Gemini 4 Carbon shows Europe will get frontier AI on Washington's timetable

Google has announced Gemini 4 Carbon, an upgrade to the Gemini 4 Argon model it had released two weeks earlier. Carbon is not yet available to the public. Google said the model's strength in agentic coding and cybersecurity calls for a cautious rollout. It is working within the US government's voluntary testing framework and plans to release Carbon through its Fairwinds cyber defence programme once that testing is done. No date for a wider release has been given.

Google said Carbon is twice as likely as Argon to write code free of vulnerabilities and that it resisted every prompt-injection attempt in the Grey Swan evaluation. In Google's own benchmarks, Carbon beats GPT-6 Astra and Claude Opus 5.5 in almost every test, though by a few points. API pricing will start at a discounted $2 per million input tokens and $10 per million output tokens, then double. Google has not said when.

A release gate set in one capital

The notable part is not the benchmark table. It is the order of events. A US company builds a model, a US government framework tests it, and only then does the wider world get access. The framework is voluntary, and Google chose to use it. Still, the effect is that the timing of a frontier release now runs through Washington.

If Carbon follows the same path as Argon, it could be available in a couple of weeks. That is short. But Google makes no promise that Carbon will follow the same path. Europe receives the result of a decision taken elsewhere.

What this means for European companies

Many European firms build products on top of models like Gemini through an API. If a new model can be announced, benchmarked and praised weeks before anyone can call it, product teams face a planning problem. They cannot test it, price it or compare it with what they use today. Open pricing adds to that. Carbon starts at a discount that will double on a date Google has declined to set.

The security angle matters too. Google says Carbon is strong at finding and fixing vulnerabilities, and it plans to route early access through its own cyber defence programme. Who gets into such programmes first will shape who can use these tools to defend their systems. European security teams, including in Sweden, should not assume they will be at the front of that line.

The strongest objection

One objection is that this is exactly how careful release should work. A model that is good at finding vulnerabilities can also help attackers, so testing before release is sensible, and a short delay is a fair price. We agree with the testing. Our concern is narrower. When the testing sits with one government and is not coordinated with others, the schedule for everyone else becomes a side effect. Safety and dependency can both be true at once.

What remains uncertain is how long the US testing will take for each model, whether other labs will follow the same route, and whether European bodies will ever be part of such pre-release checks.

Our view

We think European companies and regulators should stop treating access to frontier models as simultaneous and equal. It is staged, and the first gate is in the United States. Firms should build roadmaps that can absorb weeks of delay and unclear pricing, and avoid tying critical plans to a model that is announced but not shipped. Regulators should seek a working link to US testing, so that Europe learns what was found rather than only when the door opens.