AI News Intelligence Digest — June 25, 2026

Top Stories

• Google DeepMind brain drain reaches crisis level. Nobel laureate John Jumper (AlphaFold) announced he's leaving for Anthropic. This follows Noam Shazeer (co-author of "Attention Is All You Need") departing for OpenAI on June 18. Two more researchers — Jonas Adler and Alexander Pritzel — are also expected to join Anthropic. GOOG dropped ~6% in five sessions. Demis Hassabis insists Google is "still winning AI talent" but the market disagrees. (Fortune — "As top talent leaves Google DeepMind, some question if the lab can remain at the forefront"; TechCrunch — "Nobel laureate John Jumper is leaving DeepMind for rival Anthropic")

• Fable 5 & Mythos 5 still offline — day 13 of US export ban. The Commerce Department ordered Anthropic to disable both models on June 12 after a jailbreak was found that could bypass Fable 5's safeguards to access Mythos's cybersecurity capabilities. Anthropic argues the same jailbreak works on GPT-5.5 (which faces no ban) and calls the recall disproportionate. All other Claude models remain available. (Fortune — "Anthropic disables Fable and Mythos AI models following U.S. government export ban")

• SpaceX's Colossus becomes an $80B+ compute platform. Reflection AI signed a $6.3B deal for GB300 chips at Colossus 2 ($150M/month through 2029). This follows Anthropic's ~$45B deal for Colossus 1 and Google's ~$30B commitment at Colossus 2. SpaceX has quietly built one of the world's largest third-party compute platforms. (CNBC — "SpaceX signs computing power deal with open-source AI startup Reflection worth up to $6.3 billion")

• June 2026 is the most packed model month in AI history. GPT-5.6, Gemini 3.5 Pro, and Claude 4.8 are all shipping or imminent. Polymarket prices GPT-5.6 release by June 30 at 89%. (centerbit — "AI Rumors June 2026: GPT-5.6, Gemini 3.5 Pro, Claude Mythos")

Model Updates

• GPT-5.5-Cyber launched June 22 — a cybersecurity-focused model scoring 85.6% on CyberGym, beating Mythos 5. Designed for full defensive security: vuln finding, validation, patching, and fix testing. (unrot.co — "Top 10 AI News Stories: June 25, 2026")
• GPT-5.6 imminent — leaked routing reference found in Codex backend. Rumored 1.5M token context window (up from 1.05M on GPT-5.5), improved UI generation, sharper long-horizon coding. Internal OpenAI memo calls it "a meaningful improvement." (centerbit — "AI Rumors June 2026")
• Claude Opus 4.8 shipped May 28 — just 6 weeks after Opus 4.7. Anthropic's release velocity is accelerating.
• Gemini 3.5 Pro announced at I/O 2026 (May 19) with "next month" timeline — expected any day now.

Open Source

• MiniMax M3 — first open-weight model combining frontier-tier SWE capabilities with 1M token context and native multimodal computer use. Tops open-weight SWE-Bench Pro at 59.0%. (devFlokers — "Open-Source AI June 2026")
• GLM-5.2 (Zhipu AI) — released June 13 under MIT license. 1M context window, built for agentic engineering and long-horizon reasoning.
• IplanRIO Rio 3.5 Open 397B — MIT license, outperforms DeepSeek V4 Pro on Terminal-Bench 2.1.
• NVIDIA Nemotron 3 Ultra 550B-A55B shipped June 4; Google posted DiffusionGemma 26B-A4B June 10.
• Zyphra ZAYA1-8B — Apache 2.0, sparse routing with only 760M active params per token out of 8B total.

Industry Moves

• Anthropic now valued higher than OpenAI — $965B Series H post-money (late May) vs OpenAI's $852B (March). First time Anthropic has held the top spot.
• DeepMind exodus accelerating Anthropic's talent pipeline — Jumper, Adler, and Pritzel all heading to Anthropic. Combined with Shazeer to OpenAI, Google is losing foundational researchers.
• AI agent startup consolidation underway — underfunded teams seeking soft landings via acqui-hires as VC funding concentrates at the top.
• Q1 2026 VC records — AI received $242B (80% of total global venture funding). AI startups now attract 33% of all VC funding.

Safety & Policy

• US export controls on frontier AI models remain the dominant policy story. Fable 5/Mythos 5 ban highlights tension between national security and commercial AI access. The ban extends to foreign nationals inside the US, including Anthropic's own employees.
• Trump executive order (June 2) — "Promoting Advanced AI Innovation and Security." Asks AI developers to voluntarily share new models with the federal government up to 30 days before partner access. Creates AI-cybersecurity clearinghouse.
• Colorado AI Act was set for June 30 but was repealed and replaced with SB 26-189 — a narrower law favoring consumer notices and human review rights over risk management programs.
• EU AI Act deadline approaching — August 2, 2026 compliance required for transparency rules and high-risk AI systems in critical infrastructure, employment, law enforcement.

Worth Watching

• "June 2026 AI: Claude 4.8, GPT-5.6, Gemini 3.5 Pro All Landing at Once" — Comprehensive overview of the simultaneous model releases and what they mean for the competitive landscape.

Developing stories: The DeepMind brain drain and Fable 5 export ban are both multi-week stories worth tracking. The Colossus compute platform is emerging as a structural shift in AI infrastructure. Previous digests failed due to auth errors — this is the first successful digest in the current run.