Benchmarking inference at scale: coding agents
On Monday, May 18, 2026, the AI landscape saw a flurry of activity, with a mixed bag of updates from industry leaders and innovators. From benchmarks and new product releases to partnerships and learning pathways, there's a lot to cover.
Together AI published a benchmarking study, showing that their coding agents outperform TensorRT-LLM by 31% in terms of transactions per second, and achieve 2× better time-to-first-task at saturation, all while reducing costs by 76% compared to Claude Opus 4.6.
Vercel AI launched its Flat Rate CDN in limited beta, and also introduced Consolidated Commit Status on GitHub, making it easier for developers to manage their codebases. Additionally, the company announced that firewall-mitigated traffic is now free on its platform.
Hugging Face introduced the Ettin Reranker Family, and also published a case study on fine-tuning NVIDIA Cosmos Predict 2.5 with LoRA/DoRA for robot video generation. The company also released PaddleOCR 3.5, which enables running OCR and document parsing tasks with a transformers backend.
NVIDIA CEO Jensen Huang spoke at Dell Technologies World, stating that demand for AI is "going parabolic, utterly parabolic". The company also announced the arrival of its first CPU built for agents, Vera, which has already been delivered to top AI labs, including Anthropic, OpenAI, and SpaceXAI. Vera is said to enable agentic AI inference at one-tenth the cost per token, and runs 50% faster than traditional CPUs.
On the infrastructure side, ClickHouse published a blog post comparing its database to Prometheus for high cardinality, and also shared a success story from ChatFeatured, which migrated from PlanetScale Postgres to Postgres managed by ClickHouse, reducing analytics query times from 2.5 minutes to under a second.
AWS AI/ML published several blog posts, including one on prompting Amazon Nova 2 for content moderation, and another on how Aderant transformed its cloud operations using Amazon Quick. The company also announced integrations with Atlassian Confluence Cloud and Amazon Quick.
Google DeepMind published a study on using Co-Scientist to find novel factors that successfully rejuvenate human cells, which could potentially lead to breakthroughs in genetic research.
In funding news, there were no major announcements, but several companies, including NVIDIA and AWS, continue to invest heavily in AI research and development.
A quieter note from Databricks: the company announced the launch of its analytics engineer learning pathway, which provides a structured approach to learning data engineering and analytics. Databricks also published a blog post on how NASDAQ built a governed intelligence layer using dbt and Databricks, and another on AI-assisted analytics engineering.
Meanwhile at Google, the company's DeepMind division continues to push the boundaries of AI research, with a recent study on using machine learning to analyze genetic data.
And one for the security folks: OpenAI and Dell partnered to bring Codex to hybrid and on-premise enterprise environments, enabling secure deployment of AI coding agents across data and workflows.
Apache Airflow launched its registry, a searchable catalog of every official Airflow provider and its modules, making it easier for developers to find and integrate the tools they need.
That's the Daily AI Brief for Monday, May Eighteenth, Two Thousand Twenty-Six. More tomorrow.
Sources covered
1 story · 25 sources (0 news, 25 company, 0 voices)