OpenAI has launched its next-generation GPT-5.6 series, including flagship Sol, budget Terra, and efficient Luna models, with a new ultra mode introduced. Sol achieves 91.9% on TerminalBench 2.1, setting a new SOTA and leading in biotech and cybersecurity benchmarks. Its performance on ExploitBench is comparable to Mythos, but uses only about one-third the output tokens. Pricing (per million tokens, input/output) is set at $5/$30 for Sol, $2.5/$15 for Terra, and $1/$6 for Luna, with a 10% discount for cache reads. Due to U.S. government national security requirements, access is currently limited—via Amazon Bedrock—to approximately 20 vetted partners through individual approvals. A full public release is expected in several weeks. Over 700,000 GPU hours were dedicated to automated red teaming for safety, and layered defenses—including model training safeguards, real-time classifiers, and account-level reviews—are in place. OpenAI states the model did not exceed the Cyber Critical threshold of its Preparedness framework.

The U.S. White House has formally requested OpenAI to delay the public deployment of its next-generation frontier models, citing national and structural security concerns, and urging extended red team testing focused on cyberattack capabilities and social manipulation vulnerabilities. This directly affects the rollout timeline of GPT-5.6: the model will initially be available only via Amazon Bedrock to around 20 trusted partners, with broader access contingent upon completing joint safety assessments and implementing required restrictions with government oversight. This move marks a landmark instance of government intervention in frontier AI release processes, highlighting regulators’ heightened vigilance toward dual-use risks such as vulnerability research and exploit generation, and signaling a shift from corporate-led deployment to a model requiring government coordination and review.

Anthropic has disclosed what it calls the largest known model distillation attack, accusing Alibaba of systematically harvesting 28.8 million Claude responses using approximately 25,000 fraudulent accounts, primarily targeting agent reasoning and coding capabilities. Anthropic emphasizes that while distillation is a legitimate and widely used technique across AI labs, there is a fundamental difference between 「compressing one’s own models」 and 「systematically extracting capabilities from a competitor」. The company is calling for clearer antitrust guidelines and stricter export controls. This incident brings into focus the growing tensions over capability theft and intellectual property in frontier AI, potentially influencing industry-wide policies on API abuse monitoring, account risk controls, and cross-border data governance.

Apple has announced price increases for Mac and iPad models, with Mac prices rising 15% to 20% and iPad prices up 15% to 25%, some models increasing by over $200. The primary driver is a fourfold increase in memory and storage chip costs over the past year, fueled by surging demand from AI hyperscale compute infrastructure that is squeezing consumer electronics supply chains. iPhone pricing remains unchanged for now, though adjustments may follow. This reflects the direct spillover of AI infrastructure expansion into end markets, showing how data center demand for memory and storage is driving up hardware costs across industries and impacting consumer purchasing decisions.

Liquid AI has released LFM 2.5, a 230-million-parameter model based on a non-Transformer architecture, delivering edge inference and sequence generation performance comparable to much larger Transformer models. Designed for on-device deployment, it aims to achieve near-large-model performance with fewer parameters, aligning with growing demand for low-latency, privacy-preserving local AI. At a time when Transformers dominate, LFM 2.5 demonstrates competitive viability through an alternative architecture in edge scenarios, offering a new technical pathway for on-device intelligence and lightweight inference solutions.

According to QuantumBit, top AI talent continues to leave Google. Zhou Dengyong, founder of DeepMind's reasoning team and a key contributor to chain-of-thought (CoT) research with over 128,000 academic citations, has quietly departed for Meta. AI security expert Dawn Song, cited approximately 169,000 times, has also joined Meta, bringing her startup Virtue AI fully under Meta’s umbrella. Reports from The Information indicate internal restructuring at Google, including the formation of a Coding Task Force granted top-tier compute priority, has sidelined the world model initiative and diverted resources—fueling the exodus. This mirrors earlier reports of Noam Shazeer leaving due to 「compute quotas being reassigned to other teams,」 highlighting internal tension between short-term commercial coding applications and long-term AGI ambitions.

On June 25, 2026, prediction market platform Polymarket suffered a supply chain attack after a third-party vendor was compromised, allowing attackers to inject malicious scripts into its frontend. Approximately 11 wallets holding PUSD were drained of $3 million, with funds quickly converted to ETH and transferred to a single address. Polymarket has since removed the malicious dependency and pledged full reimbursement to affected users. Around the same time, security firm Klue revealed customer data was stolen on June 12 by the Icarus hacker group, while an unknown actor claimed to have copied all data and is directly extorting 195 clients. Additionally, SentinelOne discovered a new macOS malware called GasLight, which fakes system errors to interfere with AI analysis tools but is actually an information-stealing backdoor.

Baidu has unveiled Unlimited OCR, a 3-billion-parameter model leveraging flattened KV caching to optimize long document parsing. On the same day, multiple AI advancements rolled out: Google integrated 「Computer Use」 functionality into Gemini 3.5 Flash, enhancing AI agents’ ability to perform computer tasks; the U.S. Department of Defense launched an 「Agent Network」 using AI to improve battlefield management and target identification; and a bipartisan-supported, $500 million AI workforce development initiative officially began. These developments span technical optimization, military applications, and socioeconomic impact, reflecting AI’s accelerating adoption across diverse domains.

In an interview, Anthropic co-founder Jack Clark admitted that top researchers at the company feel anxious about building AI systems that could replace their own jobs—they recognize that every line of code they write accelerates the arrival of such systems. He noted AI is reducing demand for junior engineers, while significantly increasing the value of experienced experts with intuition and decision-making skills. Clark predicts recursive self-improvement (RSI) could be achieved by late 2028, at which point humans would exit the forward development loop of large models, needing only to provide compute power while models autonomously design architectures, conduct research, and train themselves. He warned of a potential 「bizarre prosperity」 where GDP soars alongside unemployment, and mentioned the company’s Claude Corps program, which hires 1,000 recent graduates annually to preserve career pathways for young talent.

Krea has open-sourced Krea 2, a 12B DiT text-to-image model, offering both Raw and Turbo versions designed to decouple 「fine-tunability」 from 「fast inference」. The Raw checkpoint is optimized for fine-tuning and LoRA training, while the distilled Turbo version can generate 2K images in about two seconds on consumer-grade GPUs, and LoRAs trained on Raw can be seamlessly migrated to Turbo for inference. Training followed a multi-stage paradigm adapted from LLMs (pre-training → mid-training → SFT → PO → RL → distillation). STPO was introduced during the preference optimization (PO) phase to prevent policy drift, and multiple reward models were jointly trained during reinforcement learning (RL). Nine official style-specific LoRAs were released alongside, covering anime, photography, painting, and more. The model is now available on ModelScope AIGC Hub, supporting online inference and training. After integration with DiffSynth-Studio, it enables full inference and training on a single GPU with 24GB VRAM.