🤖 AI News — July 7, 2026

🧠Model & Product Launches

• 🆕 Claude Sonnet 5 now live as default model across all Anthropic plans — Anthropic / TechCrunch

  Claude Sonnet 5 is now the default across all Anthropic plans, replacing Sonnet 4.6 entirely.
  Key details: $2/M input tokens (intro) → $3/M, $10/M output → $15/M after Aug 31, 2026.
  Performance: near-Opus 4.8 on benchmarks (BrowseComp, OSWorld-Verified, coding evals) at a fraction of Opus pricing.
  Enterprise traction: insurance tech Pace running live multi-step insurance workflows on Sonnet 5; ClickHouse reports faster time-to-insight; Lovable using it for agentic app building.
  ⚠️ Breaking change: Sonnet 5 removes temperature and top_p parameters — API integrations that pass these will throw errors.

  Framing Launched June 30; now the default model for Free/Pro/Team/Enterprise users; Sonnet 4.6 retired
• Google launches Nano Banana 2 Lite and Gemini Omni Flash for developers — Google DeepMind Blog / Google Cloud

  Google DeepMind opened access to two new models on June 30.
  Nano Banana 2 Lite (gemini-3.1-flash-lite-image): text-to-image in ~4 seconds, $0.034 per 1K images, available in AI Studio, Gemini API, and Gemini Enterprise Agent Platform. Rolling out to consumer surfaces (AI Mode in Search, Gemini app).
  Gemini Omni Flash: conversational video editing and generation, $0.10/second of video, C2PA credentials + SynthID watermarks on by default.
  Google also announced a $75M partnership with A24 bringing AI into filmmaking workflows.

  Framing Image and video generation models open to developers; image outputs in ~4s
• Kimi K2.7 Code becomes first open-weight model in GitHub Copilot — GitHub Blog / Moonshot AI

  Kimi K2.7 Code is now available in GitHub Copilot's model picker — the first open-weight model on the platform.
  Enterprise admins must enable it manually. Uses provider list rates under Copilot's usage-based billing. This positions Copilot as a multi-model coding platform rather than just OpenAI-powered.

  Framing Moonshot AI's open-weight coding model gains direct Copilot integration
• Nvidia Nemotron-Labs-TwoTower — open-weight diffusion language model — arXiv / Nvidia / Hugging Face

  Nvidia released Nemotron-Labs-TwoTower, an open-weight diffusion language model that generates text in parallel rather than left-to-right. Trained on ~2.1 trillion tokens. Achieves 2.42x higher throughput while maintaining 98.7% of baseline quality. Available on Hugging Face as nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16.

  Framing Parallel text generation at 2.42x throughput vs standard autoregressive models
• Google TabFM — zero-shot model for tabular data classification/regression — Google AI

  Google released TabFM, a zero-shot foundation model for tabular data. Handles classification and regression without requiring task-specific training or feature engineering. Potentially significant for enterprise ML workflows where structured data dominates.

  Framing Eliminates task-specific training and manual feature engineering

🔧Infrastructure & Chips

• OpenAI and Broadcom unveil "Jalapeño" — custom LLM inference chip in 9-month tape-out — OpenAI / Broadcom

  OpenAI and Broadcom unveiled Jalapeño, OpenAI's first Intelligence Processor — a blank-slate design optimized specifically for LLM inference (not a general-purpose accelerator).
  Key facts: developed from initial design to manufacturing tape-out in just 9 months (fastest ASIC cycle ever claimed in advanced semiconductors). Designed with flexibility for all LLMs, not just OpenAI's own models. Early testing shows "performance per watt substantially better than current state-of-the-art." Deployed at gigawatt scale with Microsoft and other data center partners starting late 2026. Multi-generation roadmap with Broadcom and Celestica.

  Framing Game-changing: first custom AI accelerator from design to production in 9 months; delivered to Sam Altman by Hock Tan
• Meta launching cloud business to sell excess AI compute — Bloomberg / TechCrunch

  Meta is developing plans for a cloud infrastructure business ("Meta Compute"), selling access to both AI compute power and models — directly competing with AWS, GCP, and Azure. Led by Santosh Janardhan, Daniel Gross, and Dina Powell McCormick. Follows SpaceX/xAI's model of selling off excess compute capacity from Colossus data centers. Meta has committed $183B in future AI infrastructure spending. The Ohio data center project (size of Manhattan) is coming online this year.

  Framing Meta Compute initiative aims to turn $183B data center capex into a cloud revenue stream
• Nvidia RTX Spark — 1-petaflop AI superchip for Windows PCs — Nvidia

  Nvidia announced RTX Spark, a 1-petaflop AI superchip designed for Windows PCs, enabling local inference of large models without cloud dependency. Part of a broader push to make AI hardware accessible at the desktop level for developers and researchers.

  Framing Local AI inference at 1 PFLOPS on a desktop GPU

💰Funding, Deals & Market

• Global startup funding hits record $510B in H1 2026 — AI accounts for 70%+ — Crunchbase / TechCrunch

  Crunchbase data shows global startup investment hit $510B in H1 2026, an all-time record. AI-related companies account for more than 70% of total deal value. Exit activity surged: IPOs, SPACs, and M&A all accelerated dramatically. Q2 alone saw record-breaking rounds across AI infrastructure, model providers, and agent platforms.

  Framing AI boom drives all-time high in venture investment globally
• 8090 Labs (Chamath Palihapitiya) raises $135M Series A for AI coding platform — TechCrunch

  8090 Labs, founded by Chamath Palihapitiya, closed a $135M Series A led by Salesforce Ventures with participation from WndrCo, Craft Ventures, and fellow All-In hosts. Its "Software Factory" platform is an agentic coding system for regulated sectors (healthcare, aerospace). EY claims a 70% boost in software development productivity after rolling it out across tens of thousands of US consultants. Palihapitiya announced he will lead as CEO.

  Framing Salesforce Ventures leads round; Palihapitiya takes CEO role
• Reflection AI activates $6.3B compute lease at SpaceX Colossus 2 facility — TechCrunch / aiApps

  Reflection AI activated a $6.3 billion compute lease at SpaceX's Colossus 2 facility in Memphis on July 1. The deal locks in Nvidia GB300 chips through 2029 to train American open-weight frontier models. Positions compute as a strategic asset in domestic AI infrastructure development.

  Framing Nvidia GB300 chips locked in through 2029 for open-weight frontier model training
• South Korea commits $880B 10-year AI/semiconductor investment plan — Bloomberg / LinkedIn

  South Korea announced a 10-year, ~$880 billion investment plan covering semiconductors, AI infrastructure, and robotics. Samsung and SK Hynix are committing $518 billion of that toward new fabrication sites. The largest national AI infrastructure pledge outside the US, signaling Asia's race to secure AI chip supply chains.

  Framing Samsung and SK Hynix alone pledge $518B toward new chip fabrication sites

📄Papers & Research

• PACE: A Proxy for Agentic Capability Evaluation — arXiv (2607.02032)

  Paper proposes PACE, a proxy-based evaluation framework designed to measure agentic capability in LLMs without expensive full-task rollouts. Could become a standard metric alongside existing benchmarks like SWE-bench and BrowseComp.

  Framing New benchmark framework for measuring how agentic LLMs really are
• CausalMix: Data Mixture as Causal Inference for LLM Training — arXiv (2607.01104)

  CausalMix treats data mixture in LLM pre-training as a causal inference problem, proposing a principled framework for optimal data composition. Could have significant practical impact on how training datasets are curated for future foundation models.

  Framing Formalizing data mixing strategy through causal inference lenses
• AgenticSTS: Bounded-Memory Testbed for Long-Horizon LLM Agents — Hugging Face / arXiv (2607.02255)

  New testbed for measuring how LLM agents perform under bounded memory over long-horizon tasks. Addresses a key gap: most agent benchmarks test short-horizon tasks where context limits don't matter.

  Framing Evaluating agents under realistic memory constraints over extended tasks
• Understanding Large Language Models (comprehensive survey) — arXiv (2607.01006)

  A comprehensive survey paper covering the full landscape of LLM understanding — architecture, training methodologies, scaling laws, emergent capabilities, safety, and limitations. Primarily a reference/educational contribution.

  Framing 100+ page survey spanning architecture, training, capabilities, and limitations

🌐Open Source & Community

• 🏆 GitHub Trending this week — caveman, strix, OmniRoute lead — GitHub Trending

  Top trending repos this week:
  1. caveman (86K ⭐) — Claude Code skill that cuts 65% of tokens by speaking like a caveman. Viral utility for token-conscious developers.
  2. strix (38K ⭐) — Open-source AI penetration testing tool for app vulnerabilities.
  3. OmniRoute (12.8K ⭐) — Free AI gateway to 231+ providers (50 free), token compression (15-95%), MCP/A2A support.
  4. codex-plugin-cc (26.5K ⭐) — OpenAI's Codex plugin to use Codex from Claude Code.
  5. codebase-memory-mcp (27.7K ⭐) — High-performance MCP server that indexes codebases into a knowledge graph in milliseconds. 158 languages, sub-ms queries.
  6. alibaba page-agent (24.8K ⭐) — In-page GUI agent that lets you control web interfaces with natural language.
  7. huggingface/speech-to-speech (5.5K ⭐) — Build local voice agents with open-source models.

  Framing Claude Code ecosystem dominates; security and token optimization trending
• Hugging Face LeRobot v0.6.0 — robotics simulation for AI agents — Hugging Face Blog

  Hugging Face released LeRobot v0.6.0 with new simulation capabilities for training and evaluating robot learning models. Adds "Imagine, Evaluate, Improve" workflow for iterative robotics AI development.

  Framing "Imagine, Evaluate, Improve" — new simulation workflows for robot learning
• Nvidia Nemotron-Labs-TwoTower now available on Hugging Face — Hugging Face (nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16)

  Nvidia released Nemotron-Labs-TwoTower-30B-A3B under an open-weight license on Hugging Face. The diffusion-based architecture generates text in parallel (not left-to-right), achieving 2.42x higher throughput. Trained on 2.1T tokens.

  Framing Open-weight diffusion LLM for high-throughput parallel text generation

⚖️Regulation & Safety

• Microsoft lays off 4,800 workers — "AI is changing how work gets done" — Microsoft Blog / TechCrunch / ABC News

  Microsoft eliminated ~4,800 roles (2.1% of global workforce) on July 6, adding to the 120,000+ tech jobs cut in 2026. The company claimed these roles are "not being replaced by AI" but acknowledged that "AI is changing how work gets done and automating everyday tasks." Cuts hit Commercial and Xbox organizations the hardest. Microsoft has redeployed 4,000+ employees into new roles in the past year.

  Framing 2.1% workforce cut; Microsoft says roles "not being replaced by AI" but acknowledges automation
• Five Eyes warns frontier AI will "fundamentally transform" cyber warfare — timeline is months — NCSC UK (official statement pdf) / CSA

  The Five Eyes intelligence alliance (US, UK, Canada, Australia, NZ) issued a joint statement warning that frontier AI models will fundamentally transform both offensive and defensive cyber capabilities. Key line: "the timeline is not years, it is months." For businesses, this means planning for identity checks, safety audits, and default watermarking on AI-generated media. The statement establishes a new compliance baseline for AI security practices.

  Framing Intelligence alliance issues blunt warning: not years, months
• Zuckerberg tells staff AI agents haven't progressed as expected — TechCrunch / Reuters / Bloomberg

  In an internal town hall on July 2, Mark Zuckerberg told Meta staff that AI agent development had not "accelerated in the way" executives had previously expected. This follows Meta's 8,000-person layoff and 7,000-person AI reassignment. Zuckerberg said the reorg was less "clean" than it should have been, and that the AI-focused structure's upside hadn't "come to fruition yet." He expects improvements in 3-6 months. Engineers in Meta's AI unit have described conditions as a "soul-crushing gulag."

  Framing Meta CEO admits AI agent development hasn't "accelerated in the way" executives expected
• Frontier model access increasingly restricted — ID verification, credits, vetted previews — aiApps / Anthropic

  A clear pattern is emerging: frontier AI access is becoming verified, metered, and restricted.
  Fable 5 returns with mandatory identity verification through Persona (starting July 8) and shifts to usage credits instead of flat-rate subscriptions.
  Mythos 5 is limited to vetted US critical infrastructure defenders through Project Glasswing.
  GPT-5.6 Sol is restricted to ~20 approved organizations in a government-gated preview.
  The pattern: if you want the most capable models, expect identity checks, safety audits, and default watermarking.

  Framing Fable 5 requires Persona ID verification (from July 8); Mythos 5 limited to critical infrastructure defenders

🏢Industry Moves

• VMware+AWS: Codex plug-in for Claude Code lets devs use OpenAI from Anthropic's platform — GitHub / OpenAI

  OpenAI released codex-plugin-cc, a plugin that allows developers to use OpenAI's Codex agent from within Claude Code — enabling code review and task delegation across rival AI platforms. Surged to 26.5K stars on GitHub this week. Signals a multi-model future where developers mix and match frontend/backend AI agents regardless of vendor.

  Framing Interoperability move — use Codex from inside Claude Code
• Google signs $75M AI filmmaking deal with A24 — aiApps / Google

  Google signed a $75 million partnership with A24 (award-winning independent studio) to bring AI into filmmaking workflows. The deal includes use of Gemini Omni Flash for video editing and generative media. First major Hollywood studio deal explicitly for AI-assisted film production.

  Framing First major Hollywood studio partnership for AI-assisted movie production
• Anthropic Sonnet 5 removes temperature/top_p — developer migration pain point — Anthropic / aiApps

  Claude Sonnet 5 removes the temperature and top_p sampling parameters from its API. Existing code that passes these will throw errors. Developers must audit integrations before migrating. The change aligns with Anthropic's "effort" parameterization approach, which they argue produces more reliable agentic behavior.

  Framing Breaking API change as Anthropic moves to fixed-effort parameterization

🔮Trends & Analysis

• Models are getting cheaper; access is getting tighter

  Three converging trends defined this week: (1) Model pricing is dropping dramatically — Sonnet 5 offers near-flagship performance at commodity prices. (2) Frontier access is simultaneously tightening — ID verification, restricted previews, and government-gated access are now standard for top-tier models. (3) Infrastructure spending is accelerating — $6.3B compute leases, $183B data center commits, and national AI investment plans signal that the compute arms race hasn't peaked. The net effect: mid-tier models democratize capability, while frontier models become more exclusive than ever.

  Framing The July trend: useful models arriving fast, but gated behind identity, credits, and government restrictions
• AI layoffs hit 120K+ in 2026 — Microsoft, Oracle, Meta, GitLab lead

  Tech layoffs in 2026 have now exceeded 120,000, per Layoffs.fyi — the highest single-month total in years (May). Microsoft's 4,800-person cut on July 6 adds to a pattern: Oracle (-21K), Meta (-8K), GitLab (-14%), Cisco (-4K), Intuit (-17%), Cloudflare (-20%). Most cite AI either directly (automation) or indirectly (reallocation toward AI investment). The disconnect between record revenues and mass layoffs is fueling a growing backlash — TechCrunch recently covered why the "AI layoff wave is becoming a powder keg."

  Framing The industry's contradiction: record revenue + AI-driven headcount reduction
• The compute-as-a-service business model is emerging

  Both Meta and SpaceX are now selling excess AI compute capacity as a service, following CoreWeave's model. Meta Compute will compete with AWS/GCP/Azure. SpaceX's Colossus facilities have inked deals with Anthropic, Google, and Reflection AI. The winners of the AI race may be those who own the data centers, not those with the best models. Skeptics (including Burry) warn of a bubble built on rapidly depreciating chips and uncertain end-user revenue.

  Framing Meta and SpaceX are turning excess data center capacity into cloud compute businesses