AI Frontier Daily · September 11, 2026
AI Frontier Daily — 2026-09-11
Headlines
1. OpenAI launches GPT-5.5, its first agent-native frontier model with delegated task execution
In a San Francisco livestream, OpenAI demonstrated GPT-5.5 coordinating sub-agents across email, code, and analytics to complete a four-hour software migration without human intervention. The model scored 84.3% on SWE-bench Verified and 92.1% on GAIA, with API access opening September 15.
Source: The Verge
2. EU AI Office issues first high-risk model certifications under the AI Act
On September 11, the EU certified Anthropic’s Claude Opus 4.5 and Google DeepMind’s Gemini 3.1 Ultra for regulated health-care and automotive deployments, setting the first legal precedent for the bloc’s safety-case regime. The certification requires ongoing incident reporting and quarterly capability audits.
Source: Reuters
3. Runway and Luma AI complete merger to form Lyra AI, raising $1.8B for real-time world models
The combined company unveiled a 1080p video generation model with sub-200ms latency, targeting film pre-production and robotics simulation. The round was led by Bessemer Venture Partners and included Nvidia and Adobe.
Source: Bloomberg
Model Releases
4. Mistral releases Mistral Large 3: a 1.2T-parameter MoE with 1M-token context
Mistral Large 3 uses only 42B active parameters per token, runs natively on consumer-grade quantized hardware, and includes upgraded function-calling and multilingual reasoning. It immediately topped the open-weight leaderboard on LMSYS, behind only GPT-5.5.
Source: Mistral AI Blog
5. xAI opens Grok 4.1 to enterprise API and open-weights the base model
Grok 4.1 features real-time X search grounding, native image generation, and a new “hyperon” reasoning mode that achieves 82.7% on GPQA Diamond. The open-weights release has already racked up 15,000 downloads in the first 12 hours.
Source: xAI
Industry
6. Microsoft and OpenAI finalize “Stargate II” $100B compute agreement
The deal commits 5 gigawatts of dedicated data-center capacity in Texas and Wisconsin by 2028, with OpenAI gaining early termination rights if hardware delivery milestones slip. Microsoft will retain exclusive cloud distribution for GPT-5.5 until 2030.
Source: The Wall Street Journal
Research Highlights
7. DeepMind shows world-model replay buffers boost long-horizon agent success rates from 41% to 74%
The paper, posted to arXiv on September 10, introduces a memory-augmented planning layer inside a transformer policy, reducing token consumption by 38% while improving success on 100-step embodied and API tasks.
Source: arXiv
8. Stanford HAI and Anthropic find process supervision beats outcome supervision for autonomous coding agents
A controlled study of 40,000 code-repair trajectories shows that step-level human feedback prevents silent strategy drift, improving final test-pass rates by 19% compared with outcome-only reinforcement learning.
Source: Nature Machine Intelligence
Compiled from multiple media reports.







