8T-parameter mixture of experts (104B active per token) with a 1M-token context. Weights are published under Moonshot's own Kimi K3 License, so read its terms before commercial use.
Z.ai
Zhipu's current family, with a cheap Flash tier. Check the license.
Meta
The family that started the open-weight wave. Meta's newer releases are in the Muse line.
The models and papers that changed what people thought was possible. Read in order for the story.
Cornell (Rosenblatt)
The first trainable artificial neuron.
Toronto (Krizhevsky, Sutskever, Hinton)
Deep network on GPUs that won ImageNet and started the deep learning era.
The architecture behind every modern language model.
Bidirectional pretraining that reset search and language understanding.
OpenAI
175B parameters. Showed in-context learning and made prompting a skill.
DeepMind
Solved protein structure prediction at near-experimental accuracy.
Stability AI
Open image generation that ran on a consumer GPU.
OpenAI
Robust open speech recognition across languages.
OpenAI
Instruction-tuned chat at scale. Made LLMs a mass product.
OpenAI, Anthropic
Strong reasoning, image input in GPT-4, and the start of serious competition.
Meta
A commercially usable open model with a detailed training report.
OpenAI
First widely used reasoning model that thinks before answering.
DeepSeek
Open reasoning model trained with reinforcement learning on verifiable rewards.
Anthropic and others
Terminal agents that read, edit and test code made agentic coding mainstream.