Coding agents: Codex goes cloud and sells speed, Claude Code refines its subagents

Coding agents are moving fast on both sides. Here's the gist of the latest announcements, from trade press (we couldn't open OpenAI's official recap).

Codex: cloud environments and Ultrafast

At OpenAI's DevDay (late September):

• Reusable, shareable cloud environments: work continues when the laptop is off, and can be steered from another device.
• Redesigned CLI: voice control, an /agents view to delegate and track several tasks, better Git worktree handling.
• Automatic security scans of GitHub repos (Codex Security).
• Ultrafast: a premium tier up to 8x faster in Codex (about 300 tokens/s), up to 6x in the API, priced at six times the standard rate. It's opted into per call through a service-tier parameter, so batch jobs can stay on the standard tier. In Codex it comes with the Pro 500 plan ($500/month).
• GPT-6.1 Sol, aimed at coding, is listed at $2 / $10 per million tokens.

Claude Code: subagents and plugins

Per changelog aggregators (cross-check with the official changelog):

• version 2.1.292 (October 6): plugin install from a marketplace and an adjustable effort parameter for each subagent;
• following version: Haiku 5.5 becomes the default Haiku;
• since 2.1.277 (September): falls back to AGENTS.md when a folder has no CLAUDE.md, handy for multi-tool repos.

How to choose

• Speed: Ultrafast makes sense for an interactive loop where waiting costs more than tokens, not for background jobs.
• Async work: cloud environments suit long tasks; the cost is cloud cost, not your local machine.
• Real cost: compare cost per completed task, not price per token. Third-party leaderboards disagree on model names and scores; rerun a test on your own code.
• Security: an agent running in the cloud or on your machine should stay confined (see our MXC article).

Sources: The Decoder, Latent Space, AlphaSignal, Learnetto, Gradually (Claude Code changelog), Releasebot.