Google opens Claude Opus 5 to every engineer. Run two models.

The company that ships Gemini just told its own engineers they can code with Claude Opus 5. Google opened access to Anthropic's model for all engineers through Antigravity, its internal development platform, on a per-person quota. Access used to be limited to select DeepMind teams and priority projects. If the strongest possible incentive to run one model in-house is being the company that makes it, and that company just added a second — your single-vendor coding stack deserves a second look.

What actually happened

Business Insider reporter Hugh Langley broke the story on September 15, citing employees and a company spokesperson. The shape of it matters more than the headline:

• Engineers reach Opus 5 inside Antigravity, Google's own IDE — not through Claude Code
• Each employee gets an individual quota for the third-party model
• Google's line: "Gemini remains our primary and foundational model for internal development," with third-party tools "available within quota to support specialized use cases"

So this is not a defection. It is a governed second lane: Google kept the harness, kept the telemetry, kept the default, and metered the alternative. That is a procurement decision, not a benchmark result.

Why model portability matters for your business

The harness is the asset, not the model. Google did not hand engineers a rival vendor's tool. It put a rival vendor's model behind its own interface. That is exactly the position you want to be in — the IDE, the prompts, the context files, the review gates and the logs stay yours, and the model behind them is a config value. We have made this argument about BYOK coding agents and about pinning your model layer. This is the largest possible proof of it.

Quotas are the honest version of "we have two vendors." Most small teams say they are multi-model and mean they have two subscriptions nobody tracks. A per-person cap on the expensive model forces the real question: which tasks justify it? Long refactors across many files and gnarly debugging sessions usually do. Autocomplete and commit messages do not. Set the split deliberately instead of letting whoever is loudest set it.

Run the bake-off on your code, not on a leaderboard. Google has every reason to prefer Gemini and still concluded some workloads go elsewhere. Your answer will differ from theirs because your codebase is different. Take five real tickets from the last sprint, run them through both models in the same harness, and score cost per finished task — not tokens, not vibes. That is a two-day exercise that survives the next six model releases.

Plan for the reverse, too. A vendor can raise prices, retire a snapshot or gate a tier at any time. If switching models is a one-line change in your setup, none of that is an emergency. If it means rewriting prompts and re-teaching a team, it is.

Sources: Business Insider via Techmeme, Seoul Economic Daily.