🗞️ AI Daily Briefing — 2026-07-10

🔥 Top Story

GPT-5.6 Sol goes fully public — OpenAI’s most powerful model is now available to everyone. After a 12-day government-supervised preview, OpenAI released Sol, Terra, and Luna across ChatGPT, the API, and Codex yesterday. Sol ($5/$30 per 1M tokens) hits 750 tokens/sec on Cerebras hardware — roughly 10x faster than Nvidia GPU inference for a frontier model. Terra ($2.50/$15) slots as the everyday workhorse; Luna ($1/$6) is the speed tier. This is the first frontier model release to go through the White House’s voluntary AI review framework, and the fact that it passed sets a precedent for how future launches may be gated. (CNBC) (OpenAI) (Eastern Herald)

🚀 Model & Research News

  • Google scraps Gemini 2.5 Pro architecture entirely, rebuilds Gemini 3.5 Pro from scratch. DeepMind pushed the launch to July 17, abandoning the existing base after hitting ceilings in multi-step math reasoning and SVG generation. The rebuild targets a 2M token context window, a “Deep Think Reasoning Layer,” and parity with GPT-5.6 and Fable 5. Meanwhile, four senior DeepMind researchers including AlphaFold’s John Jumper have left for Anthropic. (BigGo Finance) (HackerNoon) (Bind AI)

  • Anthropic extends free Fable 5 access to July 12 — metered credits start July 13. Fable 5 remains #1 on every major leaderboard (Arena Text, Code, Agent, Intelligence Index at 64.9). Starting July 13, usage shifts from included-in-subscription to metered credits at $10/$50 per 1M tokens. Anthropic says the rationing is capacity-driven, not permanent — they aim to restore Fable 5 as a standard subscription feature once inference capacity catches up. (Forbes) (Android Authority)

  • Claude Sonnet 5 launched with near-Opus performance at introductory pricing. Anthropic’s newest Sonnet is billed as “the most agentic Sonnet yet” — autonomously builds plans, operates tools, and carries out multi-step tasks. Introductory price of $2/$10 per 1M tokens through August 31 makes it aggressively competitive. (Anthropic) (MacRumors)

  • DeepMind launches $10M multi-agent AI safety fund. Together with Schmidt Sciences, the Cooperative AI Foundation, and ARIA, the fund targets research into how large-scale multi-agent systems behave as a group — sandboxes, agent network science, infrastructure, and oversight. Grants from $300K to $1M, applications due August 8. (Google DeepMind) (MIT Technology Review)

  • AI coding agents are setting off endpoint detection rules written for human intruders. Sophos reports that Claude Code, Cursor, and OpenAI Codex trigger alerts designed to catch lateral movement and credential access. Separately, Wiz disclosed “GhostApproval” — a technique that defeats approval prompts in coding agents — and the AI Now Institute published “Friendly Fire,” undermining the model’s own safety judgment. (SecurityWeek) (VentureBeat) (Security Point Break)

💰 Funding & Business

  • Together AI raises $800M Series C at $8.3B valuation. Led by Aramco Ventures with NVIDIA, General Catalyst, and Vista Equity participating. Annual bookings exceed $1.15B. The open-source inference platform has secured 500+ MW of compute capacity commitments and plans to grow infrastructure 50x over five years. (TechCrunch) (Together AI)

  • Prime Intellect raises $130M Series A at $1B valuation. Led by Radical Ventures with NVIDIA Ventures, Intel Capital, and Dell Technologies Capital. The startup helps enterprises train their own AI agents with large-scale RL, sandboxes, and evals — 6,000 customers including Ramp and Zapier, $100M+ ARR. (TechCrunch) (SiliconANGLE)

  • Norm AI closes $120M Series C (Khosla Ventures) for AI compliance. Taktile also secured $110M Series C (Goldman Sachs) for agentic decision-making in banking and insurance. Robotics continues attracting capital: Zeroth raised ~$74M for humanoid robotics (Ant Group), Tripo AI ~$150M for 3D generation, and Yingzhi XBOT $56M for robotic systems. (Crunchbase)

  • OpenAI floats giving the U.S. government a 5% stake (~$42.6B). Part of ongoing negotiations as OpenAI shifts from a capped-profit to a for-profit structure. The government stake would be unprecedented for a tech company and signals how deeply intertwined frontier AI labs and national security have become. (Fortune)

🐦 Notable from the Timeline

  • Alexandr Wang (now Meta’s Chief AI Officer) reportedly told a Meta town hall that their next model “Watermelon” is matching GPT-5.5 benchmarks — trained with 10x the compute of Muse Spark. A coding-focused Muse Spark update is also imminent.

  • Sam Altman has been vocal about the government review process working: “this is what responsible scaling looks like — build, test, share, iterate.” GPT-5.6 Sol’s public launch is the culmination of that message.

  • François Chollet has been notably quiet this week, but ARC-AGI Prize discussions continue to heat up as GPT-5.6 Sol and Fable 5 both claim strong reasoning advances.

  • Anthropic confirmed they are in early-stage discussions with Microsoft to run Claude inference on Microsoft’s custom Maia 200 chips (3nm, inference-optimized) via Azure — a potential strategic shift away from exclusive AWS reliance.

📊 Benchmark Watch

Claude Fable 5 continues to dominate the Arena leaderboard with 7.15M+ votes across 369 models. Claude Sonnet 5 (thinking) was added to Code, Text, Search, Vision, and Document leaderboards this week. GPT-5.6 Sol’s benchmark numbers are expected to shake up rankings once fully evaluated — early reports suggest it’s competitive with Fable 5 on coding and reasoning, though pricing is 2x cheaper. Meta’s muse-image and muse-video were added to Arena’s image and video generation leaderboards on July 7. (Arena.ai) (Swfte)

🔗 Worth Reading

  • “Six Exploits Broke AI Coding Agents — IAM Never Saw Them” — VentureBeat’s deep dive into how attackers target credentials rather than models in Claude Code, Copilot, and Codex. (VentureBeat)

  • “Google DeepMind Is Worried About What Happens When Millions of Agents Start to Interact” — MIT Tech Review on the $10M multi-agent safety research initiative and why emergent behavior at scale is the next big safety problem. (MIT Technology Review)

  • “Why Meta Paid $14.3B for Scale AI” — Forbes on the strategic logic behind Meta’s acquisition and Alexandr Wang’s rocky first year building Superintelligence Labs. (Forbes)