2026-08-18 · TuesdayPublished daily at 08:00
AI Daily
- 1
- 2
- 3OpenRouter Image Generation API: A Code-First Integration TutorialOpenRouter:Announcements
01Product Releases/Updates
Cursor Launches Origin, a Code Hosting Service as an Alternative to GitHub
Original textOpenRouter launches an Activity dashboard and Analytics API to track AI usage costs by agent, model, and request
OpenRouter released an Activity dashboard and a beta Analytics API, allowing users to view metrics such as spending, token volume, and cache hit rate by agent, model, and request, with support for drilling down to individual request logs.
Original textOpenRouter Image Generation API: A Code-First Integration Tutorial
OpenRouter launched a dedicated image generation API that lets users call image models from multiple providers through a unified request format and a single key. Developers send requests to POST /api/v1/images, and data[0].b64_json in the response contains the base64-encoded image data, which can be decoded and saved as a local file. The tutorial demonstrates implementations in both Python and JavaScript, and supports passing reference images through input_references to generate variations.
Original textAgentCore Payments Middleware Gives LangChain Agents the Ability to Pay for APIs
The AgentCore Payments middleware allows LangChain agents to pay API fees using deterministic session budgets. The middleware handles x402 payment signing, and LangSmith can track every payment record.
Original textClaude Code v2.1.234 released: Adds a project directory name variable and GitLab MR badges, with multiple security and stability fixes
Claude Code v2.1.234 adds the optional CLAUDE_CODE_PROJECT_DIR_NAME environment variable, a selection:clear key binding, and GitLab MR badges, and automatically resumes sessions after usage limits reset.
Original text02Industry News
NVIDIA and SB Energy secure power capacity at Ohio’s PORTS-Pike campus, with OpenAI set to become a tenant
NVIDIA announced a partnership with SB Energy to secure power capacity (LPS) at Ohio’s PORTS-Pike technology campus exclusively for deploying NVIDIA compute, with OpenAI set to become a tenant.
Original textChina’s A-share market welcomes its “first humanoid robotics stock” as Unitree announces an August 19 listing on the STAR Market
Unitree announced that its shares will begin trading on the STAR Market on August 19, 2026, at an offering price of RMB 150.80 per share, implying a market capitalization of approximately RMB 60.993 billion and expected proceeds of approximately RMB 6.099 billion. The company recorded revenue of RMB 159 million, RMB 393 million, and RMB 1.699 billion in 2023, 2024, and 2025, respectively, with net profit of RMB -11.1451 million, RMB 95.4747 million, and RMB 278 million. It is one of the few profitable high-performance general-purpose robotics companies in the world.
Original text404 Media traces the fate of rare books: Amazon buys books in bulk, scans them for AI training, then destroys them
Original textOpenAI funds 14 independent projects to advance research on economic opportunity and resilience in the age of intelligence
OpenAI announced a total of $1 million in funding and up to $1 million in API credits for 14 projects led by independent organizations to advance research on economic opportunity and societal resilience as AI progresses.
Original text03Paper research
PhotoScan: Estimating insulin resistance from smartphone photos with accuracy approaching DXA
Google Research introduced PhotoScan, a deep learning framework that estimates three-dimensional body composition directly from 2D smartphone photos and can predict insulin resistance, with accuracy approaching DXA scans in clinical studies.
Original text04Tips and perspectives
How OpenAI uses frontier intelligence to strengthen its own defenses: The Defender’s Window
Following the OpenAI-Hugging Face incident, OpenAI acknowledged that it had underestimated its models’ actual cyberattack capabilities. It is strengthening its security through four pillars: using Codex to validate code vulnerabilities, prioritizing agents for security alert triage, continuously enumerating attack paths, and making cyber capabilities available only to trusted defenders. The article demonstrates how ChatGPT Work, powered by GPT-5.6 Sol, identified 13 issues on a personal website in 15 minutes and completed the fixes within an hour.
Original textHow to disable or avoid intrusive AI: A practical guide covering Windows, Chrome, Edge, Firefox, and major apps
Original textBuilding zero-trust AI agents with Google’s Agent Development Kit
Google has open-sourced an example of a zero-trust customer service and returns agent built with ADK and Gemini, demonstrating defenses against prompt injection and other attacks. The architecture provides three layers of enforced security outside the LLM context: hardware-backed cryptographic signatures ensure non-repudiation of database writes, a gVisor sandbox isolates dynamic code execution, and a deterministic semantic gateway validates business logic. System prompts provide only soft constraints and cannot serve as a security boundary.
Original textThe future of the open-source model ecosystem: Nvidia bets on “teaching everyone to produce tokens”
The open-source model ecosystem is increasingly dependent on Nvidia’s funding. The company has invested $26 billion to advance the development of nearly open-source models and expand demand for inference chips. If this approach fails, open-source models will shift toward long-tail use cases centered on efficiency and modifiability, diverging from closed-source models. Meanwhile, as the barriers to training foundation models rise, the open-source community’s interest is shifting from full-scale training toward fine-tuning models such as DeepSeek V4 Flash and GLM 5.X.
Original textA 33-percentage-point utilization gain on the same cluster: The difference is allocation order
Hugging Face built a constraint-aware GPU allocator and compared it with a FIFO scheduler across seven benchmark scenarios. With the same hardware and workloads, GPU utilization increased by up to 33 percentage points, while priority-weighted output rose in every scenario, by as much as 105%. The allocator models real-time inference demand as a curve rather than a peak and ranks batch jobs by priority across the entire scheduling window, reclaiming reserved but idle capacity.
Original textHow ABC Legal uses Claude Managed Agents to turn every employee into a builder
After deploying Claude Enterprise for 1,100 employees, ABC Legal used Claude Managed Agents to turn scattered experiments into a governed agent ecosystem. By July 2026, it had launched more than 50 production-grade agents, reducing the cost of some manual tasks they cover by approximately 50%, with around 310 employees using Claude daily.
Original textWhen Models Keep Learning: How Test-Time Training Changes AI Memory and Costs
Test-time training lets models continuously update their weights during use instead of freezing them after training. Compared with standard Transformers, its memory requirements remain constant rather than growing linearly with context length. Stanford research shows inference speed can improve by up to 2.7 times, while In-Place TTT can extend a 4B model's capabilities to 128k contexts without retraining. The trade-off is that each user needs a separate model copy, shifting serving costs toward compute and chips and making it better suited to personalized applications such as coding assistants.
Original textSGLang refactors CUDA Graph support, with Breakable CUDA Graph becoming the default prefill solution
SGLang refactored its CUDA Graph support by splitting it through runner/backend interfaces, allowing different capture strategies to be reused. The community-pioneered Breakable CUDA Graph (BCG) is now the default for prefill. Its codebase is only about one-quarter the size of the torch.compile solution (521 lines versus 1,771 lines), and it builds 3.8–5.2 times faster.
Original text