Agent Atlas

Snapshot · read 11 November 2026 · stale after 11 December 2026

The Frontier, November 2026

What eight labs say about their leading models, on their own pages. Where a page does not state something, the cell says not published. Nothing here is estimated.

Eight labs, one day Factual

“Leading” is the lab's own word where it has one: Latest, flagship, strongest. Several labs do not call any model a flagship, and the note under each name says what the page does say. Models change monthly, so every cell is dated to the day it was read.

Lab and modelContext windowKnowledge cutoffWeightsParameters
Alibaba Qwenqwen4-max-previewThe text-generation page points to qwen4-max-preview when quality comes first, Token Plan only. The start-here model on the same page is qwen3.7-plus. The page does not use the word flagship. (Alibaba Cloud)1Mnot publishednot publishednot published
AnthropicClaude Fable 5.2The Fable page marks it Latest, released 9 Nov 2026, for demanding reasoning. The models overview says to start with Claude Opus 5.5 for most workloads. These pages do not use the word flagship. (Anthropic; Anthropic)1M tokensJun 2026not publishednot published
DeepSeekdeepseek-v4-proOf the two models on the current pricing page, this is the higher-priced one. The page does not use the word flagship. (DeepSeek; DeepSeek)1Mnot publishednot publishednot published
GoogleGemini 4 CarbonThe models index calls it the strongest reasoning model. The model page calls it Google's latest Gemini 4 model. On that index, flagship is used for speech, live audio, and music. (Google; Google)1,048,576Unified input and output token limit on the model page. The index card says 1M tokens.not publishednot publishednot published
MetaLlama 4 MaverickThe models page calls Maverick industry-leading. That page does not use the word flagship. Scout, on the same pages, is the class-leading model with the 10M context window. (Meta; Meta)1MAugust 2024Released as BF16 and FP8 under the Llama 4 Community License17B activated, 400B total
MistralMistral Large 4The model page calls it a state-of-the-art open-weight model. The version label beside the license is v26.10. The page does not use the word flagship for this model. (Mistral)1Mnot publishedopen-weight, Mistral Ecosystem Open Weight License52B active, 1.05T total, and a 1.6B vision encoder
OpenAIGPT-6 AstraThe models index calls GPT-6 Astra "our flagship model" in its start-here sentence. (OpenAI; OpenAI)1,050,000Apr 30, 2026not publishednot published
xAIGrok 4.7The models index calls Grok 4.7 our flagship model. The card badge says New. The docs title on these pages is SpaceXSI Docs. (xAI; xAI)500,000Model page prints 500,000. The index prints 500k tokens.May 2026not publishednot published

Reading the table

Context window is how many tokens the model can hold at once: instructions, conversation and files together. Plate IV shows what filling it costs. Knowledge cutoff is when its training text ends. Weights are the trained numbers; when they are released, anyone can run the model on their own hardware. Parameters count those numbers. Where a page gives active and total counts, as Meta's and Mistral's do, only the active ones work on each token.

Only Meta and Mistral publish weights and parameter counts for the models listed. The other six pages state neither for these models.

More in the Atlas: open weights, mixture of experts, and the chips and data centers these models run on.