AI News Today June 26 2026: 15 Biggest Stories

OpenAI revealed its first custom chip. Anthropic formally accused Alibaba of running 28.8 million fraudulent exchanges against Claude. Two more Gemini researchers announced they are joining Anthropic, making it four senior Google exits in six days. And GPT-5.6 is now a matter of days away. Here are the 15 stories that define June 26, 2026. For continuous coverage of every frontier AI development this month, the AI Industry News and Trends hub at Build Fast with AI is your reference.

 1. OpenAI and Broadcom Unveil Jalapeño: The First Custom OpenAI Inference Chip Targets 50% Cheaper LLM Serving

On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-designed AI chip. Engineering samples were physically delivered to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom President and CEO Hock Tan at OpenAI's San Francisco headquarters. The chip is purpose-built for LLM inference, not training, and was designed from initial concept to manufacturing tape-out in nine months, which OpenAI and Broadcom describe as the fastest such cycle ever for an advanced high-performance chip. OpenAI's own AI models assisted in parts of the design and optimization process. Jalapeño is manufactured by TSMC, with Broadcom supplying silicon implementation and Tomahawk networking connectivity, and Celestica handling board, rack, and system integration. Early lab testing shows approximately 50% lower inference cost per token than current-generation Nvidia GPUs, per Broadcom CEO Hock Tan's statements to Reuters and Bloomberg, with performance matching Nvidia Blackwell and Google TPUs. OpenAI's own announcement is more carefully worded, describing performance per watt as 'substantially better than current state-of-the-art' with a full technical report coming in the coming months. The chip is running ML workloads including an unreleased GPT-5.3-Codex-Spark model in lab testing. Jalapeño is not commercially available. Broadcom expects small prototype data center deployment by end of 2026, with production ramp in 2027 and full scale in the first half of 2028. OpenAI and Broadcom have committed to deploying OpenAI-designed accelerators at 10 gigawatt scale with Microsoft and other partners through 2029. The strategic context: OpenAI spent approximately $14 billion serving ChatGPT in 2025 on third-party GPUs. A 50% inference cost reduction at that scale is a profitability-defining lever, not a marginal engineering win. For the full AI chip landscape, see the AI coding tools hub at Build Fast with AI.

2. Anthropic Accuses Alibaba of the Largest Known Distillation Attack: 25,000 Accounts, 28.8 Million Exchanges

Anthropic sent a letter dated June 10, 2026, to US Senate Banking Committee Chairman Tim Scott and Ranking Member Elizabeth Warren accusing operators affiliated with Alibaba and its Qwen AI lab of conducting the largest known distillation attack on Anthropic to date. The campaign involved approximately 25,000 fraudulent accounts generating more than 28.8 million exchanges with Claude between April 22 and June 5, 2026. The targeted capabilities were specifically the domains where Claude Mythos Preview excels: agentic reasoning, software engineering, and long-horizon task performance. Anthropic alleged the Chinese government was complicit in the effort. The letter's key argument frames the threat in economic and geopolitical terms: 'These distillation attacks turn hundreds of billions of dollars in American investment and R&D into a massive subsidy for our geopolitical competitors.' Anthropic had previously accused DeepSeek, Moonshot, and MiniMax of running a combined 16 million exchanges across 24,000 fraudulent accounts in a separate campaign disclosed in February 2026. This Alibaba campaign represents nearly double that volume. Alibaba has not publicly responded to the accusations. The timing is significant: the letter was sent to Congress two weeks before the US government issued its export control directive pulling down Fable 5 and Mythos 5, and before Anthropic filed its confidential IPO S-1. For enterprise teams, the distillation threat has direct implications: a weaker model trained on a stronger one inherits its capabilities without the governance and safety controls of the original system. For more on what this means for enterprise AI security, see Anthropic's reporting on distillation threats.

3. Four Google DeepMind Researchers Leave in Six Days: Jonas Adler and Alexander Pritzel Join Anthropic

Bloomberg reported on June 24 that two more senior Google DeepMind researchers, Jonas Adler and Alexander Pritzel, are planning to leave Google for Anthropic. The exits bring the total to four senior departures from Google DeepMind's AI team in six days: Noam Shazeer (Transformer co-author, Gemini co-lead) joined OpenAI on June 18; John Jumper (Nobel laureate, AlphaFold lead) joined Anthropic on June 20; Jonas Adler (Google's AI coding lead) is joining Anthropic; and Alexander Pritzel (pretraining specialist, AlphaFold contributor) is joining Anthropic. The Adler-Pritzel pairing is structurally significant: Adler worked on the AI coding effort that directly competes with Claude Code, and Pritzel worked on the pretraining phase where frontier model capability is fundamentally determined. Together with Jumper, they represent the AlphaFold team's core leadership transferring to Anthropic. As one analyst on X summarized: 'Anthropic is 3-for-4 on Google departures this month. Jumper (science), Adler (coding), Pritzel (training). They are not just hiring talent, they are systematically hollowing out Google's Gemini team.' Demis Hassabis told reporters that movement between leading labs is expected and Google still has the largest and broadest research team in the industry. For the competitive context on AI coding tools, the Claude Code vs Codex vs Cursor comparison at Build Fast with AI covers the full developer tooling landscape.

4. Why Google's Internal Compute Politics Are Accelerating the Talent Exodus

Bloomberg's June 24 reporting on the Adler and Pritzel departures included a detail that explains the structural driver of Google's talent crisis: shortly before Noam Shazeer announced his move to OpenAI, computing power dedicated to one of his projects was reassigned to a London-based team at Google DeepMind. The reallocation was intended to improve collaboration and consolidate pretraining work, but the effect from a researcher's perspective was that scarce GPU resources were pulled from active projects. At frontier AI labs, access to compute is not a secondary concern. It is the primary constraint on research velocity. A researcher who cannot run experiments is a researcher who cannot advance. Anthropic and OpenAI are both offering something Google cannot easily match at scale: pre-IPO equity in organizations growing at triple-digit revenue rates plus, critically, a promise of the compute access needed to actually do the work. The pattern repeats: every cited case of a departure includes either a compute allocation frustration or a sense that bureaucratic coordination costs are higher than research productivity gains. Google has the resources to backfill and has weathered defections before. But losing four senior Gemini team members in six days, including the AlphaFold leadership, while Gemini 3.5 Pro misses its public June GA deadline, creates a compounding narrative that is harder to manage than any individual departure.

5. GPT-5.6 Launch Window: Kindle-Alpha RC, 1.5M Context, 83% Polymarket Odds

GPT-5.6 is now the most-tracked AI release in prediction markets, with Polymarket contracts exceeding $1.1 million priced at 83% probability for a launch before June 28, 2026. OpenAI has made no official announcement. The release candidate is internally codenamed 'kindle-alpha,' confirmed by multiple independent developer reports of the model string briefly appearing in Codex backend logs on June 12 before being pulled. The rumored feature set, triangulated from developer testing and leaked internal memos: a 1.5 million token context window (up from 1 million in GPT-5.5), significantly improved UI generation and front-end code output quality, faster Codex response times, and improved long-horizon agentic coding. The deeper structural improvement is the redesigned reward audit pipeline to prevent a recurrence of the Goblin Incident, the documented reward model miscalibration in GPT-5.5 that produced a 175% increase in creature metaphors in outputs. OpenAI's IPO quiet period since its June 8 S-1 filing constrains marketing communications, meaning GPT-5.6 will likely ship as a technical update rather than a headline launch. One important developer note: OpenAI typically updates the gpt-5.5-latest endpoint automatically when GPT-5.6 is promoted. Any production code hardcoded to that endpoint will receive GPT-5.6 behavior without a manual migration. Pin to an explicit versioned endpoint now to control the timing of any behavioral changes.

6. The Era of Cheap AI Ends in 2026: How and Why Frontier Pricing Reversed

The pricing direction that defined the AI market from 2023 to 2025 reversed in 2026. GPT-5.5 doubled GPT-5.4's per-token pricing at launch. Gemini 3.5 Flash costs approximately 3x Gemini 3.1 Flash despite improvements in capability. Claude Fable 5 launched at $10/$50 per million tokens, double Opus 4.8's $5/$25. The underlying economics explain the reversal clearly. OpenAI reportedly spent approximately $1.35 for every dollar it earned in 2025, with ChatGPT serving costs projected at $14 billion in 2026 against $3.7 billion in revenue the prior year. Running frontier models at the prices customers became accustomed to was never sustainable. The first company to prove the model can flip to profitability is Anthropic, which was tracking toward its first operating profit of approximately $559 million in Q2 2026 at an annualized revenue run rate that crossed $47 billion. The one exception to the pricing trend: Chinese open-weight models. DeepSeek made its May promotional 75% price cut on V4-Pro permanent in June, settling at $0.44 per million input tokens and $0.87 per million output tokens. That undercuts GPT-5.5 ($2.50/$15) by more than 5x on input and 17x on output. For developer teams building at scale, the price gap between frontier Western models and Chinese open-weight alternatives has never been larger. The relevant comparison for budget-sensitive pipelines is now DeepSeek V4-Pro or GLM-5.2 via the Z.ai API, not OpenAI vs. Anthropic. For the most current pricing comparison, the best AI models June 2026 guide at Build Fast with AI covers verified pricing across all major models.

7. AI Coding Market Hits $9.3B and Is Growing at 26% Per Year: Who Is Winning

Mordor Intelligence's June 2026 forecast estimates the AI code tools market at $9.3 billion in 2026, growing at approximately 26% annually to roughly $30 billion by 2031. Anthropic leads with Claude Code, which holds approximately 40% of the generative AI coding market. OpenAI's Codex holds approximately 21%. GitHub Copilot, which integrates models from Anthropic, Google, and OpenAI, commands significant enterprise developer reach through Microsoft's distribution. The market structure is unusual: Cursor, which is mid-acquisition by SpaceX, generates approximately $4 billion in annualized revenue with over 50,000 enterprise clients and has reached two-thirds of the Fortune 500. Individual enterprise buyers are deliberately avoiding long-term single-vendor commitments. MongoDB CEO CJ Desai said publicly that his company buys AI coding products one year at a time: 'If Gemini came up with something better, or Codex is better, then I want to be able to use that and not do a long-term commitment.' Analyst Gil Luria at D.A. Davidson framed the competitive dynamic precisely: 'When everyone else was getting distracted by images and videos, Anthropic knew coding was going to drive the performance of models and help with everyone else's tasks, and they're all now pivoting to coding.' The AI coding market is the single most commercially significant battlefield in the AI industry in 2026, and Anthropic's lead is the primary reason the company is tracking toward profitability before OpenAI. See the full breakdown in the AI coding tools comparison at Build Fast with AI.

8. Apple AFM 3 Foundation Models: Five Models, Gemini Distillation, and On-Device AI at Scale

Apple announced its third-generation Apple Foundation Models (AFM 3) at WWDC 2026 on June 8, a family of five models developed in collaboration with Google. The lineup: AFM 3 Core (3B dense on-device model), AFM 3 Core Advanced (20B sparse multimodal on-device model using Apple Research's Instruction-Following Pruning technique that activates only 1 to 4 billion parameters per prompt), AFM 3 Cloud (Private Cloud Compute base model for more demanding requests), AFM 3 Cloud Image (for image generation and editing), and AFM 3 Cloud Pro (for agentic tools and the most demanding tasks, running on Nvidia GPUs hosted in Google Cloud). Apple executives are careful with language: the AFM 3 models were refined using outputs from Google Gemini through a distillation process but are described by Apple as pure Apple technology and code. Gemini is a teacher signal used in post-training refinement, not the runtime model. This is the largest distribution channel to publicly adopt distillation-from-frontier as a building strategy. AFM 3 Cloud Pro runs on Nvidia GPUs in Google Cloud, meaning Apple's most powerful AI system depends on both a competitor's chips and a competitor's infrastructure. Apple has no plans to build its own AI training compute at this scale.

9. Gemini 3.5 Pro Officially Delayed to July: Alphabet Confirms Extended Testing

Alphabet has confirmed that Gemini 3.5 Pro will not achieve public general availability in June 2026, pushing past Sundar Pichai's stated 'give us until next month' commitment at Google I/O on May 19. GuruFocus reporting on June 24 confirmed the delay until July for final adjustments, citing the departure of key researchers and ongoing benchmark gap issues. The model remains in limited Vertex AI enterprise preview. The competitive context of the delay is particularly damaging: Google launched Gemini 2.5 Pro with Deep Think on June 22 to strong reasoning benchmark results (82.4% GPQA Diamond, 89.8% MMLU-Pro), giving itself a positive benchmark narrative, but the June GA miss on 3.5 Pro means Gemini loses momentum entering the final days of June. Polymarket now prices approximately 50 to 55% odds of a June 30 GA, reflecting community skepticism after the missed internal deadline. For Google, the July delay is operationally manageable but narratively costly at the same moment it is defending talent departures and benchmark comparisons.

10. DeepSeek V4-Pro Price Cut Made Permanent at $0.44 per Million Input Tokens

DeepSeek made its May 2026 promotional 75% price cut on V4-Pro permanent in June 2026, settling the model at approximately $0.44 per million input tokens and $0.87 per million output tokens. The lighter V4-Flash variant goes lower still, at $0.14 input and $0.28 output. These prices undercut GPT-5.5 ($2.50/$15) by more than 5x on input and 17x on output, and undercut Gemini 3.5 Flash ($1.50/$9) by roughly 3x on input and 10x on output. DeepSeek's V4 architecture, described in its published technical paper, uses a hybrid sparse-attention design that at 1 million tokens of context runs on roughly 27% of the per-token compute and 10% of the memory of its predecessor, explaining how the company can offer these prices without reporting losses. For enterprise teams that cannot justify frontier Western model pricing for high-volume pipelines, DeepSeek V4-Pro is the benchmark comparison that changes the cost calculus entirely. The permanent price cut is also a strategic signal: DeepSeek is not treating the promotional period as an experiment. It is establishing a price floor for the Chinese open-weight model tier.

11. Anthropic Earlier Caught DeepSeek, Moonshot, and MiniMax in 24,000-Account Attack

The Alibaba distillation accusation disclosed this week is the second major model extraction campaign Anthropic has publicly attributed to Chinese AI labs. In February 2026, Anthropic alleged that three other Chinese AI labs, DeepSeek, Moonshot AI, and MiniMax, collectively ran more than 16 million exchanges with Claude using approximately 24,000 fraudulent accounts. Moonshot and MiniMax accounted for the largest volumes in that campaign. Anthropic described the pattern as becoming increasingly sophisticated over time, with each campaign better at evading detection than the last. The Alibaba campaign represents both a larger scale (28.8 million vs. 16 million exchanges) and a higher-value target (specifically aiming at Mythos Preview capabilities rather than general Claude performance). For context on why this matters beyond intellectual property: Anthropic has argued that the distillation campaigns undermine the safety controls that frontier models are built with. A Qwen model trained on Mythos Preview outputs might approximate Mythos reasoning capabilities without Anthropic's Constitutional AI safety training, anthropic usage policies, or access control infrastructure. That is the specific concern: capability without safeguards.

12. Alibaba Sues US Pentagon Over Chinese Military Company Designation

The distillation attack accusations against Alibaba arrive as Alibaba is simultaneously fighting a US Defense Department designation that classified it as a Chinese military company. The Pentagon added Alibaba to its list of Chinese military companies, a designation the company is contesting in court, saying it has 'no basis in fact or law' and demanding removal. The designation prohibits Alibaba from landing US defense contracts and carries significant reputational damage for enterprise customers evaluating Alibaba Cloud services. The Pentagon's Chinese military company list, as cited in Anthropic's Senate letter, also includes BYD, Baidu, Unitree, and 188 other entities described as directly controlled by the Chinese military. The dual legal and regulatory exposure, a Senate-level distillation accusation plus a Pentagon military designation, creates a compounding reputational and commercial challenge for Alibaba in the US market at precisely the moment its Qwen AI lab is attempting to compete globally with frontier Western models.

13. AI Chip Wars Heat Up: Broadcom Now Designs Custom Silicon for Google, OpenAI, Meta, and ByteDance

The Jalapeño announcement cements Broadcom's position as the kingmaker of custom AI silicon, sitting behind the chip designs of Google (TPUs), OpenAI (Jalapeño), Meta (MTIA accelerators), ByteDance (in active negotiations for custom chips per June 2026 reports), and reportedly Anthropic, Apple, and Fujitsu. The business model is straightforward and enormously profitable: Broadcom provides the silicon engineering, networking connectivity (Tomahawk switching silicon), and manufacturing process expertise that AI labs lack, while the labs provide architectural direction tailored to their specific model serving patterns. For Nvidia, the competitive threat from Broadcom-designed custom ASICs is real but bounded. Nvidia's H100 and B200 GPUs remain dominant for model training, where flexible general-purpose compute outweighs the efficiency gains of inference-optimized ASICs. But inference is where the daily bill lives, and that is exactly where Jalapeño, Google's TPUs, and Amazon's Trainium are gaining ground. Broadcom CEO Hock Tan's public statement to Reuters and Bloomberg, 'you cannot, should not rely on some other third-party GPU to do it for you, because it's such a key part,' is a direct description of the customer insight that has made Broadcom the neutral chipmaking partner for every frontier lab. Broadcom stock gained approximately 2% on the Jalapeño announcement; Nvidia fell 0.26%. The market is pricing Broadcom as the infrastructure beneficiary of the post-GPU AI chip era. See the broader AI infrastructure context in the AI industry news hub at Build Fast with AI.

14. Gemini 3.5 Live Translate: Speech-to-Speech in 70 Languages Rolls Out

Google launched Gemini 3.5 Live Translate, a speech-to-speech model that enables near real-time voice translation in more than 70 languages, rolling out to Pixel 10 and Galaxy S26 devices with Android 17. The feature works at the system level, translating voice calls in real time without routing through a cloud service visible to the user. The 70-language coverage gives Live Translate broader reach than any comparable feature from OpenAI or Anthropic at launch. Apple announced a similar capability in AFM 3 Cloud Pro for iOS 27 Siri, but its language coverage details have not been published. The speech-to-speech translation market has significant commercial implications beyond consumer use: enterprise teams working across language boundaries for support, sales, and international product development represent a high-value segment. The Pixel 10 and Galaxy S26 hardware restriction at launch means Live Translate will reach a fraction of Android 17's installed base initially, but Google has indicated the feature will expand to additional Snapdragon and Tensor chipsets in subsequent quarters.

15. FrontierCode Benchmark: Fable 5 Leads Production-Quality Pull Request Evaluation at 46.3%

Cognition launched FrontierCode, a new coding benchmark designed specifically to evaluate whether AI-generated pull requests are production-quality rather than just functionally correct. Built from 150 original tasks, FrontierCode evaluates AI coding agents on scope control (does the PR stay within its stated boundaries?), regression safety (does the change break existing functionality?), and test quality (are the tests meaningful and complete?). These criteria are designed to reflect what a senior engineer would actually accept in a code review, not just whether the code runs. Fable 5 posted the highest score at 46.3%, versus Claude Opus 4.8 at 34.3% and GPT-5.5 at 25.5%. The 21-point gap between Fable 5 and GPT-5.5 on production-quality code is the widest gap on any publicly available coding benchmark between the two models. For engineering teams that have been debating which AI coding tool to standardize on, FrontierCode provides the most operationally relevant benchmark yet. The limitation to acknowledge: Fable 5 is now behind a usage-credit paywall and has returned from its export control suspension with tighter safety classifiers that increase fallback to Opus 4.8 on certain prompts. For teams using Fable 5 for coding at scale, the net capability available may be somewhat below the benchmark score depending on the coding domains involved. Track the AI coding tools comparison at Build Fast with AI for FrontierCode updates as more models are evaluated.

Frequently Asked Questions

What does Jalapeño mean for Nvidia?

Jalapeño is an inference-focused chip, not a training chip. Nvidia's strongest position is in training, where its H100 and B200 GPUs remain dominant and hardest to displace. For inference, where the daily serving cost lives, Jalapeño, Google's TPUs, Amazon's Trainium, and Microsoft's Maia are all gaining ground. Nvidia fell 0.26% on the Jalapeño announcement, a modest reaction that reflects the market's assessment that this is a meaningful but bounded competitive threat. OpenAI has confirmed Nvidia remains a key partner for training workloads.

Did Alibaba steal Claude's model weights or source code?

No. A distillation attack does not involve obtaining source code, model weights, or training data. It involves generating millions of curated input-output pairs by querying the target model through normal API access, then using those pairs to train a separate model. The attacker replicates behavioral capabilities, not the underlying code. The attack is effective because a model trained on frontier outputs can approach frontier performance on certain tasks without incurring the training cost.

What is Anthropic's ask to Congress on distillation attacks?

Anthropic's Senate letter calls for coordinated action between government and industry to combat distillation attacks. Specifically, Anthropic has been pushing for export controls on AI model access (not just hardware), mandatory screening of high-volume API usage patterns, and coordination between AI labs and government to detect and respond to distillation campaigns. The letter frames distillation as a national security issue, not just an intellectual property issue, because the distilled models can replicate advanced capabilities outside the safety controls and access restrictions of the original system.

When is GPT-5.6 officially launching?

OpenAI has not announced an official launch date as of June 26, 2026. Polymarket prices a June 28 launch at 83% probability based on developer tracking of Codex backend logs, internal memo descriptions, and OpenAI's historical pattern of shipping new models to ChatGPT and Codex first with API access following days to weeks later. OpenAI's IPO quiet period since its June 8 S-1 filing means any launch will be framed as a technical update rather than a marketing event.

What is FrontierCode and why does it matter more than SWE-bench?

FrontierCode, launched by Cognition in June 2026, evaluates AI coding agents specifically on whether their pull requests are production-ready: scope control, regression safety, and test quality. SWE-bench measures whether an AI agent can fix GitHub issues, which is a narrower proxy for coding capability. FrontierCode's criteria map directly to what a senior engineer evaluates in a code review, making it a more operationally relevant benchmark for teams deciding which AI coding tool to deploy in production.

What is Instruction-Following Pruning in Apple AFM 3?

Instruction-Following Pruning (IFP) is a technique developed by Apple Research used in the AFM 3 Core Advanced on-device model. It is a sparse architecture approach that activates only 1 to 4 billion parameters per prompt from a 20 billion parameter total, enabling frontier-class multimodal reasoning on iPhone and Mac hardware without the full inference cost of a dense 20B model. The technique is designed specifically to maximize instruction-following quality per activated parameter, which is different from standard mixture-of-experts sparse architectures that select experts based on input routing.

Why is Anthropic profitable when OpenAI is not?

Anthropic was tracking toward its first operating profit of approximately $559 million in Q2 2026 at an annualized revenue run rate of $47 billion, while OpenAI projects losses of $14 billion for 2026. The primary difference is product mix: Anthropic's Claude Code has achieved approximately 40% of the generative AI coding market, which is the highest-margin segment of enterprise AI. Enterprise coding subscriptions carry higher average revenue per user and lower relative serving cost than consumer chatbot usage. Anthropic's disciplined focus on enterprise and coding, while competitors spread attention across consumer, video, image, and voice products, is the structural reason it reached profitability first.

Recommended Blogs

• AI News Today June 25 2026: Gemini 2.5 Pro Deep Think, GPT-5.6 ...
• AI News Today June 24 2026: GPT-5.5-Cyber Launches, SpaceX ....
• AI News Today June 23 2026: Fable 5 Billing Cliff, SpaceX Cursor..
• Best AI Models June 2026: Full Ranked Leaderboard
• Claude Code vs Codex vs Cursor: AI Coding Tools 2026 Full Comparison
• AI Industry News and Trends Hub: Running Coverage of 2026

Resources and Community

Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you are a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.

• Website — buildfastwithai.com
• LinkedIn — Build Fast with AI
• Instagram — @buildfastwithai
• Founder Twitter — @satvikps
• Twitter — @BuildFastWithAI

Agentic AI Launchpad 2026

A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.

Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026

Free AI Resources

Access free tools, workshops, and micro-learning to keep building:

• AI Workshops — Free resources, upcoming events and past recordings
• Unrot — Learn AI in 5 minutes a day (free micro-learning app)

Subscribe to the Build Fast with AI newsletter and follow @BuildFastWithAI on X for sourced daily coverage of everything moving the AI frontier.

References

• TechCrunch — OpenAI Unveils Its First Custom Chip Built by Broadcom
• VentureBeat — OpenAI Unveils First Custom AI Inference Chip
• TechTimes — OpenAI's First Custom AI Chip Targets 50% ....
• The Next Web — OpenAI's Jalapeño Chip: A Way Out From Nvidia
• CNBC — Anthropic Accuses Alibaba of Distillation Campaign
• Tom's Hardware — Anthropic Claims Alibaba Used 25,000 Fake.....
• TechCrunch — AI Researchers Continue to Leave Google for Its Rivals
• Bloomberg — Google Poised to Lose Two More High-Profile
• The Next Web — Two More Gemini Researchers Are Leaving Google
• InfoWorld — Anthropic Accuses Alibaba of Using 25,000 Fake Accounts
• XDA Developers — Google's Gemini 3.5 Flash Costs 3x the Model
• AppleInsider — Apple's New Foundation Models
• CNBC — Microsoft and Google Take On Anthropic
• Build Fast with AI — Best AI Models June 2026

Build Fast with AI — AI News Today June 25 2026