AI News Weekly: June 22 – June 29, 2026

Table of Contents

• OpenAI Unveils GPT-5.6 Family and First Custom Chip
• Anthropic Exposes Alibaba's 28.8 Million Exchange Distillation Attack
• GPT-5.5-Cyber and Patch the Planet: OpenAI's Offensive Security Push
• Google DeepMind Talent Exodus: Shazeer and Jumper Depart
• SpaceX Signs $6.3 Billion Compute Deal with Reflection AI
• Colorado AI Act Takes Effect: First US State AI Law
• Fable 5 Ban Enters Day 14: Governance Questions Emerge
• Gemini 3.5 Pro Delayed to July
• AI-Driven Layoffs Reach 142,000 in 2026
• Analysis and Forward-Looking Insights
• What to Watch Next

---

OpenAI Unveils GPT-5.6 Family and First Custom Chip

OpenAI had its most aggressive week of announcements in months, launching both a new model family and its first custom-built silicon.

GPT-5.6 Sol, Terra, and Luna were announced on June 26 as a limited preview restricted to trusted partners and organizations. The three-tier lineup positions Sol as the flagship model, Terra as a balanced model for everyday production work, and Luna as a fast, affordable option. Pricing during preview: Sol at $5/$30 per million input/output tokens, Terra at $2.50/$15, and Luna at an even lower tier. The models are not yet available in ChatGPT and are accessible only through the OpenAI API and Codex to a curated group. OpenAI stated general availability would follow "in the coming weeks."

The launch was notably constrained by US government request. OpenAI limited initial access to "trusted partners" at the direction of the Trump administration, echoing the restricted rollout of Anthropic's Mythos. This marks a new normal for frontier model releases: even OpenAI, previously the most open about deployment, is now subject to government gating.

Source: OpenAI — Previewing GPT-5.6 Sol | CNBC — OpenAI limits new AI models to 'trusted partners' at request of US government | The Guardian — OpenAI staggers AI model release after Trump administration request

Jalapeño, OpenAI's first custom AI inference chip, was unveiled jointly with Broadcom on June 24–25. Designed specifically for LLM inference (not training), Jalapeño represents OpenAI's answer to the structural cost disadvantage of depending entirely on Nvidia GPUs while competitors like Google (TPUs), Amazon (Trainium), and Microsoft (Maia) have run custom silicon for years.

The chip was designed from concept to manufacturing tape-out in just nine months — which OpenAI calls the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. Greg Brockman told CNBC that OpenAI's own AI models accelerated parts of the design and optimization process, creating a recursive loop where the models help design the chips that will run future versions of themselves.

Initial deployment is targeted for late 2026, with scale-up in 2027 and full production ramp in the first half of 2028. Broadcom CEO Hock Tan described Jalapeño as the first chip in a multi-generation roadmap for gigawatt-scale AI data centers that OpenAI and Microsoft are building together.

Source: OpenAI — OpenAI and Broadcom unveil Jalapeño inference chip | TechCrunch — OpenAI unveils its first custom chip, built by Broadcom | Broadcom Investor Relations

---

Anthropic Exposes Alibaba's 28.8 Million Exchange Distillation Attack

In one of the most significant geopolitical stories of the week, Anthropic sent a letter to US Senators Tim Scott and Elizabeth Warren on June 10 (first reported by Bloomberg and confirmed by CNBC on June 24–25) accusing Alibaba and its Qwen AI lab of executing "the largest known distillation attack on Anthropic to date."

The operation ran from April 22 to June 5, 2026, using approximately 25,000 fraudulent accounts to generate 28.8 million exchanges with Claude models over six weeks. No passwords were stolen and no firewalls were breached — the attackers used Claude exactly as ordinary users would, just at industrial scale. The campaign specifically targeted Claude's agentic reasoning, software engineering proficiency, and long-horizon task completion capabilities — the exact capabilities that distinguish Claude Opus 4.8 and Fable 5 from other frontier models.

This is not Anthropic's first distillation complaint. In February 2026, the company publicly named DeepSeek, Moonshot, and MiniMax for similar operations involving 24,000 accounts and 16 million combined exchanges. Alibaba's operation is larger than all three combined.

The geopolitical dimension is critical: Anthropic's letter directly connects this campaign to the June 12 export control ban on Fable 5 and Mythos 5. The argument is that when Chinese labs appear to rapidly close the capability gap with US frontier models, US policymakers assume export controls on advanced chips are not working. If that apparent convergence is built on extracted Claude capabilities rather than independent innovation, the chip controls may actually be more effective than they appear.

Alibaba has not responded to requests for comment from CNBC, Bloomberg, or other outlets.

Source: CNBC — Anthropic accuses Alibaba of campaign to 'brazenly' and 'illicitly' extract AI capabilities | BBC — Anthropic accuses Chinese rival Alibaba of illicitly extracting AI capabilities | Ars Technica — Anthropic says Alibaba must be punished for largest Claude cloning attack | Forbes — Anthropic Says Alibaba Used 25,000 Fake Accounts To Distill Claude

---

GPT-5.5-Cyber and Patch the Planet: OpenAI's Offensive Security Push

On June 22, OpenAI released the full version of GPT-5.5-Cyber as the centerpiece of its expanded Daybreak cybersecurity initiative. The model scored 85.6% on CyberGym — the highest single-model result ever recorded on that benchmark, ahead of standard GPT-5.5 at 81.8% and Anthropic's Mythos 5 at 83.8%.

The irony is difficult to ignore: Mythos 5, the model whose offensive capabilities triggered the US government export ban, now has a rival that outscores it on the exact benchmark designed to measure those capabilities. OpenAI is deploying that capability as a defender tool, which is the strategic framing distinction that matters here.

GPT-5.5-Cyber is built on GPT-5.5 with additional tuning for the full defensive cybersecurity loop: reading large codebases, tracing attack paths, validating vulnerabilities in a controlled sandbox, writing patches, and testing those patches before passing them to human reviewers. The model also scored 39.5% on ExploitGym (vs. 25.95% for GPT-5.5) and 69.8% on SEC-bench Pro (vs. 63.1% for GPT-5.5).

Access is gated through OpenAI's Trusted Access for Cyber program, available to vetted security organizations including Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, and Zscaler.

Patch the Planet, launched alongside GPT-5.5-Cyber, partners OpenAI with Trail of Bits, HackerOne, and independent researchers to use Codex and GPT-5.5-Cyber to find and fix vulnerabilities in widely used open-source infrastructure. More than 30 projects have committed, including cURL, Go, Python, Sigstore, and pyca/cryptography. In the Linux kernel, the model identified security-relevant components across more than 30 million lines of code and generated 8 kernel pointer information-leak proof-of-concepts and 24 local privilege escalation exploits. In Chrome's V8 engine, five exploitable bugs were found. In Safari's WebKit, more than 10 vulnerabilities were identified.

Source: OpenAI — Daybreak: Tools for securing every organization in the world | WIRED — OpenAI launches full-scale effort to patch open-source bugs | Axios — OpenAI gives GPT-5.5-Cyber more powerful cybersecurity capabilities

---

Google DeepMind Talent Exodus: Shazeer and Jumper Depart

The biggest individual talent moves in the AI industry happened within days of each other in mid-June, dealing a severe blow to Google DeepMind's position as the world's leading AI research lab.

Noam Shazeer, co-author of the foundational 2017 "Attention Is All You Need" paper that introduced the Transformer architecture and co-lead of Google's Gemini models, announced on June 18, 2026 that he was leaving Google to join OpenAI. Google had spent $2.7 billion to acquire Character.AI assets and bring Shazeer and his research team back from his startup just two years earlier.

John Jumper, the 2024 Nobel Prize in Chemistry winner (the youngest chemistry laureate in more than 70 years) and the architect of AlphaFold2 — the AI system that predicts protein structures and has been used by more than two million scientists across 190 countries — announced on June 19 that he was leaving Google DeepMind after nearly nine years to join Anthropic.

Demis Hassabis responded publicly: "What we achieved with AlphaFold changed the world, and showed the field what was possible with AI for science and medicine, lighting the way for how AI can benefit humanity."

According to SignalFire's 2025 State of Talent Report, engineers at DeepMind were nearly eleven times more likely to leave for Anthropic than the reverse. The combined departures wiped approximately $270 billion from Alphabet's market cap in a single week.

Two more Gemini researchers were subsequently reported joining Anthropic on June 26, making it four senior Google exits in six days.

Source: Fortune — As top talent leaves Google DeepMind, some question if the lab can remain at the forefront | Business Insider — The AI Talent Wars Just Heated up Again | Axios — Google DeepMind loses star power

---

SpaceX Signs $6.3 Billion Compute Deal with Reflection AI

SpaceX signed a compute lease agreement with Reflection AI on June 22, 2026, at $150 million per month starting July 1, 2026, through the end of 2029 — totaling approximately $6.3 billion if the contract runs its full term. Either party can exit with 90 days' notice after the initial three months.

Reflection AI, co-founded by former Google DeepMind researchers Misha Laskin (who led reward modelling for Gemini) and Ioannis Antonoglou (DeepMind's sixth-ever researcher and co-creator of AlphaGo), is valued at $25 billion with backing from Nvidia, Sequoia, and Lightspeed. Critically, the company has not yet released a public frontier model.

The deal provides access to Nvidia GB300 chips at SpaceX's Colossus 2 data center in Memphis, Tennessee. SpaceX's Colossus now has committed compute revenues exceeding $80 billion through 2029 across Anthropic, Google, Cursor (which SpaceX is acquiring), and Reflection.

Reflection's thesis positions it as the "third option" in frontier AI: American, open-weight, and frontier-scale — a model whose weights are publicly available for inspection and self-hosting, built by researchers with provable track records, and backed by enough compute to train at frontier scale.

Source: CNBC — SpaceX signs computing power deal with open-source AI startup Reflection worth up to $6.3 billion | Bloomberg — SpaceX Signs Multibillion-Dollar AI Computing Deal With Reflection | Forbes — SpaceX's Colossus Lands $6.3 Billion Compute Deal With Reflection AI

---

Colorado AI Act Takes Effect: First US State AI Law

The Colorado Artificial Intelligence Act (SB 24-205) took effect on Monday, June 30, 2026, becoming the first comprehensive state AI law in the United States to actually go into force. The law was originally set to take effect on February 1, 2026, but was delayed to June 30 via subsequent legislation (SB 25B-004).

The law regulates high-risk AI systems used in consequential decisions affecting employment, education, housing, healthcare, financial services, government services, insurance, and legal services for Colorado residents. It requires developers of "high-risk" AI systems to conduct impact assessments, disclose material risks, and implement risk management programs.

The law's journey has been turbulent. On May 14, 2026, Governor Jared Polis signed SB 189 ("Automated Decision-Making Technology"), which repealed and replaced the original Colorado AI Act with a narrower framework just six weeks before the June 30 effective date. The revised version removed some of the more controversial provisions, including mandatory bias audits, while retaining core requirements for transparency and accountability in automated decision-making.

Despite the watered-down version, the law represents a significant precedent: it is the first time a US state has enacted enforceable requirements for AI systems in high-stakes domains, and it will serve as a template for other states considering similar legislation.

Source: TechTimes — Colorado AI Law Hits June 30 Deadline Without Bias Audits | AI Laws by State — Colorado's law, effective June 30, 2026 | ComplianceHub.Wiki — Colorado Hits Reset: How SB 189 Repealed the Nation's First AI Act

---

Fable 5 Ban Enters Day 14: Governance Questions Emerge

Claude Fable 5 and Mythos 5 remain offline as of June 26, fourteen days after the US Commerce Department's export control directive. API calls to claude-fable-5 continue to return errors, and no official restoration date exists.

The narrative around the ban has shifted considerably. Anthropic initially framed it as a jailbreak issue it expected to resolve within days. The NSA Director's Senate testimony changed that framing: the concern is now understood to be Mythos's autonomous offensive cybersecurity capability, not a patchable vulnerability.

Two key dates are emerging:

• July 8, 2026: Anthropic's updated privacy policy takes effect, requiring government-issued ID and biometric verification via Persona (a Peter Thiel-backed identity platform). This is widely understood as the mechanism for restoring Fable 5 to verified US citizens without requiring the export control directive to be fully lifted.
• August 1, 2026: The 60-day window expires under the June 2 Executive Order for NSA, Treasury, and CISA to build a classified benchmarking process and voluntary pre-release framework for covered frontier models.

On June 25, Reuters and AP confirmed that the NSA testing that informed the ban took place under Project Glasswing, Anthropic's restricted program for government and security partners. An unidentified US official told AP that Mythos identified vulnerabilities in hours but did not necessarily exploit them — a significant distinction from the earlier "breached classified systems" framing.

Source: Reuters AI News | AP News (via Reuters)

---

Gemini 3.5 Pro Delayed to July

Google has quietly pushed the general availability of Gemini 3.5 Pro from June to July 2026. The model was announced at Google I/O on May 19, where CEO Sundar Pichai committed to a June general availability date. The prediction market probability of a June 30 launch was tracking at approximately 4.5% as of June 26, down sharply from 50% earlier in the week.

The reported reason for the delay is that early testers flagged issues with token efficiency and long-horizon task performance. Google is reviewing feedback on how Gemini 3.5 Pro handles extended reasoning chains and complex multi-stage tasks before committing to a general release.

The confirmed specifications remain: a 2-million-token context window, a Deep Think reasoning mode gated to the $250-per-month Ultra tier, and frontier multimodal capability. But missing a CEO-committed June deadline after an audience of developers already groaned at the "next month" deferral at I/O creates a credibility problem.

Source: Analytics Insight (via BuildFastWithAI) | Polymarket prediction data

---

AI-Driven Layoffs Reach 142,000 in 2026

Tech layoffs in 2026 have reached 142,000 as profitable companies cut headcount to fund AI infrastructure investments. The trend has accelerated through June, with companies explicitly citing AI as the reason for workforce reductions.

Notable cuts this week include:

• Oracle laid off 21,000 employees (about 13% of its workforce) over the past 12 months, spending $1.84 billion on severance and restructuring costs. The company explicitly cited AI adoption and costly AI infrastructure ambitions.
• GitLab laid off roughly 350 workers (about 14% of its staff) to fund AI infrastructure investment and handle surging traffic from AI workflows.
• Meta plans to lay off 8,000 employees with additional cuts to follow in the second half of 2026.

Morgan Stanley projected global AI-linked debt to nearly double to $570 billion in 2026, driven by the scale of AI infrastructure financing across hyperscalers and frontier AI labs. The Stanford 2026 AI Index revealed that entry-level software developer jobs decreased by nearly 20%, as automation of boilerplate code reduces junior hires.

Source: TechCrunch — The running list: major tech layoffs in 2026 where employers cited AI | BBC — Tech giant Oracle cuts 21,000 jobs as it embraces AI | Business Insider — 16 companies, including Oracle and GitLab, that have said they're doing AI-related layoffs

---

Analysis and Forward-Looking Insights

This week reveals three converging trends that will define the AI industry in the second half of 2026:

1. The full-stack race is accelerating. OpenAI's Jalapeño chip is the clearest signal yet that frontier AI labs can no longer afford to be pure software companies. Every major player — Google with TPUs, Amazon with Trainium, Microsoft with Maia — has custom silicon. OpenAI is finally catching up, and the nine-month design cycle (accelerated by AI itself) suggests this gap will close faster than traditional semiconductor timelines would predict. The implication for developers: inference costs should drop meaningfully by 2028, but the transition period will be messy.

2. Government gating is the new normal. Both OpenAI's GPT-5.6 limited preview and Anthropic's Fable 5 ban demonstrate that frontier model releases are now subject to government review. The June 2 Executive Order's August 1 deadline for a voluntary pre-release framework will likely become the de facto standard for all major model launches. For enterprises, this means longer lead times for accessing the latest capabilities and the need to plan for model availability uncertainty.

3. The talent war is shifting from recruitment to retention. Google DeepMind losing both the Transformer co-author and the Nobel Prize-winning AlphaFold architect in the same week is not a personnel issue — it is a strategic vulnerability. Anthropic's 80% two-year retention rate (leading every frontier lab) and Alphabet's 14% ownership stake in Anthropic create a paradoxical situation where Google is indirectly funding the company that is poaching its best researchers. This dynamic will only intensify as the talent pool for truly world-class AI researchers remains extremely small.

For developers: The GPT-5.6 tiered pricing (Sol/Terra/Luna) signals a maturing model marketplace where cost-performance trade-offs matter more than raw capability. The Patch the Planet initiative also demonstrates that AI-assisted security work is moving from research to production, creating new demand for developers who can work at the intersection of AI and cybersecurity.

For enterprises: The Colorado AI Act's implementation (even in its watered-down form) is the first test case for state-level AI regulation. Expect other states to follow, and prepare for a patchwork of requirements that will vary by jurisdiction. The 142,000 tech layoffs also signal that AI investment is a zero-sum game for many companies: every dollar spent on infrastructure is a dollar not spent on headcount.

For policymakers: The Alibaba distillation attack exposes a fundamental gap in AI governance: the ability to systematically extract model capabilities through legitimate API access at industrial scale. Anthropic's letter to Congress asking to criminalize distillation will likely spark a difficult debate about the line between competitive intelligence and IP theft in the AI era.

---

What to Watch Next

• July 8, 2026: Anthropic's ID verification policy takes effect. Will Fable 5 be restored for verified US citizens?
• August 1, 2026: The 60-day deadline for the government's voluntary pre-release framework for frontier models.
• GPT-5.6 general availability: OpenAI said "in the coming weeks" — expect a broad rollout in early July.
• Gemini 3.5 Pro: Google needs to announce a specific July date to repair credibility.
• Reflection AI's first model: A $25 billion valuation with no public product is a large bet. The next 12 months will answer whether the thesis holds.
• Congressional response to distillation: Whether the US government criminalizes model distillation or treats it as a civil matter will reshape how Chinese AI labs operate.
• Jalapeño benchmarks: OpenAI has not yet released final performance-per-watt numbers. Production-scale results in 2027–2028 will determine whether the chip delivers on its promise.
• Microsoft MAI models: The seven-model suite announced on June 2 (including MAI-Thinking-1 and MAI-Code-1-Flash) is expanding to third-party platforms including OpenRouter, Baseten, and Fireworks AI — watch for competitive pricing pressure.

---

Report compiled June 29, 2026. All links verified against primary sources. Stories sourced from OpenAI, Anthropic, Broadcom, CNBC, Bloomberg, Reuters, BBC, TechCrunch, WIRED, Axios, Fortune, Forbes, and The Guardian.