Codex Ships Weekly, Gemini Previews, Claude Code Holds the Terminal

OpenAI, Google, and Anthropic all moved this week — but in opposite directions. One ships weekly, one previews a roadmap, one refuses to leave your terminal. Here's what the divergence means for your stack.

Three frontier labs made news this week, but the story for anyone actually deploying these tools isn't a single model — it's how differently OpenAI, Google, and Anthropic are now playing the distribution game. One is shipping weekly, one is previewing a roadmap, and one is deliberately staying in your terminal. If you're choosing a coding agent or a model backend right now, the cadence matters as much as the benchmark.

Google had the loudest week. Gemini 3.5 Flash landed, and per Simon Willison's notes it runs 3x the price of Gemini 3 Flash for external users (as of May 2026) — though Willison expects it is far cheaper for Google to run internally across its own products. Around it sat a cloud of I/O previews — Gemini Spark, the Antigravity agent harness, Omni — that have not fully shipped. Willison declined to cover them because he writes about products that have shipped, not "coming soon" ones, and that is a useful filter for buyers too.

OpenAI took the opposite tack: ship something every week and brand the cadence. Its latest Codex Thursday added goal mode — letting Codex work hands-off toward a goal "for hours or even days" — plus appshots that pull screen context into the app and the ability to drive a locked Mac from your phone. Separately, OpenAI introduced Guaranteed Capacity, letting customers lock in long-term compute access. That second item is not a feature — it is a signal that OpenAI is selling reliability and treating compute scarcity as a real procurement concern.

Anthropic stayed narrow. Claude Code's release notes add browser integration through a Chrome extension, so it can read console errors, network requests, and DOM state to debug — but the tool still lives in the terminal, not the IDE. Comparison roundups like tooldirectory.ai frame this as deliberate: Claude Code "intentionally doesn't compete" on IDE-and-completion, the lane where Cursor and Copilot fight.

For builders, the practical read is re-evaluation cadence. OpenAI's weekly drops mean the tool you benchmarked last month may have grown a goal mode you would actually use; Google's preview-heavy style means you should wait for general availability before rewiring a pipeline around Gemini Spark. For buyers, Guaranteed Capacity is the line item to watch — if your roadmap depends on scaling inference, long-term capacity is now something you negotiate, not assume.

The bottom line: Re-check your coding agent and model backend monthly for now — the labs are shipping and repositioning fast enough that last quarter's pick may already be the wrong one.

Sources