Google routes Claude Opus 5 to its own engineers through Antigravity under per-person quotas

The gate is per person, and it sits in the development platform. Claude Opus 5 is reachable from inside Antigravity, and each engineer has an individual usage quota ^[4]. Google's own statement puts the enforcement in its tooling. "Engineers have access to select third-party models in Antigravity, which is aligned with our external Antigravity enterprise offering," a Google spokesperson said, according to Business Insider ^[5]. The default stays where it was: "Gemini remains our primary and foundational model for internal development, with third-party models available on a quota to support specialized use cases," the spokesperson said ^[6].

The permission model moved a layer down. Before, the restriction named the tool, and employees generally could not use external coding tools such as Anthropic's Claude Code or OpenAI's Codex, with access granted to some Google DeepMind teams and certain high-priority engineering projects ^[7]. Now the restriction names a budget, and the tool is Google's own ^[8]. TechRepublic did not report the size of that budget or which use cases count as specialized ^[13].

A quota also shapes what the resulting usage data can support. Engineers spending a scarce allowance will spend it on the tasks that already went badly, so the sample is selected on failure. It can show where Gemini breaks on hard programming work, but a win rate on ordinary work is beyond it, because ordinary work never reaches the alternative. TechRepublic reports that some Google engineers have reportedly complained about Gemini's performance on difficult programming tasks ^[9], and frames the internal access as a way for engineers to compare models on the work they actually need to complete, and as a practical source of feedback for improving Gemini ^[10].

If you want to read Google's decision as a verdict on Opus 5, what has to hold is that your hard tasks resemble a Google engineer's hard tasks, in comparable repositories, with comparable review tooling. That quota-selected sample is drawn from Google's worst cases.

As policy, this still standardizes something. It standardizes the harness. Antigravity holds the routing, the approved model list and the quota, and the model name sits behind it as a selectable value ^[4]. A shop that standardizes on a provider has to migrate when the ranking changes. A shop that standardizes on a router edits a config value and then has an argument about accounting.

Google is also an Anthropic investor. It announced plans earlier this year to invest as much as $40 billion in the Claude maker, TechRepublic reported ^[11]. Amazon has taken a similar line on tooling, allowing employees to use outside AI coding tools while continuing to develop its own models and AI services ^[12].