The Fable model has generated a GPU kernel that achieves an 18.71X speedup on an RTX PRO 6000 Blackwell. This result tops the KernelBench-Mega leaderboard, beating previous records set by Claude Opus 4.8 (14.4X), GLM-5.2 (11.14X), and GPT 5.5 (4.34X). The new kernel executes exactly one launch per decoded token, whereas other top entries require between four and fourteen separate launches.
Autonomous kernel design
Developing efficient code is a core input task for AI research. If systems can autonomously design better kernels, they improve their own ability to build AI. This benchmark signals progress toward recursive self-improvement.
The official leaderboard is available here. Elliot Arledge, a benchmark maintainer, provided analysis on X here.
***
Researchers from the Center for AI Safety and Scale Labs report a rise in the success rate of AI systems automating online freelance projects. The success rate jumped from 2.5% in October 2025 to 16.1% in July 2026, based on the Remote Labor Index.
Remote Labor Index
The index tests whether AI agents can handle economically valuable projects end-to-end. Assessed areas include 3D design, architecture, graphic design, video, audio, data analysis, and web applications.
Recent evaluations covered three frontier models:
- GPT-5.5 achieved 6.3%
- Claude Opus 4.8 achieved 8.3%
- Fable 5 achieved 16.1%
The authors note the frontier has more than quadrupled in capability in under eight months. Specific tasks included recreating a ring design with swapped gem cuts, producing a 60-second animated advertisement for a tree service, and generating floor plans and renders from scanned cadastral data.
Economic impact
The question is what happens when this success rate reaches 80%. While new tasks will emerge, the speed of AI capability expansion suggests organisations may become extremely person-light. Current progress makes it difficult to reconcile rapid AI advancement with a static economy.
Humans will augment themselves and innovate, but the rate of human adaptation may not match the raw capability expansion of AI systems. Tracking metrics like the Remote Labor Index helps judge this gap.
Further reading is available on the Center for AI Safety blog here.
***
Researchers from the University of Hong Kong, UC San Diego, Columbia University, UC Santa Barbara, Mila, Snorkel AI, the University of Wisconsin, Alibaba Qwen, The Ohio State University, Simular, and NeoCognition have released OSWORLD 2.0. This benchmark evaluates how well AI systems perform multi-step tasks on computers.
Computer-use agents
The median task in OSWORLD 2.0 takes a human approximately 1.6 hours, which is about 48 times longer than the 2-minute median in OSWORLD 1.0. The suite contains 108 long-horizon tasks, including 31 self-hosted websites.
Each task is a self-contained workflow requiring a high-level goal, realistic artifacts, a stateful environment, and a scoreable final state. Nearly 70% of tasks are estimated to take a skilled user more than one hour.
OSWORLD 2.0 supports a vastly expanded set of software compared to its predecessor. New additions include Slack, LinkedIn, Shortcut, REAPER, MuseScore, WPS, GitLab, Overleaf, LabPlot, Zotero, and AWS. It also includes websites mimicking professional services like insurance claims and visa applications.
Current performance remains low. The strongest setting, Claude Opus 4.8 with maximum thinking and batched tool calls, reached only 20.6% binary accuracy and 54.8% partial-score accuracy. Agents struggle most when recovering hidden state, tracking items, or resolving conflicting information.
Performance should rise over time. In July 2025, the highest scoring models on OSWORLD 1.0 achieved roughly 30%, while recent models scored around 75% by June 2026. Similar growth is expected here.
Computer use is fundamental for AI to perform economically valuable tasks and conduct science research. These benchmarks act as a proxy for how well AI handles complex, varied workflows. The focus now shifts to tasks that require hours of human work.
The official website is here. The research paper is available on GitHub here.
***
JD.com, the Amazon of China, has published details on its Oxygen AI Item Center. This software manages inventory for a platform with 700 million users, millions of merchants, and a catalog of tens of billions of SKUs.
Deep learning meets structured systems
The Oxygen AI Item Center combines deep learning with structured systems to handle the complexity of country-scale e-commerce. It is fundamental to how the company tracks its massive inventory.
