AI News - Thu Jun 25 2026
.jpeg)
The Tech‑Reader AI Digest
Thursday, June 25, 2026
#AI #TechNews #Digest
Day 13 of the Fable 5 shutdown, and the Five Eyes just told the world what the NSA already told the Senate—this is not a software problem. Meanwhile, Google quietly shipped a benchmark-rewriting model that almost nobody noticed, and the scale of the xAI compute empire finally has a public number attached to it.
Story 1: Five Eyes Says "Months, Not Years" — and the Fable 5 Shutdown Suddenly Makes More Sense
What happened: Fable 5 and Mythos 5 remain offline on Day 13, with no public statement from Anthropic, the Commerce Department, or the White House since June 18. Polymarket's Fable 5 restoration market has now drawn over $2 million in contract volume, with prediction markets still pricing roughly 55–57% odds of restoration before July 1.
The development that reframes the entire crisis arrived Monday in a rare joint statement from the Five Eyes intelligence alliance—the US, UK, Canada, Australia, and New Zealand. "Frontier AI models are anticipated to exceed current industry expectations, fundamentally transforming both offensive and defensive cyber capabilities," the agencies stated. "The timeline is not years, it is months."
The statement was signed by the NSA's Director of the Cybersecurity Directorate, David Imbordino, and acting CISA Director Nick Andersen, among others. Crucially, it does not cite classified sources, meaning the underlying justification is based on public AI capability assessments, not secret intelligence. The timing is pointed: the warning lands squarely alongside the U.S. government's unprecedented export ban on Anthropic's flagship models, a move that cut allied governments off from frontier access while raising massive questions about enterprise liability.
Why it matters: This statement is the first time allied governments have publicly validated, at the agency level, the exact threat model that drove the Fable 5 shutdown. For thirteen days, Anthropic has framed the export control directive as a bureaucratic overreaction to a narrow jailbreak. The Five Eyes statement doesn't name Anthropic directly, but it describes the exact threat category that Mythos demonstrated in the NSA's red-team exercise. The policy crossfire is now explicit: the most capable defensive AI tools are precisely what governments are moving to gate, at the exact moment enterprise defenders need them most.
Aaron's take — "The timeline is not years, it is months." Five allied intelligence agencies do not issue a sentence like that casually. The Fable 5 shutdown spent thirteen days looking like a heavy-handed overreaction to a prompt injection. The Five Eyes statement retrospectively reframes it as the opening move in a massive governance framework the Western alliance has been quietly assembling. The cruel irony here—which the agencies flag in their own statement—is that restricting the best defensive AI tools actively disadvantages enterprise defenders while foreign adversaries continue developing equivalent capabilities with zero restrictions. The government knows this, and they shut Fable 5 down anyway. That tells you everything you need to know about how terrified they were by the NSA's red-team results.
Story 2: Google Shipped a Benchmark-Rewriting Model — and Almost Nobody Noticed
What happened: While the AI industry spent this week watching the Fable 5 shutdown and waiting for GPT-5.6, Google quietly launched Gemini 2.5 Pro with Deep Think on June 22. The benchmarks have been rewriting the leaderboard all week.
Gemini 2.5 Pro Deep Think now leads on science and reasoning: 82.4% GPQA Diamond, 89.8% MMLU-Pro, and 94.1% HumanEval+, backed by a 2-million-token context window. Fable 5 still holds the crown on software engineering (88.6% SWE-bench Verified, 88.0% Terminal-Bench 2.1)—but Fable 5 is offline. For teams building in coding, Anthropic remains the theoretical leader. But for teams doing research, life sciences, financial analysis, or hard math, Gemini 2.5 Pro Deep Think's benchmark shift is a massive real-world advantage.
The critical distinction: this is Gemini 2.5 Pro with Deep Think, not Gemini 3.5 Pro. Google promised 3.5 Pro at I/O in May with a June general availability target, and it still has not shipped. Google faces severe credibility risk if it misses its self-imposed June deadline, and only five days remain in the month.
Why it matters: Google launched a genuinely competitive, state-of-the-art model on one of the most news-saturated days of the year, and it barely generated a headline. The Anthropic shutdown and the OpenAI anticipation consumed the entire developer ecosystem. For Google, this is a catastrophic distribution and narrative failure. The lab producing real, verifiable benchmark leadership cannot cut through the noise because its model family naming conventions are a mess, and its actual promised flagship is MIA.
Aaron's take — Google dropped its most capable model in months on the exact day everyone was watching Anthropic's paywall go live and waiting for OpenAI's next move. Gemini 2.5 Pro Deep Think leading GPQA Diamond and MMLU-Pro is a massive achievement—those are the benchmarks that actually matter for serious scientific workloads. But the coverage gap points to a severe structural failure at Google. Their model family is now so violently fragmented between 2.5 Pro, 3.1 Pro, 3.5 Flash, and the still-unshipped 3.5 Pro that even engaged developers can't track what is actually available. Sundar Pichai has exactly five days to ship Gemini 3.5 Pro before his June commitment becomes a broken promise. Topping the benchmarks with a 2.5 variant will not save him if he misses the deadline on the 3.5 flagship.
Story 3: The xAI Compute Empire — $80 Billion in Committed Revenue and Counting
What happened: The full scale of SpaceX's Colossus compute business came into sharp focus this week as the Reflection AI deal details leaked. On June 22, SpaceX signed a compute lease with Reflection AI for $150 million per month, starting July 1, 2026, through the end of 2029. If the contract runs its full term, it totals approximately $6.3 billion. Reflection gains immediate access to Nvidia GB300 chips at SpaceX's Colossus 2 facility in Memphis, Tennessee.
This is the fourth major external compute lease SpaceX has signed for Colossus, following Anthropic (approx. $45 billion through mid-2029), Google (approx. $30 billion through 2029), and Cursor. SpaceX's committed compute revenues from outside clients now exceed $80 billion through 2029. Colossus was originally built simply to train Grok; it has now been converted into one of the largest commercial AI compute platforms on the planet.
Reflection AI, founded by DeepMind veterans and backed by Nvidia and Sequoia, operates on a highly specific thesis: Western governments, banks, and defense contractors want frontier-capable AI with open weights, but they refuse to use closed U.S. labs for data sovereignty reasons, and they refuse to use Chinese open-weight models for security reasons. Reflection is positioning itself as the third option: American, open-weight, and frontier-scale.
Why it matters: The Colossus revenue picture completely reframes SpaceX's AI strategy. Elon Musk didn't just build a compute cluster to train Grok; he built an infrastructure monopoly that generates $80 billion in committed external revenue while training his own models on the side. Anthropic—the company currently paralyzed by a Commerce Department standoff—is paying SpaceX approximately $45 billion through 2029 for compute. The AI industry's supply chain entanglements are now so circular that the company grounding Anthropic's flagship product is actively subsidized by the infrastructure bill Anthropic pays to xAI's parent company.
Aaron's take — Let that number sink in: $80 billion in committed compute revenues from outside clients through 2029. Colossus started as Musk's brute-force answer to a GPU shortage. It has rapidly morphed into the most valuable piece of AI infrastructure on earth, with Anthropic, Google, and Reflection all cutting checks to SpaceX just to stay in the game. The Reflection deal is the most strategically brilliant of the four. You have an American open-weight lab backed by Nvidia, positioning itself as the sovereignty-safe alternative to closed U.S. labs right as the Five Eyes are warning about AI cyber threats. Reflection just locked in the exact compute it needs to build the model that governments actually want to buy, and Musk gets paid either way.
Quick Hits — The Rest of Today's AI World
Anthropic / Claude
- Fable 5 and Mythos 5 remain offline — Day 13. No public statement from Anthropic, Commerce, or the White House since June 18.
- Five Eyes joint statement validates the threat model behind the shutdown: frontier AI cyber capability is "months, not years" from broad availability.
- July 8 Persona ID verification rollout remains the most structurally likely mechanism for U.S.-first partial restoration.
- Opus 4.8 remains the active fallback across Claude Code and the API.
OpenAI
- GPT-5.6 June window has officially closed. July is now the confirmed base case with roughly 94% Polymarket odds by the end of July.
- GPT-4.5 retires from ChatGPT tomorrow, June 27.
- GPT-5.5-Cyber remains the most capable publicly acknowledged cybersecurity AI model (85.6% CyberGym). Gated exclusively to Trusted Access for Cyber partners.
xAI / SpaceX
- Colossus committed external compute revenues now exceed $80B through 2029 across four tenants: Anthropic, Google, Cursor, and Reflection AI.
- SPCX trading. $60B Cursor acquisition pending Q3 regulatory close.
Gemini (Google)
- Gemini 2.5 Pro Deep Think leads GPQA Diamond (82.4%), MMLU-Pro (89.8%), and HumanEval+ (94.1%). Available now to Google AI Ultra subscribers.
- Gemini 3.5 Pro — five days remain in June. A miss past June 30 marks the second consecutive broken I/O commitment.
- Alphabet stock is still recovering from a 7.2% intraday drop earlier this week following the Shazeer and Jumper departures.
Microsoft / GitHub Copilot
- DeepSeek V4 evaluation for Copilot Cowork — no final decision confirmed. GitHub AWS capacity arrangement ongoing.
DeepSeek / Alibaba Qwen / Z.ai
- GLM-5.2 and MiniMax M2.5 continue to absorb displaced Fable 5 enterprise traffic.
Cohere / Aleph Alpha
- $20B merger pending regulatory approval.
Presight AI
-
Banco Santander and Presight sign memorandum of understanding to explore
strategic cooperation in artificial intelligence.
-
Abu Dhabi Chamber and Presight — a G42 company — sign a strategic
partnership to deploy agentic AI across 102,000 SMEs.
That's your AI world for Thursday. Back tomorrow. — Aaron
Aaron Rose is a software engineer and technology writer at tech-reader.blog.
Catch up on the latest explainer videos, podcasts, and industry discussions below.
