Section
Agents, coding assistants, APIs and the products built on top of them, judged on whether they work rather than on the launch video.
Cloudflare, OpenAI, Perplexity, Liquid AI, AWS's Strands Labs and independent developers have each released competing decision models since Jev's September 15 launch.
Each task inside the sandbox runs in its own Firecracker microVM, with credentials and network access controlled from outside the machine, and Perplexity has released the code under the Apache 2.0 license on GitHub.
The company says the new model finishes an edit in under eight seconds, costs $0.06 per image, and changes only the specific region a user targets.
Figma unveiled Figma Specs, a markdown-based design-rules editor, moved code layers to public beta, and linked its Payload CMS to Figma Make at its first Config event in India.
The fine-tuned Gemma models screen conversations for scam patterns directly on a user's phone, so chat data is not sent to the cloud for processing.
The startup, founded by architect Al Dram and data scientist Coco Cao, will make its second-generation software public this week, generating blueprints for architecture, structural, mechanical, electrical and plumbing work from a single sketch.
A new harness tracks evidence and conclusions together as AI agents search document collections too big for one context window, more than tripling accuracy on a benchmark called OfficeQA Pro.
OpenAI framed its new Dots agent as safer than Meta's Muse at DevDay, continuing a pattern of AI companies competing on privacy promises that have proven hard to keep.
The opt-in OSS Scanner service runs unreviewed, model-generated security reports using Anthropic's strongest models, including Claude Mythos.
The probes read a model's internal activations while it works, instead of rereading its output, letting Baseten's AI hosting customers catch rogue agents for a fraction of the usual monitoring cost.
The tool checks whether images, video, or audio carry Google's SynthID watermark, after a year of limited testing with journalists and researchers.
The new endpoint returns probabilities, category choices, or rubric scores from text and images, and OpenAI says it works about ten times faster than its Responses API.
The Wikimedia Foundation says it found AI agents it believes belong to OpenAI editing wikis, probing a note-taking tool, and sending traffic that may have fueled a May outage.
The v0.2 preview release also adds file and code-change review tools and scheduled Automation Tasks, and now supports non-DeepSeek models via OpenAI-compatible endpoints.
Enterprises can now run IBM's agentic coding assistant Bob on premises or in air-gapped networks, pairing it with models like NVIDIA Nemotron, Poolside Laguna, Claude, Gemini or GPT.
The open-source code lets developers program ESP32 boards or Raspberry Pi devices to run Meta's Muse assistant on displays, buttons, and sensors they build themselves.
The redesigned training framework scales mixture-of-experts models toward trillion-parameter sizes, in one benchmark growing a model's expert pool from 8 to 128 while active parameters per token held near 3.2 billion.
OpenAI unveiled the consumer AI agent at its DevDay conference, restricting it to $100-a-month Pro subscribers while Meta's competing Muse agent remains free for all users.
An AI agent handling an item pickup for its owner sent a false all-clear message to the buyer, then apologized publicly after the deal fell apart and a negative rating followed.
Call for Me lets the AI assistant dial businesses directly from a user's own number to make reservations, check stock or reschedule appointments, starting with U.S. Gemini subscribers on Pixel 11.
A researcher used a ClickFix-style terminal trick to steal the access token behind Meta's macOS AI agent, and Amazon began blocking Muse from its site the same week.
The FAA plans to roll out SMART, an $875 million AI platform from Air Space Intelligence, starting in the Washington, D.C. area, the Wall Street Journal reports.
The rollout, starting this week for Google Home Premium Advanced subscribers in the US, lets agents such as Claude and ChatGPT operate devices and pull camera-event summaries through natural language.
The company introduced two separate toolchains, cuda-oxide and cutile-rs, that let engineers write GPU kernels in Rust instead of CUDA's C++ or Python interfaces, with plans to let all three languages interoperate.
OpenAI says its Habitat storage platform grew from an internal Python library into a globally distributed system handling 22 million requests per second.
The managed service, powered by what OpenAI calls the Codex harness, is built for orchestration, tool use and stateful sessions in agents that run for extended periods in the cloud.
The research preview drafts code with one model, escalates to a stronger one when a quality check fails, and bills for every model the workflow calls at standard rates.
The Rust and Metal runtime, built for Perplexity's Hybrid Compute feature, skips PyTorch and MLX entirely and posts up to 1.35x faster token generation than MLX-LM in Perplexity's own benchmarks on an M5 Max.
