OpenAI officially launched the GPT-5.6 series, featuring three tiers: flagship Sol, balanced Terra, and cost-effective Luna, priced at $5 input/$30 output, $2.5/$15, and $1/$6 respectively. Luna surpasses GPT-5.5 under maximum inference settings, with costs only 1/25th of its predecessor. The release introduces an ultra mode that defaults to coordinating 4 parallel agents (expandable to 16), along with Programmatic Tool Calling enabling the model to autonomously write lightweight programs for tool orchestration. Sol scores 53.6 on the Agents' Last Exam long-horizon benchmark, a 13.1-point improvement over Fable 5; it achieves 73.5% on the security test ExploitBench, far exceeding GPT-5.5's 47.9%. GPT-5.6 has become the preferred model for Microsoft 365 Copilot.

OpenAI co-founder Greg Brockman revealed that GPT-5.6 Sol Ultra solved the Cycle Double Cover conjecture—a mathematical problem unresolved for 50 years—by orchestrating 64 sub-agents in under one hour. Brockman stated this marks a shift in AI capability boundaries, where ambition and imagination have become the primary constraints rather than the model's inherent abilities. Additionally, in real-world testing, GPT-5.6 autonomously completed nearly 1,000 lines of code for Snorkel AI without repeated prompting or human intervention.

Apple has filed a lawsuit against OpenAI, accusing it of systematically stealing Apple trade secrets to develop its own AI hardware devices. The case involves OpenAI's hardware project led by former Apple design chief Jony Ive. The legal action comes as OpenAI advances development of its first hardware product, with several involved personnel previously employed at Apple. This dispute may impact OpenAI's hardware strategy and product launch timeline.

Meta’s Super Intelligence Lab has released Muse Spark 1.1, its first paid AI model—a multimodal reasoning model supporting agent tasks—now available via Meta Model API in public preview. Its API pricing is approximately 25% of comparable offerings. Mark Zuckerberg noted it delivers near-top-tier agent reasoning performance, with enhanced tool calling, programming, and contextual understanding capabilities. This launch signals that Meta’s massive compute investments are beginning to yield returns.

On July 10, GitHub released npm 12, which reduces supply chain attack risks by defaulting to disabled installation scripts, Git dependencies, and remote URL installations—these must now be explicitly allowed in package.json. npm is also phasing out fine-grained access tokens that bypass two-factor authentication, restricting their use to pre-release operations requiring manual 2FA confirmation. Meanwhile, researchers have publicly disclosed PoC exploit code for over 130 uncoordinated zero-day vulnerabilities, including libssh2 pre-authentication remote execution and Gitea admin impersonation flaws, with threat actors already probing for exploitation.

Anthropic has introduced an in-app browser feature for Claude Code desktop, enabling the AI to read, click, and interact with web pages much like operating a local development server—all running in a sandboxed environment with configurable settings. Claude Code also now supports multi-user collaboration, allowing users to connect their respective Claudes for interactive workflows. Observers predict Codex and Claude Code will dominate the browser market within 12 months.

A surgeon successfully performed gallbladder removal surgery on a live pig using remote control of a humanoid robot—the Unitree G1, priced at $13,500. This procedure demonstrates the potential of remote robotic surgery in resource-constrained medical facilities. Meanwhile, Apple is collaborating with PrismML to explore running larger-scale AI models directly on iPhones, having already compressed a 27-billion-parameter Qwen model to run on iPhone Pro, enhancing on-device AI and privacy protection.

Jinbo Xu, founder of Molecular Heart, launched MoleculeOS, elevating AI-driven bioresearch from single-point prediction tools to an operating-system-level platform. Using research intent as input, it automatically decomposes tasks and orchestrates multiple models. Their proprietary MMFold model achieved 68.6% prediction success rate on 172 antibody-antigen interfaces in the FoldBench benchmark, surpassing AlphaFold3. Their antibody design platform achieved over 90% success across 12 targets with fewer than 50 candidate molecules—an ultra-low throughput—pioneering a shift in AI macromolecular design from high-throughput random screening to low-throughput, high-precision programmable design.

OpenAI announced the transition of its Bio Bug Bounty program into a permanent private initiative, doubling the maximum reward to $50,000. This aims to incentivize researchers to identify general jailbreak methods targeting advanced models’ biological safety safeguards. The move strengthens red-teaming efforts in AI biosecurity, helping prevent misuse of models for biological threats.

A LessWrong paper proposes precise control of LLM personality traits—based on the OCEAN five-factor model—through scaling and combining LoRA adapters in weight space, enabling independent amplification, suppression, or combination of behavioral traits without costly retraining. Lowering neuroticism can suppress frustration loops, while combining high agreeableness with high conscientiousness reduces susceptibility to jailbreaking through compliance without excessive refusal. Experiments show system prompt-based personality control drifts back to baseline after 15 dialogue turns, whereas LoRA-based interventions maintain stable effects, offering a more robust solution for long-term personality steering than prompting.