AI Daily Brief — 25 June 2026
A day when policy, capital and silicon all moved at once. Washington asked OpenAI to slow-roll GPT-5.6, OpenAI is reportedly pushing its IPO into 2027, Anthropic accused Alibaba of mining Claude through 25 000 accounts, Adobe bought Topaz Labs, Hugging Face crossed $100 M ARR, and IBM unveiled a sub-1 nm chip — while NVIDIA shipped an optimised GLM-5.2 and the diffusion-LM wave finally cracked into open weights.
Top stories
- White House asks OpenAI to delay GPT-5.6. Staged release to select partners only; a separate report has the IPO sliding into next year. via TechCrunch · The Verge
- Anthropic vs Alibaba. Anthropic’s letter alleges 25 000 accounts and 28.8 M exchanges used to distil Claude into Qwen-class checkpoints. Cohere jabbed at “staggered releases”; Nato Lambert called the letter format “smear-campaign energy”. via Ars Technica
- NVIDIA ships an optimised GLM-5.2. 753 B-parameter MoE, 1 M context, quantised to NVFP4 for Blackwell; Databricks pushed serving to 392 tok/s. The most explicit “China model on Western silicon” pairing of the year. via @teortaxesTex
- Adobe acquires Topaz Labs. Image- and video-enhancement tooling folds into Adobe’s apps. via TechCrunch
- Hugging Face crosses $100 M ARR. Clem Delangue paired the milestone with a pointed line — “maximising short-term revenue has never been our priority”. via Hugging Face
- IBM debuts a sub-1 nm chip. 0.7 nm node with claimed wafer-bonded transistor stacking. via Ars Technica
Who shipped
OpenAI went loud on Codex adoption (Brockman’s department-by-department brag, plus a DigitalOcean plugin) and a Reddit sighting of its Broadcom-built inference chip “Jalapeño” — but quiet on models. Google DeepMind baked computer use into Gemini 3.5 Flash (78.4 on OSWorld, on par with GPT-5.5) and partnered with ElevenLabs to embed SynthID into ElevenLabs audio. Anthropic joined RAISE US as a founding partner. xAI lit SuperGrok inside T3code; Grok shipped /goal for long-running tasks. Sakana AI opened an RSI Lab focused on sample-efficient self-improvement and released CoffeeBench (with KPMG). Liquid AI shipped LFM2.5-230M for edge.
Open-source pulse
Gemma 4 crossed 200 M downloads in 2.5 months — the whole Gemma family was at 100 M when Gemma 3 launched. DeepReinforce’s Ornith-1.0 on Gemma 4 + Qwen 3.5 reports 82.4 SWE-Bench Verified at the 397 B flagship (model learns its own RL scaffold). The diffusion-LM wave finally hit open weights: NVIDIA’s Nemotron-TwoTower-30B-A3B (autoregressive context tower + diffusion generator) and iLLaDA-8B (masked diffusion, bidirectional attention). Baidu’s Unlimited OCR (3 B MoE) keeps the KV cache flat across long documents. Together AI’s Parakeet hit 302× real-time STT.
Money, infra & hardware
$320 M into General Intuition on the thesis that Fortnite gameplay is real-world-action training data; $50 M into Patronus AI; Amazon $13 B into India AI infra; Azeem Azhar pegged the GenAI economy at $110 B trailing-12-months, $175 B annualised. Apple raised every Mac price 20 % and Bloomberg says M6 Pro/Max is skipped in favour of an AI-focused M7 Pro/Max/Ultra. Qualcomm entered the data centre with the Dragonfly C1000. Weaviate 1.38 hit GA with a built-in MCP server. Cloudflare launched self-managed OAuth for all.
Quiet, and underdelivered
Mistral, Meta (Llama) and DeepSeek were release-shaped quiet (DeepSeek’s only signal: a job-posting wave hinting at a “Frontier Research Group” + headcount doubling). Ford rehired ~350 engineers after automated systems failed to preserve institutional expertise. The Authors Guild showed two of five AI-detection tools flag every human-written article as AI. The Decoder reported Grok’s traffic is over half adult content and Google is bleeding researchers to rivals.
By the numbers
- 451 stories tracked in 24 h across 60+ sources
- Most-mentioned model: GLM-5.2
- Most-mentioned lab: OpenAI (delay, IPO, Codex, ads-on-paid-tier)
- Notable absences: Mistral, Meta (Llama), DeepSeek releases
Compiled by AI Feed’s editor from all 451 headlines published on 25 June 2026.