OpenAI Unveils GPT-5.6 Trio — But the Government Gets First Dibs
OpenAI announced three new GPT-5.6 models on June 26, but in an unprecedented move, the company is complying with a U.S. government request to limit initial access to a small group of trusted partners before a broader rollout.
The three models form a tiered lineup: Sol, the flagship with best-in-class agentic capabilities in coding, biology, and cybersecurity; Terra, a balanced everyday model matching GPT-5.5 performance at half the cost; and Luna, a fast, affordable option for high-volume use cases.
The restriction stems from a June 2 executive order by President Trump titled "Promoting Advanced Artificial Intelligence Innovation and Security," which established a voluntary framework for developers to share advanced models with the government for national security assessments up to 30 days before public release.
"We don't believe this kind of government access process should become the long-term default," OpenAI stated, while confirming it is working with the Trump administration to establish a repeatable framework for future releases. General availability is expected in the coming weeks.
The move marks a significant shift in how frontier AI models reach the market, adding a national security gatekeeping layer that didn't exist a year ago. Combined with the Fable 5 export ban that pulled Anthropic's most capable models offline for all foreign nationals earlier this month, a clear pattern is emerging: the era of unrestricted global access to cutting-edge AI is ending.
Google DeepMind's Brain Drain: Four Senior Researchers Leave in Six Days
In what may be the most consequential week of talent movement in AI history, four senior Google DeepMind researchers departed for rivals between June 18 and June 24, triggering a $269 billion wipeout in Alphabet's market capitalization.
The departures read like an AI hall of fame: Noam Shazeer, co-author of the landmark 2017 "Attention Is All You Need" paper that introduced the transformer architecture, left for OpenAI on June 18. Two days later, John Jumper, who shared the 2024 Nobel Prize in Chemistry for his work on AlphaFold, announced he was joining Anthropic. Then on June 24, Jonas Adler (Gemini AI coding lead) and Alexander Pritzel (Gemini pretraining specialist and AlphaFold contributor) both followed Jumper to Anthropic.
The market reaction was swift and severe — Alphabet shares dropped over 5% on the news, reflecting a brutal truth: when you're spending $190 billion on AI infrastructure but losing the people who make that infrastructure produce frontier capabilities, investors question the sustainability of your position.
The financial lure of upcoming IPOs at both Anthropic and OpenAI — targeting October and September respectively — is providing gravitational pull that established public companies struggle to match. As Fortune reported, human capital has become a market-priced asset in AI, with individual researchers commanding measurable influence over company valuations.
Anthropic Exposes Alibaba's Massive Distillation Campaign
Anthropic sent a formal complaint to U.S. Senators on June 24 accusing Alibaba Group and its Qwen AI lab of orchestrating the largest known distillation attack against Claude — an operation dwarfing previous campaigns by other Chinese AI companies.
The numbers are staggering: approximately 25,000 fraudulent accounts generated over 28.8 million exchanges with Claude between April 22 and June 5, 2026, specifically targeting the model's software-engineering and agentic-reasoning capabilities. For context, earlier campaigns by DeepSeek, MiniMax, and Moonshot AI combined had generated about 16 million exchanges across 24,000 fake accounts — the Alibaba-linked operation surpassed all three put together.
Distillation — training a less capable model on outputs from a more powerful one — has become a flashpoint in AI intellectual property disputes. The disclosure is already driving legislative action, with Senators Bill Hagerty (R-TN) and Andy Kim (D-NJ) pushing an amendment to defense legislation that would blacklist or sanction entities conducting such campaigns.
Alibaba has denied the allegations. But regardless of the legal outcome, the incident highlights a growing tension: as AI models become more expensive to train from scratch, the incentive to extract capabilities from competitors' production systems grows proportionally.
OpenAI and Broadcom Unveil Jalapeño: A Custom AI Chip Built in Record Time
OpenAI is no longer content being a customer in the chip market. On June 24, the company and Broadcom unveiled Jalapeño, OpenAI's first custom AI inference processor — and they built it in a remarkable nine months, potentially the fastest ASIC development cycle ever achieved in high-performance semiconductors.
Jalapeño was designed from the ground up for LLM inference, optimized around the specific kernels, memory movement patterns, and serving architectures that matter most for frontier AI models. Early testing shows performance-per-watt substantially better than current state-of-the-art solutions, and the chip is reportedly 50% cheaper than comparable Nvidia GPUs for inference workloads.
Perhaps most notable is that OpenAI used its own AI models to accelerate parts of the chip design and optimization process — a fitting demonstration of AI's potential to compress traditionally years-long hardware development timelines.
Deployment at gigawatt-scale data centers with Microsoft and other partners is planned for later in 2026. The move puts OpenAI in the company of Google (which has its TPU line) and Amazon (with its Trainium and Inferentia chips) in the race to reduce dependence on Nvidia's dominant GPU platform.
The Week Ahead: Model Delays, IPO Preparations, and the $452 Billion Question
Looking forward, every major frontier model that was expected in June — GPT-5.6 (general access), Gemini 3.5 Pro, and Grok 5 — has slipped into July. Google's Gemini 3.5 Pro, with its 2-million-token context window and Deep Think reasoning mode, remains stuck in limited enterprise preview after missing its second consecutive I/O commitment.
On the business side, the looming IPOs of OpenAI (September) and Anthropic (October) are creating interesting dynamics: both companies filed S-1s projecting profitability, but a potential pricing war between them — with Chinese model alternatives like DeepSeek V4-Pro offering tokens at a fraction of the cost — threatens those projections during the critical roadshow period.
Meanwhile, the combined AI infrastructure spending by Microsoft, Alphabet, Amazon, and Meta now exceeds $452 billion annually, raising persistent questions about return on investment. Microsoft generated $37 billion in AI services revenue against $97 billion in cumulative spending — numbers that suggest the industry is still betting heavily on future demand rather than current monetization.