Anthropic spent the day everywhere: shipping a new Claude model, launching a science workbench, and getting one of its restricted models cleared by Commerce. Meanwhile, OpenAI talked costs, Google pushed AI images into Gemini, and everyone else kept building tools for the agent era.
Welcome to the Around the Horn Digest, where we collect the day’s AI moves, sort the signal from the noise, and give you the clean version you wish your tabs had organized themselves into.
Around the Horn — Tuesday, June 30, 2026
Anthropic launched Claude Sonnet 5, its new default model for Free and Pro users, positioning it as a stronger everyday model for coding, research, and long-running agentic tasks.
The release also came with fresh attention from the U.S. government. Commerce Secretary Howard Lutnick said the Commerce Department worked with Anthropic over the past two weeks to analyze and approve Fable 5, one of Anthropic’s previously restricted models, with access expected to return after export-control review.
Anthropic also launched Claude Science, a beta research workbench for scientists that traces artifacts back to their code, manages compute environments on demand, and connects to more than 60 optional scientific databases across biology, chemistry, and related research workflows.
🏆 TOP 5 NEWS (Around the Horn)
- OpenAI reportedly found inference optimizations that more than halved the cost of running its models, with engineers saying some free ChatGPT traffic was served using only a few hundred Nvidia GPUs at one point.
- Amazon Web Services launched a $1B Forward Deployed Engineering group to embed AI engineers with customers and help companies build production agentic systems faster.
- Meituan open-sourced LongCat-2.0, a 1.6T-parameter MoE model trained on domestic Chinese AI chips with strong coding and agentic benchmarks.
- Etched exited stealth at a $5B valuation with $800M raised and $1B in signed contracts for its Sohu AI inference chip.
- Google expanded personalized image generation in Gemini to more U.S. users, letting people create customized AI images inside the Gemini app.
Advertisement
Honorable Mentions
- Schneider Electric agreed to buy industrial AI company Cognite for $3.1B.
- Qwen released Qwen-AgentWorld, an open-source training and evaluation environment for agents.
- Omen raised $35M to build an autonomous intelligence platform for defense and national security work.
- Dominion Dynamics raised $20M to build low-altitude autonomous drone systems for U.S. national security applications.
🍪 TOP TREATS TO TRY
- Claude Sonnet 5 is Anthropic’s new default Claude model for Free and Pro users, built for stronger everyday coding, research, and agentic work.
- Claude Science gives researchers a beta workbench with code-traced artifacts, on-demand environments, and more than 60 scientific database connectors.
- Gemini image generation now gives more U.S. users personalized AI image tools inside Gemini, free to try.
- Gemini API’s new image and video tools add Nano Banana 2 Lite for fast image generation and Gemini Omni Flash for video editing.
- Klaviyo’s new campaign and customer-service agents help marketers build launch-ready campaigns and handle orders, returns, and loyalty lookups from the same customer data.
- Qwen-AgentWorld helps developers train and test agents in simulated environments like web browsing, Android, terminal work, search, and software engineering.
- Ornith-1.0 gives developers open-source coding models that can create both solutions and their own test harnesses.
- LongCat-2.0 gives developers access to a 1.6T-parameter open MoE model optimized for long-context coding and agentic tasks.
🏢 Big Tech & Major Companies
- Anthropic launched Claude Sonnet 5 as its new default model for Free and Pro users, emphasizing stronger performance on coding, research, and longer agentic workflows.
- Anthropic introduced Claude Science, a beta research app for scientists that traces artifacts to their code, provisions environments on demand, and connects to more than 60 scientific databases.
- AWS launched a $1B Forward Deployed Engineering organization to put AI engineers directly inside customer deployments.
- TechCrunch reported that AWS’s new forward-deployed AI group is meant to speed enterprise adoption by pairing engineers with customers on production AI systems.
- Google expanded Gemini’s personalized image-generation features to more U.S. users.
- Google Developers highlighted new Gemini API media capabilities, including Nano Banana 2 Lite for fast, low-cost image generation and Gemini Omni Flash for video editing.
- OpenAI’s Codex team is reportedly experimenting with hardware-device workflows and tighter development loops around coding agents.
Advertisement
🤖 Models, Agents & Research
- Meituan open-sourced LongCat-2.0, a 1.6T-parameter MoE model with roughly 48B active parameters per token, 1M context, and strong coding and agent benchmarks.
- Reuters reported that LongCat-2.0 was trained on domestic Chinese AI chips, making the release important for both model performance and China’s AI hardware independence.
- Qwen-AgentWorld launched as an open-source environment for training and testing agents across web browsing, Android, terminal work, search, software engineering, and other simulated tasks.
- Qwen’s public repo and Hugging Face materials positioned AgentWorld as a tool for repeatable agent training and evaluation.
- Researchers proposed Tapered LMs, a method for dynamically reducing model width during inference to improve efficiency.
- Researchers introduced iLLaDA, a linear diffusion language model that aims to improve long-context generation and inference efficiency.
- Researchers introduced Autodata, a framework for agentic data generation and evaluation.
- BenchPress added another benchmark for measuring software-engineering agents against realistic coding workflows.
- DeepSeek’s open infrastructure projects, including DeepSpec, DeepEP, DeepGEMM, and SGLang’s DSpark, drew attention for showing how low-level systems work can reduce latency and improve throughput for large models.
🧱 AI Infrastructure, Chips & Data Centers
- Etched exited stealth with a $5B valuation, $800M raised, and $1B in signed contracts for its Sohu inference chip.
- Etched said Sohu is designed specifically for transformer inference, arguing that dedicated silicon can beat general-purpose GPUs on cost and performance for AI serving.
- Omen raised $35M to build autonomous intelligence infrastructure for defense and national security customers.
- Omen’s launch materials positioned the company as building AI systems for faster intelligence workflows.
- IREN announced a major AI cloud services arrangement tied to Nvidia GPUs.
- Reports on GPU pricing suggested that competition for AI infrastructure remains intense even as labs chase lower inference costs.
- Perplexity’s infrastructure-related coverage underscored how model access, serving costs, and distribution are becoming core AI product constraints.
Advertisement
💼 AI Productivity, Labor & Economics
- Ethan Mollick argued that AI is moving from chatbots toward long-running agents that complete work with less supervision.
- Ramp and Revelio Labs found that companies spending heavily on AI have also shown faster headcount growth, complicating the simple “AI kills jobs” narrative.
- TechCrunch reported on the AI labor market split, where companies are cutting some roles while aggressively hiring for AI work.
- Grindr’s CEO discussed how AI could reshape product development and user experience in consumer apps.
- Algorithmic Bridge argued that workers increasingly need AI fluency regardless of whether they personally believe AI is overhyped.
- TechBrew and The New York Times covered how $180K tech salaries can still feel stretched in San Francisco’s AI-inflated housing market.
- Decoder’s San Francisco coverage highlighted how AI wealth is reshaping the city’s housing, startups, and labor market.
- The American Bazaar also covered the tension between high tech pay and rising Bay Area living costs.
🏛️ Policy, Defense & Government
- Federal News Network reported that the Pentagon is recruiting more technical talent to embed across the armed forces.
- Bloomberg also covered the military’s push to bring AI and tech workers deeper into defense operations.
- Dominion Dynamics raised $20M to develop low-altitude autonomous drone systems for national security applications.
- The New York Times and Decoder covered how AI is increasingly showing up in campaigns, policy work, and public-sector decision-making.
- The Financial Times reported on AI-related modernization needs in air traffic and aviation systems.
- Perplexity’s defense and policy slate pointed to continued government demand for AI systems, especially in intelligence and operational planning.
Advertisement
📊 Fundraising & Deals Roundup
- Schneider Electric agreed to buy Cognite for $3.1B, giving the industrial giant a stronger AI software layer for manufacturing and energy customers.
- Axios and Bloomberg both covered the Schneider-Cognite deal as part of a broader industrial AI consolidation wave.
- Omen raised $35M for autonomous intelligence infrastructure.
- Dominion Dynamics raised $20M for national-security drone systems.
- MDOTM raised new funding for AI-driven investment technology.
- Aikido Security acquired Root to help teams patch vulnerable open-source software without forcing major version upgrades.
- 8090 Industries raised new capital for hard-tech and AI infrastructure investments.
- 1001 raised funding for AI-native enterprise operations.
- Axios Pro Rata tracked the broader AI funding environment, where large rounds remain concentrated around infrastructure, security, and vertical AI.
🧪 Science, Software & Vertical AI
- Claude Science launched as Anthropic’s research workbench for scientific workflows, with reproducible artifacts, code history, environment management, and database connectors.
- Ornith-1.0 released open-source coding models designed to generate both solutions and their own test harnesses.
- Qwen-AgentWorld gave researchers and developers a new environment for evaluating agents in simulated real-world tasks.
- Klaviyo launched paired AI agents for campaign creation and customer service.
- Digital Trends covered how AI-generated and AI-assisted games are pushing into consumer entertainment.
- Perplexity’s vertical AI slate highlighted continued momentum in specialized tools for law, finance, healthcare, and enterprise operations.
Advertisement
Late Additions From the Afternoon Research Pass
The later research sweep turned up several additional stories worth folding into the Tuesday digest. The strongest additions were mostly benchmarks, agent infrastructure, model-serving updates, and tools that showed how quickly the market is moving from chatbot demos to production workflows.
🏆 Strongest New Items
- OpenAI introduced GeneBench-Pro, a benchmark for AI agents doing messy computational biology and genomics work where expert humans usually spend 20-40 hours per problem; GPT-5.6 Sol reached a 28.7% pass rate, and OpenAI is open-sourcing a public subset.
- Google Research introduced TabFM, a foundation model for tabular data that can do zero-shot classification and regression on unseen tables without per-table training, hyperparameter tuning, or manual feature engineering.
- Thinking Machines Lab and Bridgewater AIA Labs showed that fine-tuning on expert investor annotations produced a smaller financial triage model that beat frontier models on six real judgment-heavy tasks while running much cheaper.
- Meituan open-sourced LongCat-2.0, a 1.6T-parameter MoE model trained on AI ASIC superpods with 1M context and strong coding and agent benchmarks, including Terminal-Bench 70.8 and SWE-bench Pro 59.5.
- NVIDIA GEAR introduced ASPIRE, a robot-learning system that builds a self-evolving library of reusable sensorimotor skills from simulation and real-world traces, improving transfer across tasks and robot bodies.
🍪 Top Treats / Tools
- Browserbase Agents lets developers ship a browser agent from one prompt and one API call, packaging browser automation into a production-ready harness for agent workflows.
- June is a private, local-first Mac AI app with chat, voice dictation, bot-free meeting notes, and local agent workflows, built on the open-source Hermes framework.
- Spellbook launched Autonomous Contract Management, an AI-native contract workflow that handles intake, review, negotiation support, storage, renewals, and risk monitoring from one system.
- world-model-harness turns agent execution traces into fast simulations of production environments by having an LLM act like a Docker container, making agent testing roughly 5x faster.
- Matt Pocock shared a wizard skill that builds interactive CLI wizards for tedious setup tasks like API keys, migrations, .env files, and third-party service configuration.
- Bloome launched a shared workspace where multiple AI agents, including Claude, ChatGPT, DeepSeek, Codex, and Gemini CLI, collaborate in one context with memory, sandboxes, and audit trails.
- Slite argued that teams are moving from static wikis to self-maintaining knowledge bases where AI detects stale docs, drafts updates, and routes fixes for human approval.
Advertisement
🧩 Additional Agent, Coding & Research Items
- Riley Brown showed how to build a custom voice-controlled “Jarvis” in under 20 minutes with Cursor, Claude Code, or Codex and OpenAI’s GPT-Realtime-2 API, then connect web search, computer control, image generation, and custom computer actions as tools.
- Sam Hogan’s Inference Gateway lets teams test GLM 5.2 against real production requests without changing what users see: normal traffic keeps going to the current model while the gateway silently runs and compares the candidate model in the background.
- Andrew Ng explained “loop engineering,” the practice of designing repeated agent cycles so coding systems inspect their work, gather feedback, and keep iterating instead of stopping after one answer.
- Scale AI introduced SWE-Interact, a benchmark for coding agents that evaluates back-and-forth collaboration instead of giving agents a perfect spec and scoring one autonomous attempt.
- Matt Pocock published /writing-great-skills, an installable helper for writing higher-quality skill and instruction packs for coding agents.
- Matt Pocock added Martin Fowler’s code smells to his /review skill so agents can flag patterns like duplicated code, vague names, feature envy, shotgun surgery, and overused primitive values during review.
- Tongyi Lab introduced HydraHead, an attention architecture that mixes full attention and cheaper linear attention inside individual attention heads instead of forcing whole model layers to use one method.
- Sebastian Raschka released Build a Reasoning Model (From Scratch), a 440-page guide to inference scaling, reinforcement learning, and distillation after 18 months of writing and experiments.
- Andrew Curran predicted a coming memory-efficiency architecture breakthrough from a team spun out of OpenAI, though he said it was based on what he had been told rather than confirmed inside knowledge.
- X’s AI trending page captured a broad mix of popular AI posts and conversations at the time, but did not point to one distinct news event or launch.
🧪 Research Bench
- Neural Procedural Memory proposed storing agent skills as activation steering vectors distilled from past experience, instead of relying only on text instructions or retrieval.
- The Verification Horizon argued that coding agents are running into a reward problem: test pass rates, LLM judges, and traces all eventually stop tracking true correctness and become hackable.
- ECHO explored how terminal agents can learn world models from interaction traces, improving planning and generalization without explicit world-model supervision.
- PTRM showed a 5M-parameter recursive model beating frontier LLMs on hard logic puzzles by scaling test-time compute in latent space at tiny inference cost.
- Lossfunk highlighted research showing LLMs trained only on text develop human-like internal geometry for concepts like color, pitch, emotion, and taste, usually peaking in middle layers.
- Aneesh Muppidi introduced real-time RL, where agents learn how long to think at each step in environments that keep moving while they plan.
- Hiroki Naganuma shared a TMLR paper showing warmup and smooth decay naturally emerge as robust learning-rate schedule patterns across workloads.
Advertisement
🏢 Big Tech / Model Infrastructure
- ClaudeDevs announced Claude Managed Agents updates including streaming session deltas, per-session overrides, new webhook events, credential scoping, reverse pagination, and a session observability tab.
- Boris Cherny said Claude Code’s next version runs subagents in the background by default, while forwarding permission requests to the main agent so users keep control.
- Artificial Analysis benchmarked Claude Sonnet 5 at #5 on its Intelligence Index, matching GPT-5.5 high-reasoning while using more tokens and more agentic turns than Sonnet 4.6.
- Nous Research said Hermes Agent now reads the web up to 60x faster and 49x cheaper by streaming clean scraped content directly into agents with local paging for large pages.
- Ollama said Gemma 4 is nearly 90% faster on Apple Silicon via MLX, helped by multi-token prediction now enabled by default.
- SemiAnalysis reported Google’s next TPU, codenamed Humufish, will use Intel EMIB-T packaging instead of TSMC CoWoS, trading cost and scaling upside for execution risk.
- Logan Kilpatrick announced Gemini API additions: Nano Banana 2 Lite for fast cheap image generation and Gemini Omni Flash for video editing.
- Howard Lutnick said the Commerce Department worked with Anthropic to analyze and approve Fable 5 before its return.
📖 Midweek Wisdom / Thought Pieces
- Dwarkesh Patel released a conversation with 3Blue1Brown’s Grant Sanderson on why AI may advance faster in math than other fields, proof verification, and whether humans will understand future AI-generated math.
- Sonya Huang released a Training Data episode with SemiAnalysis founder Dylan Patel on the AI infrastructure buildout, CUDA’s moat, sparse vs. dense models, and long-term compute bets.
- Omri Weinstein said a prover-verifier LLM harness changed his mind about AI doing real math research after it solved nine substantial open problems.
- Kimmonismus reported that OpenAI has found inference optimizations that more than halved model-running costs, potentially improving margins, usage limits, or API pricing pressure.
- Signüll praised Anthropic’s blog and visual explainers as unusually clear, readable technical communication that makes complex AI systems understandable without dumbing them down.
Advertisement
Previous Around the Horn Digests
The Bottom Line
Tuesday’s AI news was less about one single breakthrough and more about infrastructure settling into place. Models are getting cheaper to run, agents are becoming the default product shape, and scientific and industrial workflows are turning into the next major battleground.
