GPT Image 2.5's Price Per Token Didn't Change. Nobody Knows the Price Per Image.

GPT Image 2.5 keeps GPT Image 2's rate card, sharper output, layered PSD export, and a sketch-to-image tool. What it costs per finished image is the number OpenAI didn't publish.

LLaDA-Image Isn't the Diffusion-Language-Model Trick It Sounds Like

Ant Group's LLaDA-Image borrows its name from a diffusion language model, but only half the pipeline actually works that way. Here's the part that's real and the part that's marketing.

Google's WeatherNext 3 Is Legit. Its Independent Referee Isn't Exactly Neutral.

WeatherNext 3 really does forecast rain more accurately than Google's last model. The leaderboard proving it is run by a company that sells weather forecasts too.

Runway Says GWM Worlds 2 Trains AI Agents. The Evidence Is One Robot and Two Flags.

GWM Worlds 2 streams real-time 720p worlds with synced audio and pitches itself as agent-training infrastructure. The actual training evidence is one demo.

SolarWM Trains on Five-Second Clips, Then Runs for an Hour Without Breaking

SolarWM trains video world models on five-second clips, then runs stable rollouts for hours by fixing a training mismatch instead of scaling up the model.

OpenAI Called Astra the Start of AGI.

OpenAI's Greg Brockman said Astra might be AGI. Artificial Analysis's independent index gave it the same score as the model it replaced, five points behind Anthropic's Fable 5.1.

Qwen3.8-Max-0902 Beat Claude Opus 5 on Code. Its Open Weights Didn't.

Qwen3.8-Max-0902 tops Arena's coding leaderboard at a fraction of Claude Opus 5's price. The open weights Alibaba promised turned out to be a different, stripped-down model.

Meta's New Model Is 92% Cheaper if You Let It Train on Your Prompts. Developers Are Taking the Deal.

Muse Spark 1.3 discounts developers who let Meta train on their prompts. The math, and what Wang's 'double-digit' adoption claim actually proves.

Claude Fable 5.1 Cracked a Five-Year Bug. Mythos 5.1 Remapped Venus.

Claude Fable 5.1 debugged a five-year production crash and pushed protein-design hit rates toward 50%. Its headline pricing claim holds up less well.

Tencent Won't Call Its Newest Model 'Hunyuan'

Tencent's new 770B open model closes in on GLM-5.3 and Kimi K3, yet its own materials never once attach the Hunyuan name to it. That's not a branding accident.

MiniMax's H3 Video Model Gets a 14x Speedup, Trained on Footage Nobody Ever Filmed

FastH3 cuts MiniMax H3 from 49 sampling steps to 4 using training data the model generated for itself, no real footage involved. The 14x number is real, on 8 B200s. I checked what a 4090 actually gets you, and what the license quietly locks out.

Z.ai Didn't Retrain GLM-5.3. It Just Trained It Meaner.

GLM-5.3 ties a rival four times its size through harder RL, not a new pretrain, and its license quietly excludes the hyperscalers big enough to undercut Z.ai.

A Robot Outran Usain Bolt in Beijing. It Still Doesn't Know How to Stop.

In Beijing, a robot ran 100 meters faster than Usain Bolt ever has, then crashed into the padding. The record isn't quite what it looks like.

AI World Models Have an Amnesia Problem. This One Hired a Programmer to Fix It.

Code World Model splits the job in two: a coding agent tracks what's actually happening, a fine-tuned MiniMax H3 renders it. It's a clean fix for AI world models' memory problem, on paper.

The Training Trick That's Cutting AI Image Generator Costs by Up to 63%

ByteDance Seed's DiffusionOPSD swaps one vague final-image score for precise per-step training targets, and the GPU savings are the real headline.

MetaMotivo Never Rehearsed Standing Up. It Just Knows How.

Meta's MetaMotivo controls a virtual humanoid through hundreds of tasks it was never specifically trained on, no retraining, no planning, just a prompt.