Anthropic debuts Claude Fable 5.2 and Mythos 5.2
New frontier models priced 20% below GPT-6 Astra, featuring 30% faster token generation and improved coding/reasoning.
ddNews Β· a daily AI briefing for Dirk (generative video/ddStudio, Lit
New frontier models priced 20% below GPT-6 Astra, featuring 30% faster token generation and improved coding/reasoning.
a specialized model designed to find and exploit software flaws, already uncovering 588 unknown Linux vulnerabilities.
funding a new delivery drone factory to scale autonomous logistics.
Claude Fable 5.2 is reportedly outperforming GPT-6 Astra in key benchmarks.

new frontier models solving open physics problems with 20% lower API costs and 30% faster token generation.
New frontier models showing significant gains in coding, reasoning, and vision.

A lightweight, low-cost model optimized for repetitive agent tasks like tool calls and routing.

A lightweight, low-cost model (10 cents/1M tokens) optimized for agentic tool-calling and routing.

A professional image model supporting stable multi-pass editing over 100 iterations and native RGBA transparency.

A professional image model supporting stable multi-editing over 100 rounds and native RGBA transparency.
The company is securing its own power infrastructure to combat the AI energy crunch.

A new lightweight model optimized for agent execution at a low cost of 10 cents per million input tokens.

Anthropic introduced a background memory mechanism to improve the assistant's accuracy and recall.

a massive infrastructure move to secure energy for its expanding compute clusters.

The companies are co-developing generative AI tools for Disney's parks, studios, and ESPN.

The companies are developing brand-safe generative tools for Disney's parks, studios, and ESPN.
the startup provides AI-powered hardware-testing and engineering software.

New data shows continuous batching outperforms static/dynamic methods by up to 2.91Γ.
Growing trend of using ~27B models on consumer GPUs (8GB) for specific enterprise needs.

Vinces Yao Shunyu advocates for a middle ground between open and closed models to mitigate security risks.
introduces TPU 8tx for secure data processing and TPU 8i Max for doubled compute/memory.
The platform now includes agentic AI operations for automated troubleshooting and enhanced security.
the project is contributed to the CNCF to simplify diverse enterprise operations.
updates reduce overhead for durable workflows used in long-running AI agents.
the firm expands its enterprise AI capabilities following a previous OpenAI partnership.

Users can now generate custom Story filters directly from text prompts.
New TPU 8tx (secure inference) and TPU 8i Max (double compute/memory) target Nvidia's market share.
X users report a new upgrade to Gemini 4 Carbon currently in testing.

introduces TPU 8tx for sensitive data protection and TPU 8i Max for doubled compute/memory.

the espionage allegations may weaken Huang's lobbying efforts against tighter US AI chip export controls.
The company reaches a $1.25tn valuation in its latest funding round.

a native PyTorch backend allowing models to run on AWS Trainium chips without code modifications.
the GPU giant is joining the Series C funding round for the South Korean AI model developer.
OpenAI's new cybersecurity model is uncovering serious flaws in Firefox and the Linux kernel before hackers can.

Pinterest replaced a multi-stage clustering pipeline with a fine-tuned 4B model, increasing engagement.
The South Korean AI firm will unveil "Motif 4" in Q1 2027.
New frontier models featuring improved coding and reasoning with a 20% API cost reduction.
OpenAI's agent pipeline identified nearly 1,000 root-access vulnerabilities.
the model uses 80ms inference cycles to support natural backchanneling and web search.
A strategic shift toward researching the mechanisms of AI code generation rather than just the output.
The government released EKA-1 (26.3B) and offline-capable DhΔ«r models (2B and 5B).
OpenAI announced GPT Image 2.6, allowing users to turn generated images into GIF animations.
The open-weight model nearly matched GPT-6 Sol Daybreak in cyber-intrusion benchmarks.
Carnot's "Jinba" platform signals a move from PoC to production-grade local inference for core business operations.
the company is investing in gas-fired generators to secure 1.7 GW for its data centers.
AWS Trainium is shipping now, while Azure's Maia 300 is slated for 2027.
The model can now edit 2,048x2,048 images in under 8 seconds.
a new wiki-fication process for AI agents increased memory success rates from 13% to 48%.
the board is probing Huang's relationship with a suspected Chinese intelligence agent.
Now supports AI video generation up to 15 seconds with improved lip-sync and conversational editing.
now supports AI video clips up to 15 seconds with improved lip-sync and conversational editing.
Anthropic's coding agent has captured a 39% work-use share by mid-2026.
The model proved 70% of Lean theorems on the TCSExam benchmark but failed all 30 open conjectures.
The former Meta executive joins the AI infrastructure firm as CFO.
OpenAI's internal safety team flagged the model as a top-tier risk due to its ability to build working exploits.
A new terminal tool by Microsoft VP Scott Hanselman to monitor per-process VRAM and LLM KV-cache.

OpenAI employees used the model to design a chassis and PCB, moving from prompt to sample in under a month.
A specialized model for identifying and validating security flaws with a 51% performance increase.
Now supports AI video generation up to 15 seconds with 30% lower pricing.
Unveils a full-stack intelligent private cloud suite for enterprise agent deployment.
OpenAI's internal safety team assigned the highest risk level due to the model's ability to build working exploits.
A new feature allowing users to connect bank accounts for spending analysis.

The company is probing Huang's relationship with a suspected Chinese intelligence agent.

New framework for building multi-agent systems using fan-out, pipelines, and swarms.
A massive 2.8 trillion parameter model released for free with 1.56TB of weights.

Built in a weekend using Claude Opus 5.5 to curate 6,454 videos from 43M Hacker News comments.

CEO John Ternus is pushing for a powerful AI agent to compete with OpenAI's Dots and Meta's Muse.

using Claude Code, Karpathy filtered 43M Hacker News comments into a curated video platform.
The company is securing gas-fired generators to ensure data center energy stability.
A major agreement for gas-fired power generators to ensure energy stability for AI data centers.
A major agreement for gas-fired generators to ensure energy stability for its AI data centers.
A specialized cybersecurity agent designed to identify vulnerabilities and build working exploits.

Rights holders for Elvis Presley and Eminem are suing over illegal music generation.
A spec-first coding tool that enforces five review stages before agent execution.

Turing winner Richard Sutton argues that raw compute scaling is insufficient and fundamentally new learning methods are required.
Adds GIF animation capabilities and 50% faster image generation.

A new collaboration infrastructure designed specifically for human-AI coding workflows.
Adds GIF animation capabilities, 50% faster generation, and improved visual consistency.
An offline reasoning process that allows agents to reorganize memory, boosting complex research scores to 87%.

A shift toward simulating real-world physics to move AI from text-based logic to 3D spatial navigation for robots.

Alibaba's Co-Design Bench shows GPT-6 Astra only fully solves 26.4% of real-world tasks.

Alibaba's Co-Design Bench shows GPT-6 Astra solves only 26.4% of complex real-world user requests.
Google's latest model iteration arrives just 15 days after the Argon release.
The AI inference startup's valuation has doubled to $16 billion to meet surging infrastructure demand.

Now generally available for VMware Cloud Foundation 9.x on Z3 bare metal.

Redesigned serverless SQL infrastructure to support millions of databases without bottlenecks.
The model successfully identified 81 Firefox vulnerabilities and 948 Linux exploits.
The new GPU features 32 GB GDDR7 and a launch price of $1,999.
Open-source tool for monitoring per-process VRAM and LLM KV-cache usage.

Updated guidelines on risk categories, deadlines after the Omnibus, and enforcement fines.

The company is seeking a valuation between $1.25 and $1.75 trillion despite a broader tech selloff.

GPT-6 Sol Daybreak hits 80.8% of attack goals, while open-weight DeepSeek V4.1 Flash reaches 72.8%.

The model successfully cracked 80.8% of attack targets in simulated corporate network benchmarks.

Nous Research successfully reduced pro-Chinese bias in Alibaba's model from 22% to <1%.

In the "Into the Unknown" benchmark, none of the nine tested agents met the expert bar for open research questions.

Nous Research successfully reduced pro-Chinese bias in Alibaba's model from 22% to under 1%.

Reports suggest US chips were sold to "empty offices" in China due to poor due diligence.

Tiger Data releases an open-source extension using PRISM indexing for scalable vector search.

Google takes over a troubled data center project, highlighting the difficulty of massive AI site development.