SemiAnalysis finds top AI agents complete just 12% of hard professional tasks

SemiAnalysis finds the top AI agent completes just 12% of difficult professional tasks

SemiAnalysis launched GDPx, an evaluation of 206 tasks across 11 specialist domains and 14 data partners. Claude Opus 5.5 leads at 12.0%, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. The tasks span fields including chip design, cybersecurity, biology and legal work.

Why it matters. The results challenge sweeping claims of agent autonomy in high-stakes professions. Crucially, 12% is a score on a deliberately difficult, domain-weighted benchmark, not a measure of success across everyday office work. Tasks mix binary and continuous grading, and scores average three attempts. These are the benchmark creators' results, not an independent replication.

Meituan launches LongCat-2.5 for multimodal, long-running coding agents

Meituan introduced a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion activated per token. It handles images and text, with training aimed at extended coding, design and agent workflows. Its chat and API are available, while public model weights were described as forthcoming.

Why it matters. A new Chinese model expands the competitive field for coding agents, but total parameter count is not a direct proxy for inference cost or capability. Meituan's performance comparisons are vendor-reported; public weights and independent evaluations will be the more useful checkpoints.

OpenAI, ADNOC and G42 agree to deploy AI agents across oil operations

ADNOC announced a multi-year partnership to connect planning and operational decisions across production, refining, logistics and trading using OpenAI models and UAE-based AI infrastructure. The initial use case is field master planning across wells, reservoirs and facilities.

Why it matters. This is a consequential enterprise deployment of frontier models into industrial decision-making. ADNOC says planning work could shrink from months to hours, but that is a partner projection rather than an independently validated outcome. Governance, data access and measured performance will determine real impact.

CoreWeave expands CrowdStrike protection for production AI workloads

CoreWeave and CrowdStrike expanded their existing relationship to bring Falcon cloud protection, threat hunting and managed response to AI workloads on CoreWeave's infrastructure. CoreWeave also supports training and inference for CrowdStrike's SafeMind security models.

Why it matters. As model training and agent systems move into production, GPU clusters, containers and model-related intellectual property become security targets. The announcement describes an expanded commercial offering; it does not establish that customer incidents have fallen or security outcomes have improved.

Biological Computing Co. launches video model informed by neuron research

The Biological Computing Co. introduced mischief 1, an image-to-video model available via fal and OpenRouter. The company says insights from experiments with living neurons helped make the model five times faster than its base model and lower its price.

Why it matters. The combination of biological research and commercial generative video is unusual, but the claimed speedup is relative to the company's chosen baseline. Independent speed, quality and cost comparisons are still needed before treating it as a general breakthrough.

Watch for independently reproduced GDPx results, public LongCat-2.5 weights and evaluations, and measurable outcomes from industrial AI-agent deployments.