AI Product Updates: New Features and Releases

Qualcomm said its new flagship chip can run a 30B-parameter mixture-of-experts model locally.

LangChain's LangSmith now supports decision models including Jev and SemIf, letting developers see each step of a model's decisions to debug faster and improve their apps.

A developer announced a new release of their library so it can call the latest batch of models.

StepAudio 3 ASR ranks first on the AA-WER non-streaming speech-to-text index with a 1.7% error rate, down from 4.7% in the previous generation, and is available through the StepFun API.

The Information has launched the AI Inflection Index for Pro subscribers, quantifying themes and trends of the AI era based on its original reporting.

Arrow 2 lowers API token rates for high-throughput vector workflows, with usage-based billing determined by tokens consumed in generation and refinement.

Telos offers a 1M token context window, leaving more room for larger inputs and complex vector tasks from industrial sketches to dense illustrations.

Arrow 2 is positioned for speed, rapid iteration and small to medium-complexity tasks, while Telos targets more demanding briefs; both support SVG generation, vectorization, editing and animation.

Arrow 2 Telos adds an advanced refinement layer on top of specialized vector generation, and QuiverAI explains its capabilities, how it differs from Arrow 2 and when to choose it.

Arrow generates editable SVGs starting from code, turning shapes, paths, lines, color and composition into modifiable elements, and Telos extends this foundation to more demanding designs.

The Windows team is building AI into the OS by adding agent identities akin to user accounts, which the author calls the first opinionated approach to AI agents at the OS level.

Augment's Cosmos platform now supports Claude Opus 5.5, which the company says completes agent tasks at 40% of Opus 5's cost.