GenAI Field Guide

Getting started

The lineup on GenAI.mil

GenAI.mil launched in December 2025 as the DoD enterprise portal for approved generative AI. As of 31 August 2026 it hosts Gemini for Government, ChatGPT Mil, and Grok for Government, all accredited at IL5 for CUI.

≈4 minLandscape facts checked — re-verify before relying on them.
1 · Authorization + data What a new user feels most 2 · Harness The product around the model 3 · Vendor choices Guardrails and defaults 4 · Model design What a new user feelsleast
The four ways vendors differ, in the order a new user feels them.

Who is on the portal

Vendor Status Position, and what it is for
Gemini for Government (Google) Original offering; Gemini 3.1 Pro/Flash added April 2026; includes Agent Designer (users have built 100,000+ agents) DoD officials describe it as strong for search and multimodal, long-document work
ChatGPT Mil (OpenAI) Launched 31 Aug 2026; chat, files, projects, custom GPTs The text-heavy drafting workhorse: admin, logistics, planning, policy
Grok for Government (xAI / Starshield AI) Launched 31 Aug 2026; positioned as the third vendor to avoid lock-in Agentic multi-step workflows
Claude (Anthropic) Not on GenAI.mil. In February 2026 agencies were ordered off Anthropic and DoD designated the company a supply-chain risk; on 27–28 August 2026 a federal judge ruled that designation unlawful and ordered it removed. An appeal is possible. Use what is on GenAI.mil; do not use Claude for official work unless and until your command's guidance says otherwise. Its concepts (Claude Code, skills, subagents, MCP) still matter because every vendor has copied them.

The four axes, in the order they matter to you

Axis Gemini ChatGPT / OpenAI Grok / xAI
1. Authorization and data On GenAI.mil at IL5; FedRAMP High On GenAI.mil at IL5 On GenAI.mil at IL5
2. Harness (tools, files, projects, agents, memory) Gems and Projects, Agent Designer, Google Workspace hooks, 1M-token context Projects, custom GPTs, Codex for coding, built-in tools Deep Search, Build Mode, automations, very large contexts
3. Vendor decisions and guardrails Enterprise-minded; conservative on some topics; strongest document and video ingestion Broadest consumer base; strongest software tooling; tight controls Fewest content restrictions; thinner audit tooling (reported)
4. Model design (least visible to you) Native multimodal, fast Flash tier plus Pro GPT-5.x family with reasoning tiers Grok 4.x family, tuned for agentic work

Claude is not on the portal, so it has no column; its harness ideas (Claude Code, Skills, Subagents, MCP) show up in every other vendor's product.