AI — Page 8

Ethan Mollick says Claude Fable 5 worked for hours across research and coding tasks, offering a long-horizon outside read on the Amodeis' agent bet.

AnimateDiff stays coherent, but coherence alone doesn’t win head-to-heads when the model keeps dropping the brief. Seedance 2 Image to Video was dramatically better at actually staging the scenes it was asked to make.

Sapient Intelligence's HRM-Text claim targets the enterprise fear that custom AI means frontier-model budgets and vendor lock-in.

AuraFlow has flashes of atmosphere, but Imagineart 2.0 Preview wins this matchup decisively by following the brief, handling text, and delivering more convincing scenes. The 26.3 to 18.5 scoreline flatters AuraFlow.

Concentrate emerged from stealth with more than $5 million, while OpenRouter's recent $113 million round shows how fast the gateway layer is heating up.

This was close on aggregate, but grok-4.3 wins because it made the fewer costly mistakes in structured-output work. gpt-5.4-nano was sharper on tone and regex edge cases, yet it gave back those gains by breaking instructions where precision mattered more.

Pi is valued at $100 million and counts Navan as an early customer, while Forbes reports xAI is also using the system.

Anthropic says Fable 5 routes under 5% of sessions to Opus 4.8, while Mythos 5 keeps higher-risk capability behind trusted access.

Instacore puts five cameras and a compute backpack on workers to capture commercial tasks for AI labs, with customers still unnamed.

Across both prompt-following tests, Luma Ray 3.2 Image to Video wasn’t just better than AnimateDiff—it was operating in a different league. AnimateDiff could gesture at mood; Luma delivered the actual scene, action, and camera logic the prompts asked for.

Bagel steals one poster task, but Fibo Bbq Preview wins the matchup where it matters: prompt control, scene construction, and mood. On aggregate, B is the more reliable image model and the clear overall pick.

The new model is priced at $10 per million input tokens and $50 per million output tokens, with some requests routed to Opus 4.8.

The audio model streams speech-to-speech translation across 70-plus languages, with Meet access limited to private preview this month.

grok-4.3 takes this matchup by being the stricter, cleaner finisher on structured-output tasks, while DeepSeek-V4-Flash wins the one audience-sensitive writing test. The scoreline is close, but the deciding errors are the kind that matter in production.

The UIUC, UC Berkeley and Chroma project shifts search memory from the model context window into a structured software environment.

AnimateDiff is the steadier clip-maker, but Marey Realism V1.5 is the better prompt reader and the more convincing filmmaker. Across both tests, it delivered the details, atmosphere, and camera language the prompts actually asked for.

AuraFlow can make attractive images, but this matchup wasn’t about vibes alone. ImagineArt 1.5 Pro Preview won all three tasks by doing the harder thing consistently: following the prompt in specific, visible ways.

WSJ reports Jane Street is now an investor in Situational Awareness, whose biggest disclosed win is tied to Anthropic.

grok-4.3 wins this head-to-head 37.0 to 30.0 by being the more obedient, production-ready text model. Across four tasks, it was consistently tighter on instructions and cleaner on edge cases, while DeepSeek-V4-Pro kept drifting into unnecessary constraints or formatting mistakes.

Tencent Hunyuan and university collaborators say current models post an Exact Match Rate below 5% on the new speech and audio editing benchmark.

DeepSeek V4 Pro wins this head-to-head by being more exact where it matters: following instructions, matching schemas, and solving edge cases cleanly. GPT-5.5 Pro is still strong, but it gave away points with avoidable deviations.

In a decisive victory, Happy Horse outshone AnimateDiff in cinematic motion and dynamic environment tasks.

The Three.js visual demo points to a smaller but important market for coding agents: creators selling workflows, not software seats.

The Build Small Hackathon track asks developers to build with smaller models as open-weight systems move closer to local production use.