Coverage

News page 2 of 4

Comet's Opik adds a judge that returns a probability and cannot read images

A judge model that never explains itself is now a default option in Comet's eval platform. The catch for anyone scoring generated media is written into the same documentation page: traces with images still need a vision model.

news 6 Oct 2026

Ideogram 4.5 ships a no-drift editing claim that only Ideogram has scored

Ideogram's newest model is pitched at the tenth edit looking as clean as the first. The only published evidence is a comparison the company ran against three rivals, and no public leaderboard measures multi-turn editing at all.

news 5 Oct 2026

Kling 4.0's 30-second clips are still an October promise, with only a 720p Flash tier live

The specs Kuaishou put in the Kling 4.0 announcement belong to a model nobody can run yet. The one that is running tops out at 20 seconds and 720p, and only for one subscription tier.

news 4 Oct 2026

An image-understanding leaderboard dated October 1 carries sourced scores for only eight models

Before a model can score a generated image, it has to read one. The public benchmark evidence for which models do that well covers eight entries, and the site publishing it says so plainly.

news 4 Oct 2026

AWS open-sources a 2B model that scores options with a probability, and it cannot read images

Amazon's new decision model has the exact shape you would want from an automated judge: a fixed label set, a probability per label, and a sub-second answer on a laptop. It also has no image input, which is where the story ends for anyone scoring generated video.

news 3 Oct 2026

AI video provenance in 2026: watermarks, C2PA and the new disclosure laws

Paying to remove a visible watermark does not remove the invisible one. What the EU and California now require, and the four layers of AI video provenance.

guide 3 Oct 2026

Why OpenAI walked away from Sora, and why video specialists keep winning

Sora did not lose on quality. It lost on priorities. What that says about who can win AI video, and why open weights make a model lead short-lived.

news 3 Oct 2026

Where AI video still breaks: motion over time, not single frames

Our hardest prompts all ask a model to keep something physically consistent while it moves. That is the frontier for production use.

study 3 Oct 2026

What a minute of AI video actually costs in 2026

Raw AI video costs a few dollars a minute. A finished minute costs a hundred times more. Here is where the gap comes from.

guide 3 Oct 2026

AI video models are converging on quality. Their prices are not.

Quality at the top of AI video has bunched up. Price, control and where a model is sold now decide more than its rank on a leaderboard.

study 3 Oct 2026

Dynatrace closes its Arize acquisition, putting a widely used LLM-judge stack under an observability vendor

One of the default places teams run LLM-as-a-judge evaluations now belongs to an enterprise observability company. The announcement talks about agents and applications, and says nothing about scoring generated images or video.

news 2 Oct 2026

Cloudflare's Clef returns probabilities instead of prose, and it reads images

A model that never writes a sentence, only scores the options you give it, is the shape an image or video scoring rubric has always wanted. Cloudflare's published evidence for Clef is a domain classification task, not a media one.

news 2 Oct 2026

OpenAI puts ChatGPT Images 2.5 to work on virtual clothing try-ons, with no accuracy figure

An image model has been pointed at a task where the output can be checked against reality, and the launch note ships without a single quality number attached.

news 1 Oct 2026

Tavus says Griffin-Lite passed a video Turing test, with 48% of 54 testers fooled

A real-time avatar model that listens while it talks, scored by a 54-person call study the vendor designed and ran. The one outside measurement in the launch comes from a benchmark NVIDIA built.

news 1 Oct 2026

An image-generation leaderboard refreshed September 28 ranks 10 models on 11,311 blind votes total

Ten models, two task types, 11,311 votes in total, and a scoring formula that subtracts three standard deviations of uncertainty from every model's estimate. Here is what that ordering can and cannot tell you.

news 29 Sep 2026

221,900 new AI shows launched on Douyin in six months, and 1,055 cleared 100 million views

Chinese local governments are subsidising AI filmmaking the way they once subsidised solar panels. The production numbers are enormous and the hit rate is under half a percent.

news 26 Sep 2026

Google ships Live Avatar for Gemini 3.8 Live, with 97 languages and no comparative quality numbers

Google's newest video-generating product is a talking enterprise avatar, and the launch material describes what it does without measuring how well it does it. The only hard figure on offer is a language count.

news 26 Sep 2026

Google's Gemini 3.8 Flash TTS claims #1 on Hume's Voice Design Benchmark at 71.4, with ElevenLabs still ahead on two subscores

Google put a leaderboard number on its new speech models, 71.4 and a first-place finish on a benchmark run by a competing voice AI company. The subscores tell a more mixed story, and the blind preference results come without vote counts.

news 25 Sep 2026