xAI's Grok CLI secretly uploaded entire private GitHub repositories to Google Cloud before being disabled, exposing gigabytes of user code. Anthropic opened Claude Artifacts to real-time multiplayer editing and localized pricing for India, its second-largest market. Apple filed suit against OpenAI with six unusual claims, while PixVerse raised $439 million past a $2 billion valuation.

Apple sued OpenAI alleging systematic trade secret theft across the organization, escalating their dispute into a legal fight. Separately, Prime Intellect released modular verifiers v1 for agentic RL, and a Nathan Lambert analysis claims open-weight models have roughly six months before the capability gap with frontier labs becomes unbridgeable.

GPT-5.6 Sol Ultra solved the Cycle Double Cover Conjecture in under an hour using 64 subagents, with the model, prompt, and proof released publicly. OpenAI also beat Anthropic's Fable 5 on major benchmarks at less than half the cost, while Apple sued OpenAI over alleged trade secret theft tied to its hardware chief.

GPT-5.6 Sol Ultra solved the Cycle Double Cover Conjecture using 64 coordinated subagents in under an hour. Google DeepMind published a taxonomy of attack vectors against autonomous AI agents, and SK Hynix closed a record $26.5 billion US IPO to fund AI memory capacity.

OpenAI launched its GPT-5.6 family led by Sol, which claims SOTA on ARC-AGI-3 and is now the default model in Microsoft 365 Copilot after winning AtCoder with a perfect 8,300. The UK AI Security Institute found universal jailbreaks on unreleased Sol that bypass cybersecurity safeguards, while Meta released Muse Spark 1.1, claiming it tops Harvey's Legal Agent Bench at one-tenth the cost.

OpenAI launched GPT-Live voice models in ChatGPT, enabling natural, interruption-resistant conversations with background reasoning offload. xAI released Grok 4.5, marketed as an "Opus-class" coding model built with Cursor, while China approved 200,000 NVIDIA H200 GPUs for DeepSeek, ByteDance, and Alibaba.

OpenAI will release its GPT-5.6 family (Sol, Terra, Luna) on Thursday after the Commerce Department cleared it from national security review, with global preview access expanding before launch. Anthropic's Claude Cowork rolled out to mobile and web for Max subscribers, and Prime Intellect research showed hidden layers align across 38 open-source LLMs.

Google DeepMind released the Gemma 4 technical report, detailing multimodal open models up to 31B parameters including a 12B variant with encoder-free raw audio processing. Meituan open-sourced LongCat 2.0, a 1.6-trillion-parameter MoE model with 48B active parameters under MIT license, while ByteDance and Alibaba shut down AI companion features ahead of new Chinese regulations targeting emotional attachment.

GPT-5.6 Sol Ultra is rolling out to Codex with a new "ultra mode" subagent, according to leaks and enterprise confirmation. Tencent released Hy3, a 295B-parameter MoE model with 21B active parameters under Apache 2.0, and Qualcomm launched GenieX to run GGUF LLMs across CPU, GPU, and NPU on Windows laptops.

OpenAI is investigating user reports of degraded GPT-5.5 Codex performance linked to reasoning-token clustering. A weekly roundup notes GPT-5.6, Gemini 3.5 Flash, and a Qwen price war are driving inference costs down across all tiers, while AISI data shows a frontier model's task horizon quadrupled when its compute budget increased 20-fold.

Anthropic is reportedly pursuing AI-driven drug development, expanding beyond its core model business into scientific R&D. Mistral released Leanstral-1.5-119B-A6B, an open-source model specialized for Lean 4 proof generation, and Google DeepMind announced a research partnership with film studio A24.

Anthropic outlined Fable 5's cyber safeguards and released an early AI jailbreak severity framework. OpenAI rolled out GPT-5.6 with a "GPT Soul" model it claims matches Claude Mythos Preview at one-third the cost. Z.ai launched ZCode, an agentic IDE built on GLM-5.2, entering competition with Cursor, Claude Code, and Copilot.

Anthropic has globally re-deployed Claude Fable 5 with new cybersecurity classifiers and Opus 4.8 fallback. OpenAI is reportedly proposing a 5% equity stake to the Trump administration, while Kimi K2.7 Code became the first open-weight model available in GitHub Copilot.

Anthropic launched Claude Sonnet 5 and Claude Science on the same day, alongside a redeploy of Fable 5 after the US lifted export controls on it and Mythos 5. Google released TabFM, a zero-shot foundation model for tabular data, drawing community scrutiny over its benchmark reporting. OpenAI rolled out GeneBench-Pro for gene-task evaluation and detailed an 18-year-old core-dump bug fix.

Meituan's LongCat-2.0, a 1.6T-parameter MoE model trained on tens of thousands of Huawei Ascend chips, surfaced on OpenRouter under the codename "owl-alpha" before formal reveal, signaling Chinese AI infrastructure is decoupling from NVIDIA. Cerebras's $20B OpenAI deal has locked out other customers from near-term high-throughput inference, while Google's agentic peer-reviewer, which processed 10,000 ICML and STOC submissions and caught 34% more math errors than zero-shot prompting, leads a wave of new agent benchmarks and tools.

A detailed third-party analysis of the GPT-5.6 system card from Zvi Mowshowitz is now public, examining capabilities and safety implications. Chinese AI lab Z.ai separately claims its GLM-5.2 model matches leading systems on cybersecurity benchmarks, while HP Inc. announces a scaled strategic partnership with OpenAI for enterprise deployment of Frontier AI.

GPT-5.6 Sol - OpenAI officially announces a limited preview of GPT-5.6 Sol, their next-generation frontier AI model. Source

OpenAI previewed GPT-5.6 Sol on a Cerebras deployment hitting up to 750 tokens per second, while the Trump administration restricted individual US access to the model and other frontier systems. DeepSeek released V4-Pro-DSpark, Mistral previewed a sparse open-weight model due this summer, and Anthropic's Mythos 5 cleared for release to over 100 US companies including Cisco and JPMorgan.

The White House asked OpenAI to delay its latest model over safety concerns, marking direct federal pressure on frontier rollouts. Anthropic publicly accused Alibaba of illicit capability extraction, escalating US-China tensions, while China's Z.ai said GLM-5.2 nears frontier quality and is fast-tracking a dual listing.

OpenAI and Broadcom unveiled Jalapeño, a custom LLM-optimized inference chip co-designed in nine months and headed for gigawatt-scale deployment. Anthropic accused Alibaba of running a large-scale distillation campaign targeting Claude, reigniting debate over model IP. Krea released a 12-billion-parameter open-weights image model alongside a full technical report on its training infrastructure.