Around the Horn Digest: Everything That Happened in AI Today (Thursday, July 2, 2026)

Anthropic got Fable 5 back online, Cursor said it topped its coding-agent benchmark, and half the internet immediately started arguing whether anyone could actually feel the difference.

Fable 5 had the most suspiciously dramatic product arc of the day: launch, shutdown, government review, relaunch, benchmark win, developer hype cycle, developer anti-hype cycle. Meanwhile, the rest of AI kept sprinting in the same direction. AWS embedded engineers with customers, Meta tried to turn excess compute into a cloud business, Cognition turned security scanning into an agent swarm, xAI packaged Grok Voice into call-center software, NVIDIA pushed both faster language models and physical AI, and every benchmark suddenly wanted to prove agents can do real work. Let's get into it.

Around the Horn - Thursday, July 2, 2026

Anthropic restored Claude Fable 5 after one of the strangest frontier-model rollouts yet. The model launched as Fable 5 and Mythos 5 on June 9, went dark globally June 12 after a U.S. export-control directive tied to an Amazon-reported cybersecurity safeguard bypass, then returned July 1 with new classifier safeguards and a proposed framework for scoring jailbreak severity.

The story is not only that Fable 5 came back. It is that a frontier model can now be switched off for everyone because a government directive, a cloud partner report, or a safeguard failure changes the risk calculus overnight. Fable is broadly restored, but Mythos 5 remains limited to approved U.S.-based organizations, and AP connected the episode to OpenAI's temporary GPT-5.6 Sol restriction for approved customers. The Financial Times also framed the White House move as the restriction lift that let Anthropic re-release the models.

The practical takeaway: model roadmaps are becoming policy roadmaps. Companies buying frontier AI are not only choosing capability, price, and latency anymore. They are choosing exposure to regulation, safety gates, export controls, and emergency shutdowns.

🏆 TOP 5 NEWS (Around the Horn)

• CAIS and Scale AI Labs released updated Remote Labor Index results showing Fable 5 leading public models at 16.1% on 240 real remote-work projects across 23 domains. Chubby's summary framed it as a major jump because the benchmark asks whether a client would accept the AI's work as a real deliverable.
• Senior SWE-Bench, introduced by Henry Ehrenberg, launched as an open-source benchmark for testing coding agents on long-horizon, realistically under-specified senior engineering work instead of intern-style tasks with tidy instructions.
• Cognition launched Devin Security Swarm, a new Devin for Security system that scans large codebases, validates exploitability in sandboxes, and opens remediation PRs using an Agentic MapReduce architecture. Cognition's own launch post framed it as a way to find security bugs across entire codebases at lower cost than prior agentic scans.
• AWS created a $1B Forward Deployed Engineering organization that embeds AWS frontier teams directly with customers to build production agentic AI systems.
• NVIDIA released Nemotron-Labs-TwoTower, a diffusion language model that reports 98.7% quality at 2.42x generation throughput, and also introduced Generative Pretrained Controllers for transferable physical-AI motor control.

Honorable Mentions

• xAI launched Voice Agent Builder, a no-code Grok Voice product for building human-like phone agents at $0.05 per minute.
• Meta popped 9% after investors cheered its push to sell excess AI compute capacity through a new cloud business, a possible way to justify its huge infrastructure spending.
• Ramp Labs introduced PorTAL, a way to port learned task behavior across new base models so fine-tunes do not become obsolete every time the model layer moves. Rahul connected the point to the model-progress tradeoff: custom fine-tuning is partly a bet that a good enough base model will not arrive soon.
• Together AI raised $800M at an $8.3B valuation to scale open-model and neocloud infrastructure, with TechCrunch noting the company was valued at $3.3B in early 2025 and The New York Times framing it around cheaper AI options.
• Yoshua Bengio and Maria Ressa warned that the world has a narrow window to coordinate on AI risks before concentrated power, failing guardrails, and threats to shared reality get harder to contain.

🍪 TOP TREATS TO TRY

• Tabstack lets you give your app or agent browser tasks like clicking, filling forms, scraping pages, and pausing for human input without hosting browser infrastructure yourself. No pricing details.
• Adam CAD Copilot lets you edit Onshape and Autodesk Fusion parts with prompts, selected geometry, and feature-tree context while keeping the CAD model parametric and editable. Free tier included.
• Sequence Agentic lets you connect AI agents to real U.S. bank accounts with permissions, limits, and audit trails so they can move money safely inside workflows. No pricing details.
• Mark by Airtop researches your business, builds a GTM plan, and turns lead gen, enrichment, outbound, SEO, and Google Ads tasks into web agents you can run. Free to start.
• RunInfra turns a plain-English model workload into an optimized production inference endpoint by benchmarking GPUs, quantizing models, and generating custom kernels. Pay per million tokens.
• Stigg 2.0 lets you enforce credits, usage limits, entitlements, budgets, and governance on every AI request before spend runs away. Free forever for AI startups.
• N71 gives your agents one shared knowledge graph across tools like Notion, mail, calendar, docs, chat, and repos so they stop re-learning context from scratch. Product Hunt launch code offers 2 months off Pro.
• Gemini Omni Flash lets you generate and edit videos through Google's API with conversational prompts, reference images, uploaded clips, and multi-turn revisions. No pricing details.
• xAI Voice Agent Builder lets you build no-code Grok Voice agents for support, sales, scheduling, and workflow handoffs with playbooks, knowledge bases, 80+ voices, voice cloning, guardrails, call replay, a free phone number, or SIP transfer. It costs $0.05/minute.
• ZCode, announced by Z.ai, is the official cross-platform GLM-5.2 coding harness with BYOK support, a 1.5x GLM Coding Plan quota boost, long-running Goals, 20+ tools, multi-agent collaboration, and remote control through WeChat, Feishu, or Telegram. Paid plans start around $16/month.
• Claude Code 2.1.198 adds Claude in Chrome general availability, background agents that can auto-commit, push, and open draft PRs, a new /dataviz skill, better grep guidance, and stability fixes. Paid Claude access required.
• Devin for Security scans large codebases for vulnerabilities, validates exploitability in sandboxes, and opens remediation PRs using Cognition's Agentic MapReduce system. No public pricing details.
• Parsewise gives you an API for reasoning across large mixed document sets and returning validated CSV or JSON with citations, lineage, and uncertainty flags. No pricing details.
• DoorDash's agentic-orchestrator gives you an open-source Go CLI and TUI for coordinating AI agents across planning, research, implementation, review, and linked PRs across one or more repos. Free/open source.
• Fireworks AI now hosts GLM-5.2, a 743B-parameter MoE model with 1M-token context, multi-effort coding, and efficiency improvements for long-horizon agents and enterprise RAG. No pricing details.

🏢 Big Tech & Major Companies

• Claude's July 1 thread said Fable 5 returned with updated cybersecurity safeguards after conversations with the U.S. government. Follow-up posts explained the paid-plan promotional access, pointed users to the formal support article, and asked users to share false-positive feedback through Claude Code and claude.ai feedback paths.
• Claude's promo-access article says Fable 5 is available at no extra cost from July 1-7 for Pro, Max, Team, and premium Enterprise seat users, but it draws from existing weekly limits and can consume them faster.
• Axios covered Anthropic's Sonnet 5 positioning as the everyday lower-cyber-risk agent model for browser use, planning, coding, and knowledge work compared with Mythos and Fable.
• The Information reported that Anthropic backtracked on a controversial Claude Code location-tracking feature meant to identify users in China or affiliated with Chinese AI labs.
• ClaudeDevs said Claude reset five-hour and weekly limits after Fable 5 returned, which turned access mechanics into part of the story as users debated the July 7 promo window.
• Cursor said Claude Fable 5 is available again inside Cursor, where it leads all models on CursorBench but is also the most expensive model per task.
• Perplexity said Claude Fable 5 is available again inside Perplexity Computer as an orchestrator model for complex multi-step workflows.
• Meta is reportedly exploring a cloud business that would sell excess AI compute and possibly hosted model access, with CNBC reporting the stock jump as investors looked for a return on AI infrastructure spending.
• Meta One is Meta's limited-testing subscription for AI glasses, with availability and plan options varying by location, account type, and current tests.
• Google pushed production agents toward app plumbing with Genkit Agents, ADK 2.0 workflows, and a Google Cloud Workbench VS Code extension.
• Google also announced Gemini Spark for the Gemini app on macOS in beta for U.S. Google AI Ultra subscribers, alongside custom MCP support and new app integrations with services like Canva, Dropbox, Instacart, OpenTable, Zillow Rentals, Tasks, and Keep.
• Google open-sourced Zero-Knowledge Proof libraries for age assurance so services can verify attributes like age without exposing extra personal data.
• Google's June AI roundup collected the company's model, product, and initiative updates from June.
• GitHub made Copilot's browser tools generally available in VS Code and added Copilot CLI auto model selection based on task type, reliability, and cost signals.
• Cisco is rolling out AI agents to all 90,000 employees, with finance already using agents for first-draft MD&A sections, investor-relations tools, benchmarking, and a CFO dashboard.
• Reuters reported that Elon Musk denied a Wall Street Journal story saying SpaceX showed investors a handset-like AI device prototype before its IPO, while TechCrunch connected the reported device to SpaceX's wireless ambitions.

💼 AI Productivity, Labor & Economics

• Thinking Machines said Bridgewater used expert-labeled data and on-policy distillation to fine-tune a model that sorts financial documents worth an analyst's attention, reaching 84.7% accuracy, 29.8% fewer mistakes than the best frontier models, and 13.8x lower inference cost. The full Thinking Machines Lab article framed it as a case where expert data still beats a frontier base model on nuanced professional judgment.
• CNBC reported that some employers that cut workers citing AI are already reversing course and rehiring after deployments struggled with edge cases, inconsistent outputs, quality problems, and tasks needing human judgment.
• Top economists warned that an AI boom-bust cycle could create global fallout through hyperscaler debt, leveraged investor bets, and unemployment shocks.
• Gene Sperling argued for investing AI productivity gains into "double-dignity jobs" in care, health, education, counseling, and navigation rather than accepting a future with less work and less contribution.
• A California lawsuit accused gas station operators and pricing software company Kalibrate of using AI-powered tools to coordinate gasoline prices, testing the state's updated antitrust law against algorithmic price-fixing.
• Palantir CEO Alex Karp called the AI industry "effing insane" in a heated interview, criticizing high fees, data extraction, and the idea of outsourcing U.S. military AI to Silicon Valley consensus.
• Xiaoyin Qu argued that open-weight Chinese models have compressed profits in the model layer, leaving compute, energy, and applications as the better AI business layers, while also arguing OpenAI is better positioned than Anthropic because of Codex, consumer distribution, open-source integrations, ads, consulting, robotics, compute, and execution speed.
• Xiaoyin Qu separately argued that enterprises rejecting cheaper Chinese open models on vague security grounds while sending data to U.S. frontier labs are missing the cost and privacy advantages of self-hosting.
• Thorsten Ball, Mikhail Parakhin, Ethan Mollick, Theo, and elvis all circled the same model-routing problem: judging whether a task is simple enough for a cheaper model often requires doing the work first.
• Theo also kept the cost-performance debate going by comparing how model choice, token use, and agentic workflow shape the real price of coding work.
• Francois Chollet argued that the current AI wave will not create mass unemployment and may increase demand for software engineers.
• Siddarth Pai said Claude Code outperformed Codex across infrastructure, AI/ML, backend, and UI work based on his hands-on use of all subscription tiers.
• Greg Isenberg, MTS, Nathan Lambert, Nicolas Bustamante, and Lee Robinson added shorter developer and operator reactions around the same practical question: whether better models, better routing, or better product harnesses are doing the actual work.

🧑‍💻 Agents, Coding & Developer Tools

• Senior SWE-Bench reframes coding-agent evaluation around senior work: vague specs, long-horizon execution, taste, validation, and realistic ambiguity.
• Cognition's Agentic MapReduce uses deterministic selectors, parallel bounded-shard agents, and a reducer to reason across whole codebases without letting one agent wander through too much context.
• Kasra joined OpenAI to work on Codex and said the coding-agent GUI was the first one that pulled him out of the terminal.
• Matt Pocock shared a planning skill for large greenfield projects that maps decision frontiers, research needs, and parallelizable implementation scopes.
• Daniel Miessler shared 10 Fable prompts for harness optimization, security posture, eval design, autonomy ladders, and personal/work prioritization.
• Don't Train the Model, Evolve the Harness is a Hugging Face Space focused on improving model performance by optimizing the surrounding harness instead of retraining the model.
• Victor Taelin said Fable 5 handled a complex task well enough to catch subtle issues that he and other models missed, adding to the day's strongest first-hand developer reactions.
• Gavriel Cohen shared an overnight Fable 5 session that moved a stale private fork of a PR Factory-style agent system most of the way toward being open-sourceable, with the human mostly making taste and scope calls.
• Stephen Haney added another field report from the Fable 5 developer rush, contributing to the day's mixed picture of benchmark wins, impressive agentic sessions, and uneven user perception.
• Hacker News users debated Fable 5's return, praising its refactoring and agentic strengths while complaining about guardrails, usage limits, high cost versus Opus 4.8, and trust questions from the model's shutdown.
• atomic.chat ran a coding contest where Fable 5 produced the best HTML5 canvas physics demos but cost roughly 6x more than Opus 4.8, while GLM-5.2 delivered competitive output at far lower cost.
• Hanami 3.0 launched with mailers, i18n, Minitest support, and performance gains, while the Hacker News thread framed its appeal as intentional architecture and taste versus Rails-style convention.
• Box3D launched as an open-source 3D physics engine forked from Box2D with mesh, height-field, compound collision, continuous collision, SIMD, determinism, large-world support, recording, and replay. The Hacker News thread praised it as a compact deterministic option for games, VR, networked simulations, and reinforcement learning.

📊 Fundraising & Deals Roundup

• Venice raised a $65M Series A led by Dragonfly at a $1B valuation to scale its privacy-first, unrestricted AI platform, with TechCrunch reporting more than $70M in annualized revenue and 3.5M registered users.
• Abu Dhabi's MGX raised $49B for a new AI fund, with Bloomberg calling it one of the largest AI funds ever.
• Higharc raised a $95M Series C to scale AI for the homebuilding design-to-construction lifecycle, while the Triangle Business Journal reported the Durham company has more than 200 employees and plans to hire.
• Aligned raised $60M for B2B sales software and may seek a Series B in the next 12-plus months.
• Oxmiq raised $35M to build chip architecture and software aimed at lowering AI application costs.
• Twelve Labs raised $100M from Amazon, NEA, and Naver Ventures to advance AI search and analysis across large video libraries.
• Uber reshuffled its data-labeling business after two senior executives departed from the Uber AI Solutions and AI Data Infrastructure unit.

🔬 Research, Models & Benchmarks

• Nemotron-Labs-TwoTower-30B-A3B-Base-BF16 is available on Hugging Face and supports mask diffusion, mock autoregressive mode, and standard autoregressive mode. The paper decouples context and denoising into separate towers.
• Generative Pretrained Controllers tokenize motor skills and train a GPT-style controller on 600+ hours of motion data, with NVIDIA reporting 99.98% success reproducing motion clips and emergent recovery behaviors. NVIDIA is also pointing the robotics and physical-AI crowd to its SIGGRAPH program.
• Learning-to-Theorize and the Neural Theorizer learn executable, compositional theories from before-and-after observations instead of only predicting next states. The project page expands on the Observation-to-Theory Induction Benchmark, and Sungjin Ahn framed the ICML 2026 Oral work as a shift from prediction toward reusable theory-building.
• Baseten researchers explained neural KV cache compaction for long-horizon agents, focusing on near-lossless relevant-context retrieval and durable learned memory. Charlie O'Neill described the same work as a path toward billion-token effective context windows through compaction.
• SkillComposer treats agent skills as a structured planning problem by deciding which skills to activate, how many to use, and in what order.
• Hugging Face highlighted metacognition adapters that read model hidden states to estimate whether an answer is likely wrong while keeping the base model frozen.
• Rohan Paul shared a Meta paper arguing that post-training quantization can make reasoning models second-guess correct answers, while penalizing 50 hesitation tokens cut reasoning length 12-23% without hurting accuracy.
• Google Gemma highlighted a Gemma and Xenova kernel-optimization demo, adding another practical example of agent-written performance work running locally in browser.
• Mia said she would test NVIDIA's Nemotron TwoTower variant against Qwen 3.6 35B in agentic workflows.
• Quanta Magazine reported that University of Minnesota researchers built a synthetic cell from nonliving biomolecules in a lipid membrane that can grow, replicate DNA, and divide, though it still needs external inputs and remains far from autonomous life.

🛡️ Security, Governance & Trust

• Cognition's Security Swarm eval tested 50 real GHSA vulnerabilities across 14 languages and reported 72% recall at $90 per run, including critical vulnerabilities other tools missed.
• Factory AI introduced Droid Shield 2.0, a learned secret-detection system for safer autonomous engineering agents. The methodology writeup says the fine-tuned detectors catch missed secrets and reduce false positives across repository-level evals.
• Eugene Yan's cybersecurity-evals writeup, highlighted by Cameron Wolfe, explained why realistic cybersecurity agent evals can run for 24+ hours, cost around $40K in API credits, and need partial-credit scoring for steps like finding, reproducing, and exploiting a vulnerability.
• Yoshua Bengio and UN ODET framed the UN AI Panel report as a turning point where government choices will shape whether AI's benefits or harms dominate.
• Reuters reported that the UN independent scientific panel's first AI report warned of benefits and risks across mental health, deceptive behavior, information integrity, human rights, and control of autonomous systems.
• Axios reported that the UN launched an AI for Good Global Commission co-chaired by Salesforce CEO Marc Benioff and Rwanda's Paul Kagame, bringing tech CEOs and political leaders together to discuss global rules, infrastructure, and responsible AI deployment.
• Andrew Curran said the White House is accelerating voluntary AI model standards and benchmarks with Anthropic, CAISI, and NSA involvement around cyber-capable models, frontier definitions, release timelines, and access rules.
• Mistral added connector controls for enterprise agents, including scoped API keys, multi-account connectors, admin controls, a connector debugger, and connector reuse in Vibe Code and Workflows.

🌐 AI Infrastructure, Web & Search

• Cloudflare launched efforts to make AI search smarter and compensate creators, pairing network signals for better answers with a move from Pay Per Crawl toward Pay Per Use when AI systems use content in generated answers.
• ByteDance picked Brazil for its largest data center outside China, a reported $39B project tied to the country's push for Chinese investment.
• Bloomberg reported that Amazon's new transatlantic cable underscores Ireland's role as an AI hub while exposing security risks tied to weak defense spending.

🎙️ Voice, Robotics & Creative Tools

• Weave Robotics launched Isaac 1, a home robot focused on household chores like folding laundry, with deliveries planned for this fall. The official Isaac 1 page describes a privacy-first mobile robot with a collapsible torso, swappable fabric shells, and $250 preorder deposit, $449/month subscription, or $7,999 upfront purchase.
• UBTECH launched UWORLD U1, billed as the world's first full-size mass-produced ultra-bionic humanoid robot, at its 2026 Global Launch Event in Shenzhen.
• Hugging Face and Cerebras showed an open, modular speech-to-speech stack using Nvidia Parakeet, Google DeepMind's Gemma 4 31B on Cerebras, and Alibaba's Qwen3TTS.
• Fish Audio announced a free text-to-speech option.
• fal launched a TRELLIS.2 LoRA Trainer for turning a folder of 3D assets into style-matched GLB generations.
• Paper Shaders went fully open source and removed prior reselling restrictions for plugins, templates, sites, apps, libraries, and Figma tools.
• Acti was Product Hunt's top July 1 launch: an agentic mobile keyboard that fetches links, docs, schedules, profiles, and meeting links from inside a text field.
• Mistral OCR 4 added bounding boxes, block classification, confidence scores, 170-language support, self-hosting, and API pricing at $4 per 1,000 pages, or $2 through batch API.
• Microsoft's AI For Good Lab built AI features for the Theodore Roosevelt Presidential Library, including a lifelike Roosevelt avatar and a Campfire Reading Room search tool over digitized documents. The Hill reported that President Trump interacted with the AI Roosevelt during the library opening events.

🧠 Intelligent Insights

• Ethan Mollick argued that working with Fable feels less like step-by-step collaboration and more like commissioning work from an autonomous system: you give ambitious instructions, wait, then judge the result.
• Mollick also argued that AI work is moving from chatbots toward autonomous agents, making experts more like AI managers who delegate through harnesses and tools rather than prompt every step.
• Steve Krouse said that after using Fable nonstop and returning to Opus, he did not notice a meaningful difference in agentic coding, then ran a blind-style test asking users to spot Fable among Claude models.
• Aniket Panjwani recommended saving the short Fable 5 access window for planning, project review, and hardest-problem diagnosis, then delegating implementation to cheaper models or Codex-style tools.
• Avi Chawla and Akshay Pachaar argued that many RAG failures come from treating raw text chunks as retrieval units, pointing instead to structured IdeaBlocks with source, version, and governance metadata.
• Pratyush Choudhury argued that high-return AI biology work may sit downstream of target discovery, including toxicology, preclinical translation, trial design, regulatory evidence, and closed-loop wet-lab data.
• Gergely Orosz and dax turned Spotify engineers' public Claude usage into a broader point about AI bragging colliding with product complaints from engineers who are also paying customers.
• Katie Parrott used frontier models to test whether they could design a coherent visualization of the pre-exilic kingdoms of Israel and Judah.
• eglyman added a Ramp-side thread to the day's fine-tuning and task-adapter discussion, reinforcing the argument that portable task representations matter when the base-model layer keeps changing.

Previous Around the Horn Digests

Catch up on everything you missed:

• Tuesday, June 30, 2026: Anthropic launched Claude Sonnet 5 and Claude Science, AWS built a $1B forward-deployed AI push, and Etched hit $1B in AI chip sales.
• Monday, June 29, 2026: AI pressure hit billable hours, data centers, chip policy, government adoption, elections, jobs, and coding agents.
• Monday, June 22, 2026: Sakana launched Fugu, OpenAI expanded Daybreak, AI infrastructure debt accelerated, and Getty struck an OpenAI display deal.
• Friday, June 19, 2026: OpenAI helped solve rare pediatric disease cases, Z.ai's GLM-5.2 shook up open models, and Amazon aimed Trainium at Nvidia.
• Thursday, June 18, 2026: Noam Shazeer left Google for OpenAI, OpenAI pushed deeper into life sciences, and Anthropic's access fight became geopolitical.
• Tuesday, June 16, 2026: SpaceX reportedly pushed deeper into AI coding, CoreWeave trained DeepSeek-V3 in two minutes, and OpenAI tested deployment simulation.
• Monday, June 15, 2026: Anthropic's Fable and Mythos fight spilled into cyber policy, Salesforce agreed to buy Fin, and Qualcomm eyed Tenstorrent.

That's a wrap! See you on the next one.