Skip to content
Source · Daily Brief

AI Daily Brief — 22 June 2026

Sakana AI launched Fugu — a multi-agent orchestration LLM that calls a pool of frontier models (including itself recursively) through a single API and matches Anthropic’s Fable / Mythos on engineering and reasoning benchmarks. OpenAI rolled out Daybreak with GPT-5.5-Cyber + Codex Security + the Patch the Planet OSS initiative. NVIDIA used its ISC keynote to ship a wall of scientific-AI software (DAQIRI, cuPhoton, ALCHEMI), the JUPITER exascale system, Vera CPU at Los Alamos, and 45°C liquid cooling. And SpaceX inked a $150M/month compute deal with Reflection AI.

Top stories

  • Sakana Fugu launches. A “full multi-agent orchestration system accessible via a single model API”; Fugu Ultra was live on Vercel AI Gateway hours after announcement. Sakana published the technical report and Nikkei carried it as a Japanese tech-story headline. via Sakana AI
  • OpenAI’s Daybreak push: GPT-5.5-Cyber + Patch the Planet. Codex Security plugin for security teams (deep scans, validated findings, codebase-specific patches); GPT-5.5-Cyber claimed state-of-the-art on CyberGym; Daybreak now patches critical vulns in major browsers and network infra. via OpenAI · Patch the Planet
  • Anthropic + Micron co-design memory. Micron joins Anthropic’s Series H and gets a multi-year deal to supply memory for Claude’s training and serving stack. Tom Brown calls memory the next infra bottleneck. via The Decoder
  • SpaceX × Reflection AI: $150 M / month. Open-source lab Reflection AI gets immediate access to NVIDIA GB300 capacity at SpaceX’s Colossus 2 Memphis site through 2029. via TechCrunch
  • Google DeepMind × A24: $75 M for Hollywood tools. Research partnership to “ensure the tools of the future are shaped by the creators who use them”. via TechCrunch
  • Google ships Interactions API as default for Gemini. Typed steps replace role-based generateContent schema across Gemini models and agents. via The Decoder

Who shipped

Sakana shipped Fugu (Vercel-live, multi-orchestrator). OpenAI shipped Daybreak + GPT-5.5-Cyber + Codex Security + Patch the Planet (combined cybersec push). NVIDIA ran an ISC-conference wall: JUPITER exascale system online, Vera CPU at Los Alamos, DAQIRI for high-speed data acquisition, cuPhoton + ALCHEMI scientific libraries, NVIDIA Halos full-stack robotics safety, CCCL CUDA C++ runtime, 45°C liquid cooling, NVQLink for quantum, NAIRR scientific program update, and announced GTC returns to San Jose March 2027. xAI‘s Grok Build added /goal for long-running autonomous execution with subagent verification — and a Grok × Interactive Brokers integration for portfolio queries. Anthropic brought the full Claude Desktop experience to AWS, GCP and Microsoft Foundry. Google shipped the A2A protocol (Python ↔ Go cross-language agents) plus a “Measuring What Matters with Jules” telemetry post. Mistral hit 1 000 employees worldwide. Runway‘s Aleph 2.0 expands video scenes to fit any aspect ratio.

Money & capacity

Groq confirmed $650 M raise + re-staffing after the Nvidia $20 B “not-acqui-hire” — leaning into the neocloud business. Together AI + 5C announced GB300 NVL72 deployment for purpose-built inference. GM installed factory robots after 1 300 layoffs; the UAW called the “dark factory” pattern out. Anthropic may have “talked itself into an AI export ban” per Ars Technica. Hugging Face is days away from 3 M public models + 1 M datasets.

Open-source pulse

OpenMed hit 7.6 M downloads on HF (+100 k/day), trended on GitHub. Moebius: a 0.2 B image-inpainting model with claimed 10 B-level performance was ported by Simon Willison to run entirely in the browser via Claude Code. GLM-5.2 reached 131 tok/s on Together. GLM-5.2 × AWS Marketplace went live via Z.ai’s API. Nathan Lambert calls GLM-5.2 the “DeepSeek moment for agents”.

Quiet corners

Apple absent again. DeepSeek silent on releases. Meta’s Llama camp produced nothing release-shaped. Cohere only had a CEO talking-head moment about sovereign AI at FII Institute.

By the numbers

  • 109 stories tracked in 24 h across 50+ sources
  • Most-mentioned model: GLM-5.2 (AWS, Together, Perplexity, Z.ai)
  • Most-mentioned lab: NVIDIA (ISC keynote wall: scientific AI + cooling + supercomputers + GTC return)
  • Notable absences: Apple, DeepSeek, Meta (Llama)

Compiled by AI Feed’s editor from all 109 headlines published on 22 June 2026.