16 AI stories explainx.ai reported on September 2, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguard levels — Fable 5.1 is generally available and roughly 25-45% cheaper per task than Fable 5, mainly from a 75% cut to cache-read pricing.
OpenAI confirmed Astra reached the Critical cybersecurity threshold in its Preparedness Framework and is previewing a defense-first release: strict safeguards for offensive cyber capability, broad access for everything else.
H3 Max can generate short clips quickly enough for continuous AI television, but a nonstop stream costs thousands per day and still lacks narrative memory.
Google Pics is a Nano Banana 2-powered image editor built into Workspace and Docs/Slides that lets teams do object-level edits and text translation collaboratively, but it still lacks the layers, masking precision, and brand-asset management that make Canva and Adobe sticky for professional work.
ChatGPT for Healthcare can now pull authorized, read-only Epic patient data into a chat or embed inside Epic's own interface, plus query nine public clinical databases directly — evaluated by physicians at 99.1% safe across 4,363 ratings on 27 clinical use cases.
A fake "Claude Opus 5 Free Desktop" app distributed on GitHub and game-cheat sites installs RevStealer malware, which targets 50+ crypto wallets and 12 password managers, then deletes itself before you notice.
AlphaEvolve is a Gemini-powered evolutionary coding agent that optimizes code against user-defined, client-side evaluator metrics like latency, memory, or score.
World Labs' Atlas generates video with an exact, hand-designed camera path and directly outputs 3D point clouds or Gaussian splats from the result, combining two capabilities — generation and reconstruction — that have mostly lived in separate models until now.
Gemini's new agentic video mode lets the model navigate a video's timeline and fetch only the frames, audio, or transcript it needs to answer a question, cutting tokens up to 88% instead of sampling every frame at a fixed rate.
Community contributors on an open mlx.fast leaderboard pushed Gemma 4 26B A4B's Apple Silicon inference speed up 130.3% over baseline, which Google rounded to "2x faster" in its own announcement.
The Fable 5.1 leak is the claude.ai chat system prompt (roughly 180,000 characters, not 270,000) describing how the memory feature works — not an actual leak of any real user's private memories.
Qwen3.8-Flash is the hosted, production-featured version of the same Qwen4-architecture-preview model Alibaba open-weighted as Qwen3.8-Flash-Next on August 26 — now priced at $0.16/1M input and $0.47/1M output tokens with 1M context and built-in tools by default.
DeepSeek's first vision-capable model is now free MIT-licensed weights on Hugging Face, not just an API — a 305B-parameter multimodal model that DeepSeek's own benchmarks put close to Opus 4.8, a model two generations behind Anthropic's current lineup.
Gemini 3.8 Flash and Gemini 3.8 Flash Cyber are now officially confirmed by Google, at the same $0.75/$3.75 per-million-token pricing as 3.7 Flash, with published DeepSWE, HLE-Verified, and CyberGym benchmark numbers.
Grok 4.6 was the only model LatchBio tested that scored above 50% on both refusing disguised bio-hazard tasks (59.2%) and completing routine biology research (64.8%), in an independent evaluation xAI commissioned and published alongside Anthropic's and OpenAI's own biosecurity disclosures the same week.
Muse Voice Transcribe is Meta's first real-time audio perception model, scoring 3.1% streaming word error rate and leading diarization benchmarks by folding ASR, speaker separation, and endpointing into one model instead of three.
Dan Luu fact-checked years of Ed Zitron's specific, dated AI predictions and found most did not come true — a case study in why confidence and citation count are not the same as a good track record.
Use hosted H3 Max for its sub-playback latency; use base MiniMax H3 locally when privacy and control matter more than matching fal's speed.
Neural rendering is a learned stage after the render, not a generator — it excels where lighting is under-authored and fails where art direction is deliberate.
Recurrent depth makes models reason more by looping the same layers repeatedly on hidden vectors instead of writing extra text, which boosts capability but makes that reasoning invisible to chain-of-thought monitoring.
+18 more updated posts
The Computable GPU Index is the first open-source price index for GPU compute, providing valuable market insights.
Gauth AI Course offers a comprehensive platform for learning AI through videos, quizzes, and content creation.
Sider Code allows users to reshape any website using simple language through its browser extension.
Kilo Code is a fully native, open-source coding agent designed specifically for JetBrains, enhancing the coding experience.
A unified API for accessing multiple AI models with enhanced privacy features.
Get each day's AI news in your feed reader: daily RSS · every post
