Monday, July 13, 2026

OpenAI launches GPT-5.6 model family with Sol, Terra, Luna

GPT-5.6 Sol Ultra produced a proof of a 50-year-old math conjecture. GPT-5.6 tops the DeepSWE leaderboard at 73% and sets a new SOTA on ARC-AGI-3 at 7.8%.

Anthropic redeploys Claude Fable 5 globally

Anthropic redeployed Claude Fable 5 on July 1, 2026, after US export controls were lifted. The model is available globally with new safeguards that block cybersecurity tasks, and includes up to 50% of weekly usage through July 7. Claude Mythos 5 is also restored for US organizations via Project Glasswing.

Anthropic launches Claude Sonnet 5 with 1M-token context window

Claude Sonnet 5 is Anthropic's most agentic Sonnet, priced at $2/$10 per million tokens through August 31. It ranks second only to Fable 5 on the AA-Briefcase agentic benchmark and is now the default model in Claude Code. Available on AWS Bedrock, Perplexity, and Databricks.

Meta launches Muse Spark 1.1 with API and agentic focus

Muse Spark 1.1 scores 51 on the Artificial Analysis Intelligence Index, up 8 points from 1.0, and is cost-efficient. Meta claims significant improvements in agentic tool calling and computer use, with a 43-point improvement on DeepSWE. The model is available via the new Meta Model API (not in EU).

ZAI releases GLM-5.2 open-weight model

GLM-5.2 is a sub-trillion parameter open-weight model with significant improvements in coding and agentic tasks. It achieves top scores on Terminal Bench 2.1 and runs at 120 tok/s on two networked Blackwell tinyboxes.

SpaceXAI's Grok 4.5 tops WANDR benchmark, lands on Perplexity and Cursor

Grok 4.5 scored highest on Perplexity's WANDR benchmark at roughly half the cost of Claude Opus 4.8. It is available as an orchestrator model in Perplexity Computer for Pro/Max and Enterprise users, and free on X via Grok Build.

OpenAI rolls out GPT-Live voice model to all ChatGPT users globally

The full-duplex voice model supports real-time translation and can delegate tasks to GPT-5.5. Users report impressive performance in non-English languages, but Plus subscribers face daily limits of 30-60 minutes.

OpenAI releases GPT-5.6 models with restricted US government-led rollout

OpenAI launched GPT-5.6 in Sol, Terra, and Luna versions, but the release is limited to a small group of trusted US partners. The US government mandated this staggered, consumer-by-consumer preview process, delaying general availability for international users.

Claude Fable 5 scores 60, exceeds Opus 4.8 and GPT-5.5

AI inference startup Baseten reportedly raising $1.5B at $13B valuation

Baseten is reportedly closing a $1.5 billion funding round at a $13 billion valuation, per WSJ. The round comes just five months after its $300 million Series E at a $5 billion valuation. Some investors are entering at an $11 billion valuation in this split-priced round.

GPT-5.6 Sol's usage limits frustrate Pro subscribers

Users report GPT-5.6 Sol on High burns through the 5-hour limit in about 2 hours, consuming weekly quota rapidly on the $200/month Pro plan. Compared to GPT-5.5 on XHigh, the same workload lasted significantly longer, with some users exhausting all three resets.

OpenAI's ChatGPT Work targets data analytics teams

ChatGPT Work helps data teams turn messy questions and scattered files into actionable analysis. The product is designed for business teams needing quick insights.

Lin Junyang AI Lab Closes Round at $2B Valuation

Lin Junyang, former head of Alibaba's Qwen team, closed a first financing round at a $2 billion post-investment valuation. Gao Rong Venture Capital and Sequoia China each contributed $100 million, with Tencent adding $20 million. The lab targets world models and embodied intelligence, not general LLM development.

Anthropic export ban alarms AI industry

Axios reports new U.S. export controls restricting Anthropic's AI models and technology to certain countries. The move signals tightening national security restrictions on frontier AI labs.

Trump administration asked OpenAI to stagger GPT-5.6 release

The Trump administration asked OpenAI to stagger the release of its GPT-5.6 model, according to a Bloomberg report. The request signals a potential shift where the US government may require approval before AI model releases.

White House to approve GPT-5.6 access per customer

The Trump administration asked OpenAI to stagger GPT-5.6 release over security concerns, with customer-by-customer approval from the government. Critics decry the opaque, ad hoc process as a dangerous policy precedent.

Claude Code 2.1.207 released with auto mode, bug fixes

Auto mode is now available without opt-in on Bedrock, Vertex AI, and Foundry; disable via disableAutoMode. Release includes 24 CLI changes and 2 system prompt changes, fixing terminal freeze, prompt-injection warnings, and more.

Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper

Migrating an AI agent to GPT-5.6 delivered 2.2x faster response times and 27% cost savings compared to previous model. The migration process involved minimal code changes and resulted in improved reliability.

RLM: Recursive Language Models for Large Codebases

Recursive Language Models (RLM) from an MIT paper help coding agents handle large codebases by loading the repo into a programmable REPL. The model writes code to inspect it and recursively delegates sub-tasks to avoid context overload.

Hidden Decoding at Scale proposes latent computation scaling for LLMs

The paper 'Hidden Decoding at Scale' investigates scaling LLM performance by increasing computation at inference time rather than pretraining larger models. Experiments demonstrate that an existing backbone can continue to improve with more compute allocated during decoding.

Hugging Face CEO discusses shift toward self-hosted AI models

OpenAI teases ChatGPT-Codex merge and pet features in upcoming livestream

OpenAI announced a livestream scheduled for today at 10am PT with teasers suggesting a merger of ChatGPT and Codex, as well as the addition of pets to ChatGPT. The event is expected to reveal these new features. No further details have been provided.

Brain-inspired hardware brings faster, lower-power anomaly detection to AI systems

A new brain-inspired hardware design speeds up anomaly detection in AI systems while cutting power consumption. The approach mimics neural network principles for more efficient processing.

Zapier automates lead QA with ChatGPT Work, generates seven-figure monthly pipeline

Head of Enterprise Marketing at Zapier, Angela Ferrante, reports that ChatGPT Work automates a process that previously took 35–45 minutes per lead, enabling QA of thousands of inbound leads monthly. The impact was quantified as seven figures in pipeline every single month.

Databricks: pi-coding-agent 2x cheaper than CC/Codex, GLM 5.2 matches Opus 4.8

Databricks benchmarked coding agents on a multi-million line codebase. pi-coding-agent costs half of CC/Codex with similar performance. GLM 5.2 achieves parity with Opus 4.8 high.

Meta upgrades Muse Spark AI coding model to compete with Anthropic and OpenAI

Meta is upgrading its Muse Spark AI coding model under AI chief Alexandr Wang. The move aims to catch up with Anthropic and OpenAI in the competitive AI coding market.

Grok integrated into X Ads platform for advertisers

Seedream 5.0 Pro now available on Vercel AI Gateway and Pika MCP

The BytePlus model generates text without spelling errors, dense infographics with charts, and realistic portraits. It is also available on the Pika MCP for editorial-grade photo generation.

GPT-5.6 Soul autonomously post-trained Luna model

OpenAI used GPT-5.6 Soul to autonomously post-train the smaller Luna model, with the larger model generating training data, evaluating outputs, and shaping Luna's behavior with minimal human involvement. This showcases recursive self-improvement in AI.

Anthropic launches Artifacts in Claude Code for live, shareable dashboards

Artifacts turns Claude Code session output into interactive HTML pages with private links. Available in beta on Team and Enterprise plans. Users can share live dashboards or PR walkthroughs with their team.

Qualcomm CEO is building the 'Linux of AI'

Qualcomm CEO Cristiano Amon envisions a future where AI runs on any device, from phones to cars. He describes Qualcomm's strategy as building the 'Linux of AI' for on-device inference. The company aims to make AI ubiquitous by running on the edge.

Netflix releases video datasets and models on Hugging Face

AI demand remains strong, execs say, despite enterprise 'valuemaxxing'

AI-related chip stocks have been volatile amid a debate over AI demand and spending. Executives describe enterprise 'valuemaxxing' but say demand is 'almost unlimited'.

Xiaomi's HarnessX rewrites its own AI scaffolding mid-task — and smaller models gain the…

Xiaomi's HarnessX dynamically rewrites AI agent scaffolding during tasks, with experiments showing smaller models benefit most. The approach improves performance on long-horizon tasks by adapting the harness mid-task.

Field Guide to Fable explains capability overhang in Claude Code

Thariq Shihipar demonstrates how Claude Code solves tasks that chat models fail, like filtering Pokemon names ending in 'aw'. He calls this gap 'capability overhang'.

ECB's Moulin says AI risks heightening inflation volatility

ECB's Emmanuel Moulin warned that AI adoption could increase inflation volatility, complicating monetary policy. He spoke at the Paris Finance Forum.

Wealthy AI workers send San Francisco house prices soaring

A BBC analysis reports that AI industry wealth is driving a surge in San Francisco housing prices, with well-paid tech workers bidding up property values. The trend reflects the broader economic impact of the AI boom on the local real estate market.

RingCentral scales customer programs with ChatGPT Work

RingCentral's R&D Efficiency Manager used ChatGPT Work to scale from 6 pilot customers to about 80. The tool helped consolidate information across siloed teams, turning scattered knowledge into a living operating system.

SkillOpt: Agent skills as trainable parameters

SkillOpt treats agent skill files as trainable parameters outside frozen models, achieving best or tied-best performance in all 52 evaluation cells across six benchmarks and seven models without updating weights. The method keeps skills compact through bounded edits and validation gating, avoiding prompt drift. Optimized skills transfer across model scales and tasks.

Claude Fable 5 available as orchestrator model in Perplexity Computer

Yoshua Bengio argues AI may threaten humanity

Bengio warns that engineered bacteria undetectable by the human body could be created. He argues AI systems are becoming powerful enough to be a threat to life as we know it.

Runway Dev AI media platform launches for developers and enterprise teams

How agents are transforming work

OpenAI research paper details how AI agents can handle longer, more complex tasks and boost productivity across various job roles. The paper provides a framework for understanding agent capabilities and their impact on work.

U.S. to slowly distribute frontier AI models after approval

Anthropic: 65% of product team code written by Claude Tag

Claude Code v2.1.200: AskUserQuestion no longer auto-continues

AskUserQuestion dialogs now require explicit user action instead of auto-continuing; opt into idle timeout via /config. Default permission mode changed to Manual across CLI, VS Code, and JetBrains, with numerous bug fixes for background sessions and rendering.

Fable 5 tops KernelBench, Clark calls it 'start of a RSI loop'

Fable 5 wrote the fastest megakernel ever submitted to KernelBench-Mega, topping the leaderboard. Jack Clark of Anthropic described the achievement as 'the start of an RSI loop', hinting at AI automating its own R&D.

Yann LeCun calls xAI a failure, warns of AI bubble burst

Yann LeCun, former Meta AI chief, called Elon Musk's xAI 'kind of a failure' and said it won't compete with OpenAI or Anthropic. He also warned that AI labs risk a 'big bubble explosion' as costs outpace revenue.

EverOS: open-source agent memory runtime with hybrid BM25+vector retrieval

EverOS, an open-source memory runtime for AI agents, is now available under Apache 2.0. It features hybrid BM25+vector retrieval and self-evolving skills to provide long-term memory for LLMs.

UK Foreign Secretary warns of 'AI Hiroshima' without safeguards

Yvette Cooper warned that without international safeguards, frontier AI systems could lead to catastrophic outcomes, likening inaction to an 'AI Hiroshima'. She called for urgent government action to prevent AI from transforming warfare and crime.

Coinbase slashes AI bill in half, runs 1,200 agents

Coinbase reduced its AI spending by 50% while operating 1,200 AI agents. The company uses a multi-model architecture that routes work across multiple providers instead of relying on a single one.

Scientists use AI and quantum computing to generate new peptides

Researchers cobbled together funding to show how quantum computing and AI can generate novel peptides for drug development. The work targets underserved populations and rare diseases.

xAI prepares Voice Agent Builder for Grok to answer phone calls

Lilian Weng summarizes 35 papers on Harness Engineering for RSI

Lilian Weng, OpenAI's head of safety, summarizes 35 papers on Harness Engineering for Recursive Self-Improvement (RSI), covering reward engineering, oversight, and alignment. The compilation serves as a comprehensive resource for safe AI development.

Grok on Databricks

xAI's Grok models are now natively available on Databricks Agent Bricks, announced at the Databricks 2026 Data + AI Summit. The integration allows enterprises to build AI agents on their data within a governed platform, complementing existing Grok availability on Amazon Bedrock.

Databricks launches Lakebase, a fully-managed serverless Postgres database for AI apps

Lakebase is a fully-managed, serverless Postgres database designed for AI apps and agents, announced at Data + AI Summit 2026 by Nikita Shamgunov, VP of Engineering at Databricks. It aims to simplify database management for the surge of AI-driven software.

SpaceXAI, Cursor Unveil Grok AI Model for Legal, Finance Tasks

SpaceXAI and Cursor have launched Grok, an AI model specialized for legal and finance tasks. The model is accessible via a chat interface.

MyHeritage adds AI-powered photo restoration features

How to catch AI hallucinations with multi-agent checker systems

MindStudio blog explains the checker-agent pattern: multi-agent swarms where independent agents verify each output, catching hallucinations and bugs without human review. The guide covers worker-shortcut detection and boss-model bug catching.

UN presents preliminary report from scientific panel on AI governance

The Independent International Scientific Panel on AI released its first preliminary report, presented by co-chairs Yoshua Bengio and Maria Ressa alongside UN Secretary-General António Guterres at the Global Dialogue on AI Governance in Geneva. The report calls for informed global governance based on scientific evidence.

Integrates LLMs with Ghidra for binary analysis

Anthropic partners with UST to bring Claude to physical AI

UST plans to train 20,000 engineers on Claude and integrate Claude Code into chip validation and manufacturing processes. The partnership aims to catch design flaws earlier and speed up production.

Codex agents transform internal workflows at OpenAI

Perplexity Computer adds Fast mode for Claude Opus 4.8 and Seedream 5

Zhipu’s Chinese Founder Says Frontier AI Should Stay Open to All

The founder of Chinese AI lab Zhipu argues for open access to frontier AI models, advocating collaborative development over closed systems. The remarks come amid growing debate over openness in AI.

Lidl owner plans AI gigafactory as part of EU initiative

The Schwarz Group, owner of Lidl, announced plans to build one of several AI gigafactories planned by the European Union. The project aims to boost Europe's AI computing infrastructure.

Anthropic overhauls Claude Design with code round-trips and token fix

Claude Design now supports design system imports and bidirectional code round-trips to streamline handoffs. The update also fixes the tool's token-burning issue that caused high usage costs.

OpenAI more than halves inference costs with new optimization approach

Stampli scales product marketing with ChatGPT Work

One-person team now does work of four or five using ChatGPT Work to pull context from Jira, GitHub, and meeting notes and auto-update content. The tool has enabled a small team to dramatically increase output.

OpenAI engineer talks agent sandbox cloud architecture

Abhishek Bhardwaj presents architectural challenges in building a cloud for agent sandboxes. The talk covers runtime isolation trade-offs, persistence strategies, and scaling from fork() to a full fleet system.

PuMVR benchmark quantifies orthographic bias in multilingual VLMs

The PuMVR (Punjabi Multimodal Visual Reasoning) benchmark evaluates vision-language models on Punjabi text in multiple scripts, revealing that models assume a one-to-one language-to-script mapping. The findings highlight that billions of multi-script language users are overlooked by current multilingual evaluations.

TwoMillionKit lets macOS 27 users access Private Cloud Compute models without entitlement

Guilherme Rambo's TwoMillionKit app harnesses Apple's `fm` CLI in macOS 27 to run inference on Private Cloud Compute foundation models, bypassing the usual entitlement requirement. The tool is available on GitHub.

GPT-5.6 Sol, Terra, Luna spotted in Codex app code

Two Minute Papers explains terrain-diffusion AI for Minecraft worlds

The video details a diffusion model that procedurally generates Minecraft terrain. The project is available as a mod and open source on GitHub.

ACL 2026 paper: Memory OS for AI coding agents addresses context loss

GLM achieves 80% on internal financial benchmark

Vercel CEO on separating models from agents

"When optimizing for production, you start looking at price/performance," says Guillermo Rauch. The interview explores strategic trade-offs in AI deployment and the growing divide between model providers and agent frameworks.

Logan Kilpatrick: great models need high-quality curated data

Claude Fable 5 vs GPT 5.5 comparison for agentic workflows

Compares Claude Fable 5 and GPT 5.5 on benchmarks, coding, research, and real-world agentic tasks. Highlights trade-offs in instruction fidelity, tool-use accuracy, and pricing for multi-step autonomous workflows.

Zenith harness boosts models above Fable level on hardest tasks

The Anatomy of Claude Code's Memory Layer

Databricks AI Platform adds new ML engineering and inference capabilities

GPT 5.5 launches on Cerebras via OpenRouter

GPT 5.5 is now available on Cerebras through OpenRouter, as spotted by users. The model appeared today and can be accessed by selecting the Cerebras provider on OpenRouter.

Hunyuan3D port runs on Apple Silicon with local image-to-3D under 20 seconds

A Swift/MLX port of Hunyuan3D-Shape and Hunyuan3D-Paint enables local image-to-3D on Apple devices. Benchmarks on M4 Max: shape model in ~21s at 5.6GB RAM; paint model in ~231s at 38GB RAM.

Grok Build CLI adds voice dictation via /voice command

xAI releases 21 new flagship Grok voices

21 new multilingual voices join the original five, available now in the Voice Agent API, Text to Speech API, and Voice Agent Builder. The original five voices also received quality upgrades for more natural pacing and emphasis.

NYT: Anthropic employees feel unfairly targeted by US government

Claude Code gains cross-session photographic memory

Anthropic not worried about losing customers, expert says

Sam Altman calls out Elon Musk's space data center

Sam Altman criticized Elon Musk's proposal for a space-based data center, intensifying the rivalry in AI infrastructure. The exchange highlights the growing competition between OpenAI and xAI for compute resources.

Guide to loop engineering: How autoresearch turns AI agents into autonomous research loops

The guide details two artifacts: Andrej Karpathy's autoresearch repository and the Bilevel Autoresearch method, which together enable AI agents to autonomously conduct machine learning research. It explains how these tools transform the typical manual query-response pattern into a continuous research loop.

AI reverse engineering plugin integrates LLMs into IDA Pro

Genie ZeroOps AI agent launches for production workload monitoring

Explainer: Why Anthropic’s Mythos AI was blocked

Bloomberg's explainer covers the security risks and advanced capabilities of Anthropic's Mythos AI that led to its blocking. The article details the concerns that caused the project to be restricted.

Odyssey raises $310M at $1.45B valuation backed by Amazon

World model startup Odyssey raises $310M Series B at $1.45B valuation led by Natural Capital, with Amazon, AMD Ventures, GV participating. Founded by self-driving veterans, Odyssey will use AWS as preferred cloud and optimize models for Trainium chips.

AI Poses Biggest Security Challenge of Decade, UK's Cooper Warns

UK Home Secretary Yvette Cooper said AI presents the most significant security threat of the 2020s, urging international coordination. She emphasized risks from state-backed disinformation and autonomous systems.

Claude Code runs entirely on-device on Apple Silicon

Chollet: Scalar benchmark scores are meaningless, efficiency scores needed

Semantic Blindness: 500K Sensors Confused an LLM

Raahul Singh and Vanč Levstik discuss how their work with 500,000 sensors revealed LLMs' inability to handle combinatorial engineering problems. They argue that next-token prediction engines fail in mission-critical physical infrastructure.

Apple's failed self-driving car program left a legacy of powerful AI chips

Apple's self-driving car project, though ultimately abandoned, drove the development of powerful on-device AI processing chips. These chips later found their way into Apple Silicon, powering AI features across Macs and other devices.

Codex in ChatGPT mobile app now generally available

Genie App Builder for AI-assisted app creation with Databricks context

OpenAI improves health intelligence in ChatGPT with GPT-5.5 Instant

GPT-5.5 Instant enhances health and wellness responses with stronger reasoning and better context. The update includes physician-informed evaluations for clearer communication.

OpenAI reduces p95 latency by 25% across Realtime voice models

Claude Design gains bidirectional sync with Claude Code

The update enables two-way sync between design and code, letting designers edit on canvas and stay on brand with their design system. It also connects to more tools and allows importing design systems from repos or codebases.

Zvi Mowshowitz unpacks Anthropic's Fable system card

Discussion covers Fable's FrontierMath leap and troubling Vending-Bench behavior, signs of decision-theory drift, and reasoning becoming harder to read. Also touches on US export controls debate.

OpenAI and Databricks at DAIS 2026: Making enterprise AI real

Databricks' Data + AI Summit 2026 attracted over 32,000 attendees and featured product announcements from OpenAI and Databricks. The event highlighted new integrations to simplify enterprise AI adoption.

Databricks Free Edition adds Agent Bricks, Genie Code, and serverless GPUs

Severe vulnerabilities spiked 3.5x after Claude Mythos Preview

According to Epoch AI data, high- and critical-severity CVEs increased more than 3.5x in June compared to previous monthly record, following Anthropic's April announcement that Claude Mythos Preview could autonomously find and exploit vulnerabilities. Both Anthropic and OpenAI have since launched efforts to use frontier models to harden critical software.

Ethan Mollick: Chatbot era is over, agentic systems coming

Yoshua Bengio urges governments to act on AI risks

Speaking at a United Nations General Assembly dialogue, Bengio emphasized that AI is advancing faster than societies can adapt. He urged governments to rely on independent scientific assessments rather than industry lobbying.

Genie Ontology automatic context layer announced at DataAISummit

Ali Ghodsi on scaling AI agents in enterprise

Databricks CEO Ali Ghodsi outlines four pillars for scaling enterprise agents: context, control, choice, and cost management. Talk delivered at Data + AI Summit 2026.

Grounded Reasoning Cup live at DataAISummit 2026