The first week of September 2026 gave every major lab a new model story. It also made a quieter catalog problem obvious: Bookmarkit already listed Cursor, Devin, and GitHub Copilot, and other products already pointed at Claude Code and Antigravity as alternatives, but those two had no profiles. Teams were comparing ghosts.
We added first-class pages for Claude Code, Google Antigravity, OpenAI Codex, OpenCode, and Devin Desktop. The collection is AI coding agents to shortlist. This post is the sorting rule, not a ranking.
Split the job before you split the vendors
A coding agent is software that can read a repository, change more than one file, run a command, and come back with a diff or a pull request. That loop now ships as a terminal, an IDE, a desktop manager, or a cloud worker. Buying two surfaces is normal. Scoring them as if they were one SKU is how you get a second seat tax and no logs.
- Terminal-first: Claude Code, OpenCode, Aider. You already live in git.
- IDE plus agent manager: Cursor, Devin Desktop, Google Antigravity.
- Same agent inside a chat plan: OpenAI Codex with ChatGPT, or Copilot if GitHub is already the seat.
- Unattended cloud worker: the Devin cloud agent, not Devin Desktop.
- Prompt-to-app, no existing repo: Lovable. Do not put it on the same scorecard as Claude Code.
What the first week of September changed
Anthropic’s Fable 5.1 / Mythos 5.1 post says Fable 5.1 defaults to High effort in Claude Code. OpenAI’s Astra launch is also a Codex story: the system card talks about internal Codex tasks and Daybreak toggles inside Codex. Google’s Gemini 3.8 Flash announcement tells developers to build in Antigravity. If you last scored these tools in June, the default model behind the agent moved.
- Write the job: local edits, parallel agents, or an overnight cloud run.
- Pick one surface per job. Then pick the vendor that already owns the identity (Claude, ChatGPT, Gemini, GitHub, or Cognition).
- Run the same messy ticket on two agents. Score the log, the diff, and whether you could stop the run.
- Ask which MCP servers and git remotes the agent can reach. Cursor Origin makes that question louder.
- Save the shortlist in the coding agents collection so the next model week does not reset the argument.