Claude Fable 5.1 and Mythos 5.1 land cheaper and less locked down

AIpositive +0.8#model-release #fable #mythos #agents #anthropic

Anthropic's new Fable drop beats Fable 5 on agent evals at lower cost, Mythos 5.1 ships alongside it, and Perplexity already wired Fable in as the high-stakes orchestrator.

What they showed / shipped

Anthropic published Claude Fable 5.1 and Claude Mythos 5.1. The Verge notes Fable is up to 45% cheaper for agentic work and less restrictive than prior Fable.
Perplexity says Fable 5.1 ranked first on its August WANDR eval at 0.601 and $12.76 per task — 21% higher score and 37% lower cost than Fable 5.
Arav is running Fable as the frontier planner on Perplexity Computer, with GPT-5.6 Terra as cheaper subagents.
The HN thread on the Anthropic post hit ~900 points the same day.

Why it matters

A frontier planner that is both better and cheaper is the rare release that changes your default stack the same afternoon.

Sources

Perplexity Mac app gets hybrid cloud-to-local compute

AIpositive +0.8#agents #tooling #perplexity #hardware

Computer can start a task in the cloud, then hand sensitive steps to a local model on your Mac so bloodwork, taxes, and case files never leave the machine.

What they showed / shipped

Arav announced hybrid compute for every Perplexity Mac app user: Computer orchestrates local models for private files.
The official Perplexity post: cloud start, local model for sensitive data, available today.
Same day Arav flagged the next risk: agents will spin on-demand GPU nodes to train themselves, so guardrails matter before that becomes normal.

Why it matters

This is the privacy primitive agents have been missing. Cloud for horsepower, local for secrets.

Sources

OpenAI previews Astra at Critical cyber capability

AImixed +0.1#model-release #safety #openai #astra

OpenAI says Astra hits the Critical threshold on its Preparedness Framework for cybersecurity, Sam is pacing the release for safety, and the Path to Astra writeup is public.

What they showed / shipped

OpenAI's Astra post says the model reaches Critical cybersecurity capability under the Preparedness Framework, with eval and safeguard previews.
Sam Altman: Astra has been done training for a while, is a step up in capability and alignment, and they are pacing the launch while slowing later models for more safety work.
The companion essay Path to Astra landed on HN. TechCrunch frames it as a model that is very good at breaking into computer systems.
The Verge adds that OpenAI delayed development after the Hugging Face hack.

Why it matters

Critical cyber is not a vibe. It is OpenAI admitting the model is dangerous enough that release pacing is the product decision.

Sources

World Labs ships Atlas, a spatial world model

AIpositive +0.8#world-models #video-gen #world-labs

Fei-Fei Li's lab launched Atlas: text and image in, up to a minute of spatially consistent video with camera control, including re-shoots of real phone footage from new angles.

What they showed / shipped

World Labs published Atlas: A World Model for Spatial Intelligence.
Venturetwins: Atlas takes text, image, and camera paths, and generates up to 1 minute of video with spatial consistency.
A separate demo reconstructs a scene from three phones on tripods and re-frames the clip from a new angle.
a16z clipped Fei-Fei: world models are a fundamentally different problem from language, because the 3D world has physics and materials, not words in the sky.

Why it matters

This is spatial intelligence as a product, not a research slide. Camera-path control is the API surface creators and agents will actually use.

Sources

Gemini gets agentic video understanding at 88% fewer tokens

AIpositive +0.8#model-release #video-gen #agents #gemini

DeepMind says the latest Gemini models can analyze video with better accuracy while burning up to 88% fewer tokens.

What they showed / shipped

Google DeepMind: agentic video understanding on the latest Gemini models, better accuracy and up to 88% fewer tokens.

Why it matters

Video agents were token-expensive. An 88% cut turns long-form video into something you can afford to run in a loop.

Sources

x/@GoogleDeepMind X@GoogleDeepMind
♥ 1.1K↻ 103👁 83K

A small transformer trained in 1.5 hours beats many LLMs on ARC

Sciencepositive +0.7#science #benchmarks #open-source

An HN writeup shows a tiny transformer trained in about ninety minutes outscoring much larger LLMs on ARC-style tasks, a reminder that scale is not the only lever.

What they showed / shipped

Front-page HN: I trained a small transformer in 1.5hrs and it beats many LLMs (~566 points).
The claim is narrow and checkable: a small, purpose-trained model on ARC-style reasoning, not a general chat model.

Why it matters

Specialized small models still win on structured puzzles. Useful counterweight to 'just call the frontier API.'

Sources

hn/small transformer beats LLMs on ARC mvakde.github.io

AI sitcoms are getting watchable, and Higgsfield remakes the whole scene

Mediapositive +0.7#video-gen #creator-economy #higgsfield

Multi-episode AI sitcoms like Bad Cat are holding attention, while Higgsfield Genjutsu swaps character and scene without breaking the original acting and edit.

What they showed / shipped

Venturetwins is four episodes into Bad Cat and calling AI sitcoms 'incredibly watchable.'
Min Choi: Higgsfield Genjutsu changes character and entire scene while keeping original acting, camera movement, and edit.
📌 On the radar: Runway added ACES EXR export in Ruby for pro color pipelines. Grok Bot template roundups keep spreading (minchoi, catalog site).

Why it matters

Genjutsu is instruction-level scene surgery. If acting and edit hold, reshoots get cheap.