Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Coding playbooks · Reference table

Model specs: context, tools, and cutoffs in one table

Written November 6, 2026

Context windows, tool support, modalities, and knowledge cutoffs for the models coding agents actually use — each cell linked to its source.

Every cell below traces to the vendor's own model page or model card. A dash means the vendor page does not state that value. It is a gap in the vendor's page, not a number filled in from a second-hand summary. Specs move at release time, so open the linked source before planning a workload around a row.

How to read this table

A context window is the total budget the model works inside, not the amount you can usefully fill. Input, thinking, and output can share one window: Gemini 4 Argon counts all three against the same 1,048,576 tokens, while some vendors split the window between input and output instead. A window is also not the same as usable context: the advertised number is the ceiling, and hosts and harnesses can cap it far below. The Ripple-65B card advertises 1,048,576 tokens, but the hosted endpoint enforces a 16,384-token limit, per our Ripple-65B page. Check the live route, not just the card, before planning long-context work. The table shows specs, not a ranking. A bigger window says nothing about whether the model does the job well.

The table

ModelContext windowMax outputToolsModalitiesKnowledge cutoffSource
GPT-6.1 Sol1.05M tokens128K tokensFunctions, web search, file search, computer useText, image in; text outApr 30, 2026OpenAI API models
GPT-6.1 Astra1.05M tokens128K tokensFunctions, web search, file search, computer useText, image in; text outApr 30, 2026OpenAI API models
Gemini 4 Argon1,048,576 tokens (unified input + output)65,536 tokens by default; up to 1,048,576 with long decodeFunction calling, code execution, file search, computer useText, image, video, audio, PDF in; text outMar 2026Gemini 4 Argon model page
Kimi K3.1 Pro1,048,576 tokens—Tool calls, tool choiceText, image—Kimi K3.1 Pro model card
Mistral Large 41M tokens—Native function calling, JSON outputText, image in; text out—Mistral Large 4 model card
LongCat-2.5—256K tokensFunction callingText, image—LongCat API FAQ
Ripple-65B1,048,576 tokens—Yes; MCP Atlas score on the model cardText—Ripple-65B Preview model card
Hy41M tokens—Tool-call parser in the serving recipeText—Hy4 preview model card
Claude Sonnet 5.51M tokens128K tokensAgentic tool useText, imageJun 2026Claude Sonnet 5.5 overview
Claude Opus 5.51M tokens128K tokensAgentic tool useText, imageJun 2026Claude Opus 5.5 overview
GPT-6 Cyber1,050,000 tokens128,000 tokensFunction calling; web search, file search, code interpreter, MCP, computer useText, image in; text out—GPT-6-Cyber model page

What this table does not do

It does not rank models by context window alone, and it does not tell you what a model costs. For that, see Model prices: what the major models cost and, if you pay through a subscription instead of the API, Plan allowances: what each agent plan lets you do. It does not cover licences: open-weight rows here are described by their specs only, and the licence terms live in Open-weight model licences, in one list. A dash is a missing vendor statement, not a missing capability. Where a vendor does not publish a value, the honest entry is blank. And a context window is a ceiling on the vendor's page, not a measurement of how much of it your setup will actually let you use. Some agent setups assume a smaller window for models they do not recognize and compact early, so confirm the working window in your own setup before designing a long session around a row.

Related tools

Sources

Back to: Coding playbooks hub · Also: Model prices: what the major models cost · Plan allowances: what each agent plan lets you do · Open-weight model licences, in one list