Top 10 AI News: June 26 2026 Daily Roundup

AI News Today June 26 2026: Top 10 Stories

OpenAI unveiled its first custom chip. Alibaba harvested 28.8 million Claude interactions using 25,000 fake accounts to train its own model. And Gemini 3.5 Pro, which Google's CEO promised in June, has been quietly delayed to July.

Today is the last Thursday of June 2026. The Colorado AI Act takes effect on Monday. SK Hynix is filing for a $29 billion Nasdaq listing next month. Alphabet just joined the Dow Jones Industrial Average. And Fable 5 is still offline. There is a lot to track. Here are the 10 stories every AI learner needs to know.

1. OpenAI and Broadcom Unveil Jalapeño: The First OpenAI Custom Chip

OpenAI and Broadcom unveiled Jalapeño on June 25, 2026, OpenAI's first custom AI inference chip and the first tangible output of the partnership the two companies announced in October 2025. The chip was delivered physically to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom CEO Hock Tan and President Charlie Kawwas.

Jalapeño is specifically designed for inference, the process of running a trained AI model to generate responses to users. It is not a training chip. OpenAI's AI models have been entirely dependent on Nvidia GPUs for inference up to this point, putting the company at a structural cost disadvantage compared to Google (which uses TPUs), Amazon (Trainium), and Microsoft (Maia). Every major cloud provider that competes with OpenAI has been running custom inference silicon for years. Jalapeño is OpenAI's answer to that gap.

Built in Nine Months with AI Help

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: "The degree to which our models have been able to accelerate it was very surprising to us." OpenAI's models are helping design the chips that will run future versions of those same models. That loop is genuinely interesting.

Early testing shows Jalapeño will deliver substantially better performance per watt than current Nvidia alternatives for inference workloads, though OpenAI has not yet released final benchmark numbers. Initial deployment is targeted for the end of 2026, with scale-up in 2027 and full production ramp in the first half of 2028. Broadcom CEO Hock Tan said Jalapeño is the first chip in a multi-generation roadmap designed for gigawatt-scale AI data centers that OpenAI and Microsoft are building together.

OpenAI still depends on Nvidia for training runs, which are far more compute-intensive. But inference is where the day-to-day cost of serving ChatGPT and Codex to hundreds of millions of users accumulates. Reducing inference cost per token is directly connected to OpenAI's path to profitability, which matters given its IPO timeline.

My take: Nine months from design to tape-out is genuinely fast. If the performance-per-watt numbers hold at production scale, this is a meaningful structural improvement for OpenAI's economics. The full impact will not be visible until 2028. The story right now is that OpenAI is serious about owning its stack, not just renting it.

2. Anthropic Accuses Alibaba of 28.8 Million Claude Distillation Attacks

Anthropic sent a letter to US Senators Tim Scott and Elizabeth Warren on June 10, 2026, accusing Alibaba and its Qwen AI lab of running what it calls "the largest known distillation attack on Anthropic to date." The letter, first reported by Bloomberg and confirmed by CNBC on June 25, became public this week.

Distillation is an AI training technique where a company sends millions of carefully crafted prompts to a rival's model, collects all the outputs, and uses that data to train its own model. No passwords were stolen. No firewalls were breached. The attackers used Claude exactly as an ordinary user would, just through 25,000 fraudulent accounts over six weeks, running 28.8 million exchanges between April 22 and June 5, 2026.

What Alibaba Was Targeting

According to the Wall Street Journal's reporting on the letter, the specific Claude capabilities Alibaba's campaign sought to extract were agentic reasoning, software engineering proficiency, and long-horizon task completion. Those are precisely the capabilities that distinguish Claude Opus 4.8 and the now-offline Fable 5 from most other frontier models. Anthropic also said the campaign was designed to help Alibaba's Qwen model approach Mythos Preview capabilities, the most restricted version of Anthropic's technology.

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

The geopolitical dimension is the part most commentary has underweighted. Anthropic's letter directly connects this distillation campaign to the June 12 export control ban on Fable 5 and Mythos 5. The argument: 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 look. The distillation attack is what makes the gap seem smaller than it is.

Alibaba did not respond to requests for comment from CNBC, Bloomberg, or other outlets. Alibaba is also fighting a separate federal lawsuit against the Pentagon to remove itself from the 1260H military companies list.

My take: The mechanics of what Alibaba did are technically legal under most frameworks, which is exactly why Anthropic is asking Congress to criminalize it. 28.8 million exchanges over six weeks is not an accident or a coincidence. That is a systematic program, and one that Anthropic says continued even after the White House issued a memo in April warning foreign entities to stop.

3. Fable 5 Ban: Day 14, Anthropic Staff Confirm Zero Traffic

Claude Fable 5 and Mythos 5 remain offline on June 26, 2026, fourteen days after the US Commerce Department's export control directive. As of this morning, API calls to claude-fable-5 still return errors. No official restoration date exists.

On June 25, 2026, viral posts on X claimed that users of Claude Code v2.1.190 could access Fable 5. Anthropic staff responded directly and specifically. Sam McAllister, writing as @sammcallister, stated: "We are currently serving exactly 0 traffic to Fable 5." Amol Avasare, Anthropic's Head of Growth, described the access reports as categorically false. The likely explanation for what users were seeing: a front-end UI bug showing Fable 5 in the historical model picker, where selecting it produces a "Claude Fable 5 is currently unavailable" message rather than any actual response.

The July 8 and August 1 Deadlines

The most concrete near-term dates to watch are July 8 and August 1. Anthropic's updated privacy policy, requiring government-issued ID and biometric verification via Persona (a Peter Thiel-backed identity platform), takes effect July 8. 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. International users would remain on Claude Opus 4.8 under that scenario.

August 1 is when 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. Anthropic's structural path back into the government's good standing involves agreeing to that framework for future model releases. Whether it also covers restoration of existing models is the open question.

Also 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. Critically, 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 that had been circulating.

My take: Fourteen days in, I think the restoration question has become secondary to the governance question. The export control ban is less a product decision and more a preview of what frontier AI regulation looks like when there is no established process for it. That matters for every AI lab, not just Anthropic.

4. Gemini 3.5 Pro Delayed to July, Google Needs to Refine Long-Task Performance

Google has quietly pushed the general availability of Gemini 3.5 Pro from June to July 2026, according to insider reports covered by Analytics Insight and prediction market data from Polymarket. The official 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. According to Analytics Insight's coverage, 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. Google declined to comment on the revised timeline.

Gemini 3.5 Pro was announced at Google I/O on May 19, 2026, where CEO Sundar Pichai committed to a June general availability date. That commitment drew audible groans from developers who had expected the model that day. Not shipping in June after a CEO commitment creates a credibility problem that will need to be addressed with a clear July date, not a vague updated window.

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. The competitive context is no longer as favorable as it was two weeks ago. GPT-5.5-Cyber has demonstrated OpenAI's execution cadence. Jalapeño shows OpenAI is building long-term infrastructure. Gemini 3.5 Pro missing June adds to a pattern of announcement ahead of delivery that developer communities are beginning to call out explicitly.

My take: Missing a CEO-committed June deadline is a bigger deal than most Google coverage acknowledges. 'Give us until next month' from a company stage is a promise, not a hedge. The technical reason for the delay sounds legitimate: long-horizon task performance is exactly where you do not want to ship early. But Google needs to say something officially and give a specific July date. Silence makes the credibility gap wider.

5. Colorado AI Act Takes Effect Monday June 30: The First US State AI Law

The Colorado Artificial Intelligence Act takes effect on Monday, June 30, 2026, becoming the first comprehensive state AI law in the United States to actually go into force. 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.

The journey to this point has been turbulent. The law was originally set for February 1, 2026, but a special legislative session in August 2025 extended it to June 30. Then in May 2026, Governor Jared Polis signed SB 189, which amended and narrowed the law substantially, pushing its effective date to January 1, 2027, while scaling back several original requirements. But that amendment was signed on May 14. As of today, June 26, it is the amended version with the January 2027 date that reflects Colorado's current regulatory posture for most covered entities.

What the Amended Law Actually Requires

The original Colorado AI Act required high-risk AI developers and deployers to conduct impact assessments, implement risk management programs, submit annual reports to the Attorney General, and avoid algorithmic discrimination. The amended SB 189 significantly narrowed these requirements, eliminating the duty of care for algorithmic discrimination, removing deployer obligations to maintain risk management programs, and dropping certain reporting mandates. What remains is a transparency-focused framework centered on disclosure requirements when automated decision-making tools are used in consequential decisions.

For businesses: the January 1, 2027 effective date of the amended law is what most compliance teams should be planning toward. The June 30 original effective date is now effectively superseded by the May 2026 amendment for companies in Colorado. The carve-out for algorithmic discrimination liability is the most significant change. Consumer rights groups have criticized the amendment as gutting the original law's protections.

My take: Colorado's AI Act becoming the first US state AI law to go into force, even in significantly amended form, is a landmark. What is more significant for the national picture is what Colorado's quick retreat signals: the EU regulatory model, with its mandatory risk assessments and duty of care, is not going to be the dominant US state AI framework. The US is converging on disclosure and transparency, not substantive risk management. Whether that protects consumers adequately is a separate debate.

6. SK Hynix Plans $29 Billion Nasdaq Listing as Soon as July 10

South Korean chipmaker SK Hynix plans to raise $29 billion through a Nasdaq listing targeting as early as July 10, 2026, according to CNBC reporting. If completed at the target raise, it would be the largest tech IPO since SpaceX's $75 billion listing on June 12, 2026.

SK Hynix is the world's second-largest memory chip manufacturer and the leading supplier of high-bandwidth memory chips (HBM), which are the specialized memory components that Nvidia's H100 and H200 GPUs require for AI training. The company's market cap passed Samsung Electronics earlier in 2026, making it South Korea's most valuable company. According to Reuters, SK Hynix's soaring share price reflects the fundamental shift the company's CEO described: "The emergence of customized AI memory fundamentally changed the industry's economics and allowed SK Hynix to establish itself as the market leader."

The Nasdaq listing, if it proceeds, would make SK Hynix the first major Korean chipmaker to dual-list in the US. It also arrives in the context of Samsung supplying HBM4 memory for OpenAI's Titan chip project, with mass production targeted for late 2026. Both Korean chipmakers are positioning themselves as critical supply chain infrastructure for the AI build-out, and US listings give them direct access to the capital markets where AI infrastructure spending is being priced.

My take: HBM memory is one of the least-discussed but most genuinely critical bottlenecks in AI infrastructure. You cannot run a Nvidia H100 cluster without it. SK Hynix's Nasdaq listing is, in a sense, AI infrastructure investing coming to Main Street. Whether retail investors should own memory chipmakers as an AI play is a separate question I am not qualified to answer, but the strategic logic for the listing is clear.

7. Alphabet Added to the Dow Jones Industrial Average, Replacing Verizon

Alphabet, Google's parent company, has been added to the Dow Jones Industrial Average, replacing Verizon. The change reflects the Dow's periodic rebalancing to ensure the index represents the current state of the US economy rather than its industrial-era composition.

The timing is notable given everything else happening at Alphabet this week. The company lost Noam Shazeer to OpenAI and John Jumper to Anthropic in the same week. Gemini 3.5 Pro has missed its June launch target. Alphabet stock fell approximately 5% on Monday, June 22, 2026, its steepest single-day decline since May 2025, in what analysts attributed directly to the compounding talent departures.

Yet Alphabet being added to the Dow is a recognition of its fundamental position in the US economy. The company's $422 billion in annual revenue includes Google Search, YouTube, Google Cloud, and the Pixel hardware line. Alphabet's 14% stake in Anthropic also means that Google indirectly benefits from every dollar of revenue Claude generates, including the commercial activity of the researchers it just lost.

My take: Joining the Dow is a symbol, not a business result. But it is an interesting week for a symbol. The company is simultaneously being recognized as one of the most important companies in America and losing the architects of its two most significant scientific AI achievements in the same seven-day period. Those two facts can both be true.

8. Qualcomm Reveals Dragonfly C1000 CPU for AI Data Centers, Meta Signs On

Qualcomm announced the Dragonfly C1000 at its shareholder meeting on June 25, 2026: a data center central processing unit built specifically for agentic AI workloads. Meta has signed on to use the Dragonfly C1000 when it starts production in 2028.

The Dragonfly C1000 is built on the open RISC-V instruction set architecture, the same choice as Tenstorrent, the AI chip startup Qualcomm is in acquisition talks with at $8-10 billion. Qualcomm's CEO Cristiano Amon told investors the new CPU targets computing performance without excessive power draw, specifically designed for the kind of persistent, multi-step reasoning loops that agentic AI systems run. Qualcomm also said it has secured two custom chip deals with hyperscalers and acquired Modular, a startup that built software enabling AI applications to run across multiple chip architectures, which Amon described as "equivalent to Nvidia's CUDA."

The financial signal: Qualcomm updated its 2029 non-handset revenue guidance from $22 billion to $40 billion, with $15 billion specifically from data center sales. Qualcomm stock jumped 15% in extended trading on those numbers. The company's primary business remains smartphones, which represented two-thirds of product revenues in the most recent quarter. But the AI data center push is now the company's explicit diversification strategy.

My take: The Meta-Qualcomm deal is the detail that makes this more than an announcement. Meta operates at a scale where it needs hundreds of thousands of chips and has strong incentives to reduce Nvidia dependency. Qualcomm building a CPU (not a GPU) for agentic AI is also interesting: the bet is that the next wave of AI compute is persistent, sequential reasoning rather than massively parallel matrix math, which is a different architecture challenge.

9. Anthropic ID Verification via Persona Goes Live July 8

Anthropic's updated privacy policy, requiring government-issued ID and biometric verification for all Claude users, takes effect July 8, 2026. The verification is handled through Persona, a Peter Thiel-backed identity verification platform that has become the standard provider for fintech and crypto companies requiring KYC (Know Your Customer) compliance.

The rollout requires users to submit a passport, driver's license, or national ID, plus a live selfie. Anthropic will retain this data under its updated retention policy. Critics of the change have raised surveillance concerns, pointing to the involvement of Thiel, a prominent tech investor with ties to both Palantir (a government data analytics company) and the current administration. Supporters note that enterprise-grade identity verification is standard practice for any platform with regulatory obligations, and that the Fable 5 export control situation created exactly the kind of regulatory obligation that requires it.

For most consumers, the July 8 change is the most directly personal AI news of the week. Whether you want to continue using Claude, you will be required to verify your identity. No exceptions are described in the public policy for free-tier users. API users may face different requirements under the developer terms, which Anthropic has not yet detailed separately.

My take: I understand why Anthropic is doing this. The export control directive created a legal obligation to verify who is accessing its models, and Persona is a credible implementation partner. What I find worth watching is how Anthropic communicates the data retention implications to users who have never had to hand over a government ID to use a chatbot before. The gap between 'this is legally necessary' and 'this is what happens to your data' is where trust problems develop.

10. Fable 5 Held a 70% DeepSWE Score Before the Ban, the Highest Ever Recorded

Before the June 12 export control ban pulled it offline, Claude Fable 5 held a 70% PASS@1 score on DeepSWE, the most challenging real-world software engineering benchmark currently in operation, according to Datacurve's verification of the results. That is three points above the second-highest score, held by GPT-5.5.

DeepSWE is different from the older SWE-Bench benchmarks that most model leaderboards use. Where SWE-Bench Pro tests on curated GitHub issues, DeepSWE tests on fresh, real-world software repositories where the problems are not part of any known training set. A 70% score means Fable 5 successfully solved 70 out of 100 novel, real-world programming tasks on its first attempt, without seeing the task before.

The significance of this number has been growing as the ban drags into its second week. The model that was banned, and that Anthropic's own staff this week confirmed is serving exactly zero traffic, was the single best software engineering AI ever tested at the time of its removal. Developers who had pipeline dependencies on Fable 5 are not working around a mediocre model. They are working around a model that, for a brief window of four days, was objectively the most capable AI coding tool available to any developer on earth.

My take: The 70% DeepSWE number is a useful reference point for evaluating everything else in this week's news. The Jalapeño chip is designed to run models like Fable 5 more cheaply. The Alibaba distillation attacks were targeting Fable's agentic and coding capabilities specifically. The NSA testimony was about what Mythos, which shares its architecture with Fable, could do when fully unleashed. All the threads of this week's AI news connect back to what was briefly the most capable AI model ever deployed, and why it is now offline.

Frequently Asked Questions

Q: What is the biggest AI news today, June 26, 2026?

OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom AI inference chip, designed and built in nine months using OpenAI's own models to accelerate the design process. Simultaneously, Anthropic accused Alibaba of running the largest known distillation attack in AI history: 25,000 fraudulent accounts generating 28.8 million Claude interactions between April and June 2026 to train Alibaba's Qwen model.

Q: What is the OpenAI Jalapeño chip?

Jalapeño is OpenAI's first custom-designed AI inference chip, built with Broadcom and unveiled June 25, 2026. It is specifically designed for inference, running trained AI models to serve ChatGPT, Codex, and API users, rather than for training. OpenAI's own AI models helped accelerate the nine-month design cycle. Initial deployment targets end of 2026, with full production scale in early 2028. Early results show substantially better performance per watt than current Nvidia alternatives for inference.

Q: Did Alibaba steal Claude AI data?

Anthropic has accused Alibaba of running the largest known distillation attack on its Claude models. According to a June 10, 2026 letter Anthropic sent to US Senators Tim Scott and Elizabeth Warren, operators affiliated with Alibaba and Alibaba Qwen used approximately 25,000 fraudulent accounts to generate 28.8 million exchanges with Claude between April 22 and June 5, 2026. The goal was to train Alibaba's Qwen model on Claude's outputs. Alibaba did not respond to requests for comment.

Q: Is Fable 5 back online on June 26, 2026?

No. Claude Fable 5 and Mythos 5 remain offline fourteen days after the US export control ban issued June 12, 2026. Anthropic staff confirmed on June 25 that the company is serving exactly zero Fable or Mythos traffic. Viral claims that Claude Code v2.1.190 users could access Fable 5 were confirmed false by Anthropic's Head of Growth. The July 8 ID verification rollout and August 1 EO framework deadline are the next structural dates to watch.

Q: Has Gemini 3.5 Pro been delayed to July?

Yes, according to insider reports and prediction market data. Google has reportedly postponed the general availability of Gemini 3.5 Pro from June to July 2026 to refine token efficiency and long-horizon task performance based on early tester feedback. The prediction market probability of a June 30 launch fell to approximately 4.5% as of June 26. Google has not officially confirmed the delay or announced a new date.

Q: What is the Colorado AI Act and when does it take effect?

The Colorado AI Act is the first comprehensive state AI law in the US, originally enacted in 2024. It regulates high-risk AI systems used in consequential decisions affecting Colorado residents across employment, education, housing, healthcare, and other domains. The original effective date of February 1, 2026 was delayed to June 30, 2026. However, a May 2026 amendment (SB 189) significantly narrowed its requirements and moved the effective date to January 1, 2027. Most businesses should be planning toward the January 2027 timeline.

Q: What is a model distillation attack in AI?

A model distillation attack, also called model extraction, involves sending millions of carefully crafted prompts to a rival AI company's model, collecting all the outputs, and using those outputs as training data for your own model. No system is hacked. No code is stolen. The attacker interacts with the target model like an ordinary user, but at industrial scale with prompts designed to extract its most valuable capabilities. Anthropic accused Alibaba of doing this with 28.8 million Claude interactions via 25,000 fraudulent accounts.

Q: What is Alphabet's addition to the Dow Jones?

Alphabet, Google's parent company, was added to the Dow Jones Industrial Average, replacing Verizon. The change reflects the Dow's periodic rebalancing to keep the index representative of the current US economy. Alphabet has annual revenues of approximately $422 billion across Google Search, YouTube, Google Cloud, and hardware. The addition comes the same week Alphabet stock fell roughly 5% after the departures of Noam Shazeer to OpenAI and John Jumper to Anthropic