OpenAI Unveils GPT-5.6 Family: Luna, Terra, Sol with Tiered PricingChatGPT Work: New Agent That Acts Across Apps and Files, Works for HoursSpaceXAI Launches Grok 4.5, First Opus-Class Model After Cursor AcquisitionOpenAI Unveils GPT-5.6: Smarter per Token, Better Performance per DollarMeta Launches Muse Spark 1.1, First Spark Model with API, Boosts Tool CallingGPT-5.6 Becomes Default Model in Microsoft 365 CopilotBuilding a Real-Time AI Tutor for 5-Year-Olds in Under 1000msTriRoute: Unified Learned Routing for Adaptive Attention, Experts, KV-CacheAgentLens: New Benchmark Evaluates Coding Agents with Production Trajectory DataCost-Effective Agent Achieves Breakthrough in ARC-AGI-1 Abstract ReasoningLLM Agents Evolve: From Atomic Actions to Standard Operating ProceduresMILES: Modular Instruction Memory Helps LLMs Self-Improve During ReasoningLLMs Silently Correct African American English; Activation Steering Reduces BiasMozilla AI: Next AI Wave Is About Infrastructure, Not Just ModelsAI-Generated Content Floods Social Media, Especially LinkedInDeepSeek Plans to Build Its Own AI Chip, Aiming for Compute IndependenceOpenAI Launches GPT-5.5 Bio Bounty Program for Security TestingGitHub Gave Every Active Repository a Validated Owner in 45 DaysDify v1.16.0-rc1 Introduces Experimental Agent Experiencellm-meta-ai 0.1 Released: Adds Support for Muse Spark 1.1 to LLM ToolOpenAI Unveils GPT-5.6 Family: Luna, Terra, Sol with Tiered Pricing

OpenAI launched its GPT-5.6 family today, featuring three sizes: Luna, Terra, and Sol. Pricing per 1M input/output tokens is Luna $1/$6, Terra $2.50/$15, Sol $5/$30. For comparison, Claude Opus costs $5/$25. The new models are now available as the default option in select applications.

SpaceXAI has released Grok 4.5, its first Opus-class model following the acquisition of Cursor. The model brings significant improvements in reasoning, code generation, and long-context processing, positioning it as one of the fastest-iterating frontier models in the industry.

OpenAI released GPT-5.6, calling it ‘frontier intelligence that scales with your ambition.’ It offers more intelligence per token, stronger performance per dollar, and more capability on demand for the hardest tasks.

Meta released Muse Spark 1.1, the first Spark model to offer an API. Compared to the initial April release, it features significant improvements in agentic tool calling and computer use capabilities. Developers can now access the model via API; detailed performance metrics are available in the official evaluation report.

Microsoft has adopted GPT-5.6 as the default model for Microsoft 365 Copilot. The upgrade brings stronger AI capabilities across Word, Excel, PowerPoint, Chat, and Cowork, enabling faster and higher-quality work outputs for enterprise users. This marks a deeper integration of OpenAI’s flagship model in productivity software.

Elo shares insights on building a real-time AI tutor for 5-year-olds, with a core requirement of under 1000ms response time to maintain attention. The post details how personalized teaching is achieved under low-latency constraints, offering valuable lessons for children’s educational AI products.

A new paper introduces TriRoute, a unified learned routing approach that jointly optimizes adaptive attention, Mixture-of-Experts, and KV-cache allocation. Unlike methods that act on a single axis (MoE for FFN sparsification or MoD for block skipping), TriRoute coordinates multiple conditional computation dimensions within one framework, potentially reducing per-token inference costs.

Researchers introduced AgentLens, a production-assessed benchmark for interactive code agents. Unlike traditional single-pass/fail metrics, it evaluates the entire execution trajectory, capturing intermediate steps and errors for a more realistic user experience assessment.

A new study presents a cost-effective agent framework that advances on the ARC-AGI-1 abstract reasoning benchmark. Unlike prior approaches requiring heavy test-time compute or benchmark-specific training, this work finds a more economical middle ground.

A new arXiv paper proposes a framework for LLM agents to self-evolve by iteratively optimizing their tools. Agents move from executing granular atomic actions to following Standard Operating Procedures (SOPs), improving performance on complex tasks. The approach reduces reliance on human-designed workflows and enables more autonomous problem-solving.

A new arXiv paper introduces MILES, a framework with modular instruction memory and learnable selection that enables LLMs to accumulate and reuse experience across sequential problems. Unlike isolated problem-solving, MILES allows models to autonomously select relevant past experiences for new tasks, achieving self-improvement during reasoning. Experiments show strong gains on multiple reasoning benchmarks.

A new arXiv study reveals that major LLMs (14B-70B) systematically ‘correct’ African American English (AAE) to standard English, leading to misinterpretation and dialect bias. The researchers propose an activation steering intervention that reduces this bias without harming overall model performance. The work highlights AI fairness issues and offers a practical technique for developers serving diverse user populations.

Mozilla AI published a blog post arguing that the next era of AI will be defined by infrastructure, not just models. It emphasizes that the ‘control layer’—including data pipelines, monitoring, and safety guardrails—is crucial for real-world AI deployment. Developers and enterprises are advised to prioritize robust infrastructure alongside model performance.

An analysis reports that AI-generated content is flooding social media, with LinkedIn being the worst offender. The article discusses its impact on user trust and platform ecosystems, scoring 175 points and 154 comments on Hacker News.

AI startup DeepSeek is reportedly planning to develop its own AI chips, according to Proactive Investors. The move aims to reduce reliance on external suppliers like NVIDIA and achieve greater compute independence. Success could allow DeepSeek to optimize hardware for its models and potentially disrupt the current chip market.

OpenAI announced the GPT-5.5 Bio Bounty program, encouraging researchers to find and report potential biosafety vulnerabilities in its models. Details on rewards and participation are available on the OpenAI website.

GitHub resolved ownership for over 14,000 repositories, with fewer than half having clear ownership initially. The team gave every active repository a validated owner in under 45 days and archived the rest. This process became the foundation for all subsequent repository management improvements.

Dify released v1.16.0-rc1, introducing an experimental Agent experience. The new feature leverages a shell-based LLM agent paradigm, bringing a significant leap in agent capabilities. Dify warns the service should only be provided to trusted, non-malicious users.

Simon Willison released llm-meta-ai 0.1, a plugin that lets users run prompts against Meta’s new Muse Spark 1.1 model via his LLM command-line tool.