On the ranking by AI analysis portal Artificial Analysis, Claude Opus 4.8 has held first place among all commercial language models since May 28, with an index score of roughly 61 points. The model comes from San Francisco. European systems do not appear in the top ranks. Yesterday afternoon, Germany's Federal Innovation Agency SPRIND closed the application deadline for a program meant to change that: €125 million for Europe's first internationally competitive frontier AI labs.
What SPRIND funds
The program is called Next Frontier AI and runs in three stages. In the first stage, up to ten teams can each receive up to €3 million, plus two years of computing capacity, infrastructure, and business development support. Up to six teams advance to the second stage and receive €8 million each. The three funding winners of the third stage each get €15.5 million. Those who win this round are expected to be able to raise a billion-euro financing round from private investors afterward.
Jano Costard, who oversees the program at SPRIND, explained the initiative to Euronews with a simple statement: "We have no time to waste." The jury, composed of international researchers, frontier lab veterans, and technology investors, will hold selection interviews on June 24 and 25, 2026. Final decisions are expected by end of June. The first ten teams are scheduled to begin work in July 2026.
Europe's gap and why it's widening
Since DeepSeek R1 appeared in January 2025, Europe has been under particular pressure to explain itself. The Chinese startup trained a powerful model at a fraction of the cost American companies had charged. Since then, OpenAI, Anthropic, and Google have shifted the benchmarks further. Anthropic's Claude Opus 4.8, released May 28, 2026, leads according to Artificial Analysis with roughly 61 points on the Intelligence Index, 1.2 points ahead of OpenAI's GPT-5.5. Google's Gemini 3.1 Pro achieves 77.1 percent on the ARC-AGI-2 benchmark, which tests reasoning.
Europe in this race is an observer. French AI startup Mistral has positioned itself as an efficient challenger, but in absolute model performance lies far behind American top models. A European institution that could be directly compared to OpenAI or Anthropic does not exist.
How far the gap becomes noticeable in everyday life is shown by a detail from the technology calendar: On June 8, Apple will present at its developer conference WWDC what is expected to be the integration of Google's Gemini in iOS 27. The operating system running on hundreds of millions of devices in Europe will deliver AI from one US corporation operated by another. European AI systems play no role in this infrastructure.
EU AI Act: Two months until obligation
Parallel to the competition for better models, the clock is ticking for all companies operating AI systems in the EU. Starting August 2, 2026, the EU AI Act applies fully. Then, among other things, the transparency obligations from Article 50 take effect: AI-generated content must be clearly marked as such, interactions with chatbots must be disclosed. Violations carry fines of up to €15 million or three percent of global annual revenue.
For European applicants at SPRIND, the regulatory framework is ambivalent. On the one hand, native EU compliance gives European providers a home advantage over non-European systems that must adapt. On the other hand, compliance infrastructure costs burden early teams in a growth phase. Which side of the equation dominates depends on how quickly teams can scale. The Code of Practice that the EU Commission is developing as implementing rules for the AI Act is expected in June 2026 and should clarify exact requirements.
By end of June, the jury decides
What SPRIND can achieve with €125 million is not obvious. The budget lies far below what American frontier labs invest per training run: industry estimates suggest large model training costs in the double-digit to low triple-digit million range depending on architecture and size. SPRIND argues that centrally provided computing capacity and the search for more efficient architectures change the equation. Whether that's true can only be answered once the first teams produce results. That's at earliest 2027.
The real test comes in 2028: Then the three finalists must convince private investors to invest roughly one billion euros each in them. Whether European venture capitalists are willing to invest this sum in AI labs competing against San Francisco and Beijing is the actually open question. The applications that came in to Berlin yesterday show, at least, that teams in Europe want to take on the challenge.