August 2026 Summaries

Mux Robots Directives is a new API and dashboard feature that lets developers combine multiple video-processing workflows into a single managed sequence, reducing the need to chain API calls, track job states, handle retries, or manage dependencies themselves. Directives can be assigned when a video is uploaded or applied later to existing video catalogs, making them useful for both new content and backfill operations. A highlighted use case is user-generated content platforms that need to generate premium captions and translate them into several languages, as Directives can wait for captions to become available before automatically launching translation jobs. By coordinating workflow order and fault-tolerant backend processing, the feature aims to simplify integration and allow developers to focus on their applications rather than video-processing infrastructure.

Mux Data’s Playing time by rendition metric shows which adaptive-bitrate video renditions viewers actually receive during playback, helping teams evaluate quality of experience and encoding efficiency beyond the renditions merely available in an ABR ladder. Because players determine and report delivered renditions based on device, bandwidth, congestion, and other conditions, the metric uses player-collected details such as bitrate, resolution, frame rate, codec, and rendition changes. It measures wall-clock viewing time per rendition while excluding time spent paused, buffering, or seeking, and presents proportional consumption over time through stacked charts of the most-used renditions. Users can identify underused ladder rungs, assess the effects of encoding or resolution-limit changes, monitor adoption of higher-resolution content, and filter results by factors such as title, country, or device category. The dashboard also supports comparisons across categories like countries or operating systems and allows rendition grouping by selected attributes, such as resolution alone, and is available in the Mux Data dashboard, with API access planned for the future.

Mux updated its basic video encoding ladder from 270p, 480p, and 1080p to 270p, 480p, 720p, and 1080p after analyzing viewer experience, infrastructure costs, cache behavior, and the needs of both human viewers and AI systems. Its per-title encoding model continues to optimize settings for individual video segments while maintaining stable rendition options, and internal analysis found that removing 270p would offer only marginal cache benefits while increasing costs and reducing service for low-bandwidth, small-screen, and machine-inference use cases. Rolled out gradually in April 2026, the new ladder shifted many viewers from both 480p and 1080p to 720p, while 270p usage remained stable, demonstrating continued demand for the lowest tier. The change increased delivered bytes per minute by about 6.5%, partly offset by encoding improvements, but raised Video Quality Scores from roughly 99.1 to 99.7 without harming seek latency or startup performance. Because 720p is priced below 1080p and within the same lower-cost bracket as 480p, the adjustment improved visual quality for many viewers while slightly reducing customer delivery charges despite Mux serving more data.

Mux’s free, anonymous video-hosting demo app stream.new has evolved its moderation system from a 2021 approach based on sampled thumbnails and separate Google Vision SafeSearch and Hive API integrations to a more capable, asynchronous workflow powered by Mux Robots. The original system helped identify NSFW and violent material but required custom score reconciliation, synchronous webhook processing, and substantial human review because frame-level classifiers could not reliably recognize contextual issues such as pirated television, watchalongs, or cycling-race footage. After modernizing the application stack, Mux first adopted the open-source @mux/ai toolkit, which simplified provider integrations and enabled multimodal models to evaluate plain-language policy questions using video context and automatically generated multilingual transcripts. Stream.new then migrated to Mux Robots, a first-party Mux API that runs video AI analysis near the hosted assets and delivers results through webhooks, removing the need to manage third-party model keys and much of the application-side orchestration. The current system automatically removes high-confidence violations such as full-length professionally produced films and watchalongs, while sending other flagged material to human reviewers, with thresholds and sampling rates refined through production monitoring.

Mux Robots’ Ask Questions workflow uses both video frames and transcripts to answer customizable questions about uploaded videos, extending beyond standard moderation scores for sexual or violent material. One example from stream.new asks whether a video is mostly of feet, content that may not trigger sexual-content filters but can violate platform rules. The feature can also verify listing completeness for rental or marketplace videos, check required advertising disclosures, document completed field-service work or returned equipment condition, identify static-image videos, and route support tickets based on whether footage shows software, hardware, or something else. Its multimodal approach is intended to overcome the limits of transcript-only analysis by interpreting visual content and narrative context, as illustrated through questions about the silent animated film Big Buck Bunny, where it identified characters, violence, child suitability, and plot outcomes from pixels alone. When a question is irrelevant or cannot be answered confidently, such as asking which programming language appears in the film, the system skips it rather than guessing.

After YouTube banned CraftingTable’s brand account, founder Austin Eckman lost access to all embedded course videos, disrupting lessons for schools and students. He used Claude Code to scan the website, extract YouTube links, download the videos, select Mux as a new AI-focused hosting provider, upload the content, and replace sitewide embeds with only a few prompts. Beyond restoring the courses quickly, the move to Mux provided transcripts that support a chatbot capable of directing learners to relevant video timestamps, along with expanded analytics on watch time, drop-off points, and student engagement to help improve course content.

Mux Robots Summarize can generate video titles, descriptions, and tags from transcripts, but production-quality results benefit from configurable guidance rather than default prompts alone. Users can set an overall tone, output lengths, tag counts, language, and granular output-steering settings such as summary style, intended audience, approved brand terms, and a controlled tag taxonomy. A NASA Artemis video example shows that default results may be accurate but include overly broad or duplicate-like tags, while steering can produce editorial copy, correctly formatted names, fewer user-friendly tags, and canonical categories that consolidate aliases such as “SLS” and “liftoff.” The workflow can also be refined with a coding agent that reviews an application’s video use cases, tests parameter combinations, presents results for review, and recommends settings before they are applied across a video library.

Inspired by television “squeezeback” effects that shrink live video to reveal adjacent content, the piece demonstrates how interactive web video layouts can be built with CSS Grid, Video.js v10, and WebVTT metadata cues rather than pre-rendered motion graphics. Animating grid track sizes lets a video resize while sponsor cards, product panels, forms, alternate views, or other elements appear in newly opened cells, with cue identifiers and payloads in a VTT file defining when layouts activate and allowing non-developers to retime experiences. The approach uses hidden text tracks and cuechange events to handle playback, seeking, looping, and scrubbing reliably, while a low-resolution canvas can generate blurred ambient video colors and dynamic accent colors when CORS and content-protection limitations permit. It also identifies implementation details such as matching CSS Grid fractional totals for smooth interpolation, properly registering animatable custom properties, using file-based rather than programmatically added VTT tracks because of an hls.js issue, and preserving form panels while users have focus. The discussion situates the technique within the history of broadcast digital video effects and suggests extensions including sign-language windows, synchronized presentation slides, alternate camera angles, persistent video handoffs into page-like layouts, and shoppable-video interfaces.

The journey of CDN selection at Mux highlights the challenges and evolution in finding an efficient and reliable method to choose content delivery networks. Initial attempts relied on third-party real user metrics (RUM) services, which proved insufficient due to limited control and data scope, leading to the development of a Mux-owned CDN selection process. This process initially faced issues with data dependence and latency-focused decision-making, ultimately leading to a more nuanced approach that prioritizes ruling out underperforming CDNs rather than finding the singular best option. The current system incorporates independent data sources and pre-calculated availability scores to enhance speed and adaptability, achieving a p99 latency of 9.9ms. Mux continues to improve this system by considering additional data sources and exploring content steering to further enhance resilience and performance.