Made a tiny video editor, then hired a little orange guy to run it . Cuts, trims, gets confused sometimes, but always does his best. Built with Claude Opus 5.5, music too.

I asked Opus 5.5 to make a watch in Blender, just code, no downloaded models or textures. It built the whole movement too. @claudeai

Opus hasn't hit the same since 4.6. 5.5 finally does.

I expected Opus 5.5 to land near Fable 5.1. Didn't expect it to land above it. 58 vs 53 on the Artificial Analysis index, and it does the same work 40% cheaper than Opus 5. Running my own tests on it now, results coming soon.

🚨 MiMo V2.6 Flash is free in OpenCode for a week. Good week to find out what it actually does.

🚨 Grok 4.7 got a week of hype. MiMo V2.6 Pro just put this out. Grok's first real tests aren't backing the buildup. MiMo came with none of it. Open weights, and this is what it does on game dev. That's a wild place for an open model to be.

Grok 4.7 is out. The table looks fine. Terminal-Bench goes 20.3 → 38, it tops legal and electrical, and the price stays where 4.6 was. Two things though. 4.7 is measured on xHigh, 4.6 on High. And outside the benchmarks, nothing looks different yet.

GPT-6 Sol slips a week. Opus 5.2 leaks keep piling up: routing changes, faster outputs, and silence from Anthropic. The pace between labs right now is relentless.

New stealth drop: Union Alpha. > Free on OpenRouter and OpenCode > Claims "frontier-level" performance Any thoughts who it could be? Maybe Kimi?

Leaked Opus 5.2 output and the visual jump is real. Astra's had a big lead here, this might close the gap.

Opus 5 drew all of this, frame by frame, in code. The detail is wild

What GPT-6 Astra built here is a full motocross simulator, not just a scene. The line between "AI demo" and "AI built a game" keeps getting thinner.

One-shot. Single HTML file. No harness. DeepSeek-V4.1-Flash one-shotted this writing automaton in Three.js. Impressive.

Every "slow down AI" call deserves one follow up question: slow down what, exactly? Training runs? Nobody will unilaterally. Deployment? That's just a release calendar, not a slowdown. Compute buildout? Ah, now we're at the real thing: a bidding war for GPUs and gigawatts that's wrecking everyone's margins. A slowdown that never names its axis isn't safety policy. It's an oligopoly asking to stop bidding against itself, with a safety logo on it.

Introducing voice, music, image, and video generation in the ElevenLabs MCP. Generate speech, transcripts, dubs, music, sound effects, images, and video from the assistant you already work in.

Not long ago, picking a model was easy: Anthropic for quality. Then GPT-6 Astra shows up, matches it in some areas, outperforms it in others, and beats it on speed and price. So what does the premium buy you now? A bigger bill.

Today, Anthropic quietly cut Claude Code's weekly usage limits by ~17%. No new model, no counter-announcement to GPT-6 Astra. Just less usage, right as the competition heats up. Not sure how that math works out for them.