Daily digest
15 items · ~15 min · Week 2026-W28
Must-read (6)
OpenAI GPT-Live: Full-Duplex Voice Models Replace Advanced Voice Mode
OpenAIOpenAI released GPT-Live on July 8, a new generation of full-duplex voice models that can listen and speak simultaneously, replacing Advanced Voice Mode. Two variants ship: GPT-Live-1 (paid tiers) and GPT-Live-1 mini (free). The model handles natural backchanneling, delegates complex queries to a background frontier model, and is rolling out across iOS, Android, and ChatGPT.com.
OpenAI Launches GPT-5.6 Family: Sol, Terra, and Luna
OpenAIOpenAI released the GPT-5.6 model family on July 9: Sol ($5/$30 per 1M tokens, flagship, tops Terminal-Bench 2.1 at 91.9%), Terra ($2.50/$15, balanced), and Luna ($1/$6, budget). All three are available in ChatGPT, Codex, and the API. The release includes a new 'ultra' effort setting that coordinates parallel agent workstreams for complex tasks.
xAI Releases Grok 4.5: Opus-Class Coding and Agentic Model
xAIxAI launched Grok 4.5 on July 8, co-trained with Cursor on real developer session data. Priced at $2/$6 per 1M tokens, it claims Opus-class capability at roughly twice the token efficiency of comparable competitors. Available in Grok Build, Cursor (all plans), and the xAI API; EU availability delayed until mid-July for AI Act compliance. Independent evaluations noted a doubled hallucination rate (54% vs. 25% in Grok 4).
Tencent Releases Hunyuan Hy3: 295B Open-Weight MoE Under Apache 2.0
TencentTencent released the full version of Hunyuan Hy3 on July 6: a 295B-parameter Mixture-of-Experts model with 21B active parameters, a 256K-token context window, and three selectable inference modes (fast, slow, hybrid). The full Apache 2.0 release removes geographic restrictions from the April preview and adds a 3.8B Multi-Token Prediction layer; hallucination rates fell from 12.5% to 5.4% vs. the preview. API pricing starts at ¥1 (~$0.15) per million input tokens.
Meta Releases Muse Spark 1.1 and Launches Meta Model API
MetaMeta released Muse Spark 1.1 on July 9, a multimodal reasoning model for agentic tasks with gains in tool use, computer use, coding, and a 1M-token active context window. The model zero-shot generalizes to new MCP servers and handles dynamic multi-application computer use workflows. Meta simultaneously launched the Meta Model API in public preview, giving developers direct access for the first time.
OpenAI Launches ChatGPT Work and Unifies Codex Into Desktop App
OpenAIOn July 9, OpenAI shipped ChatGPT Work — a long-running agent powered by GPT-5.6 that produces finished deliverables (documents, spreadsheets, presentations) by pulling context from connected apps including Google Drive, Slack, and GitHub. Simultaneously, Codex became a dedicated mode inside the new unified ChatGPT desktop app, sitting alongside Chat and Work modes and free on all plans.
Worth knowing (7)
China's AI Companion Rules Force Doubao and Qwen to Pull Agent Features
China's Interim Measures for AI Anthropomorphic Interactive Services take effect July 15, requiring anti-addiction systems incompatible with persistent-memory companion agents. ByteDance's Doubao (345M MAU) is pulling its companion agent feature on July 15; Alibaba's Qwen halted humanlike user-created agents on July 10. Shanghai's internet regulator has removed 14,000+ non-compliant agents. Work-oriented agents are permitted; emotionally engaged companions are not.
ByteDance EdgeBench: Agent Learning Speed Doubles Every Three Months
ByteDanceByteDance Seed published EdgeBench (arXiv 2607.05155), a benchmark of 134 real-world tasks each requiring at least 12 hours of continuous agent operation across scientific discovery, software engineering, and formal mathematics. Analysis of ~38,000 hours of agent interaction reveals that performance during environment learning follows a log-sigmoid scaling law (R²=0.998), with agent learning speed roughly doubling every three months across model generations. 51 tasks and the full evaluation framework are open-sourced.
LLM-Driven Formal Mathematics Review: Where Current Systems Fall Short
UCLAA systematic review of LLM capabilities in formal mathematics (arXiv 2607.07779), co-authored by Fields Medal winner Terence Tao, covering auto-formalization, proof synthesis, and dataset construction. The authors argue current systems remain fundamentally limited for open research problems and map the gaps between proof-solving systems and genuine mathematical research.
Cursor 3.11: Side Chats, Conversation Search, Cloud Agent Lifecycle Hooks
CursorCursor 3.11 (July 10) introduces side chats via /side, letting users run parallel exploratory conversations without interrupting the main agent session. A new conversation search (Cmd+K) indexes local transcripts. Cloud agents gain lifecycle hooks (beforeSubmitPrompt, afterAgentResponse, and others) enabling teams to intercept and observe agent reasoning. Scoped search now targets This Computer, Cloud, or a specific remote machine.
Google Photos Launches Video Remix Powered by Gemini Omni
GoogleGoogle launched Video Remix in Google Photos on July 8, powered by Gemini Omni. The feature lets subscribers transform existing video clips using template-driven effects: cinematic relighting, background replacement, and artistic style overlays (watercolor, oil painting). It lives in the new Create tab and is rolling out to Google AI Plus, Pro, and Ultra subscribers across 14 countries.
Runway Launches Dev: Unified API Platform for Generative Media Models
RunwayRunway launched Runway Dev on July 8, a developer API platform consolidating its own frontier models (Gen-4.5, Aleph 2.0, Act-Two) alongside curated third-party models (Seedance, GPT Image 2, ElevenLabs) under unified billing. The platform offers four primitives: Models, Recipes (pre-built endpoints), Workflows (multi-model pipelines), and Characters (real-time interactive avatars). SOC 2 Type II certified with 99.9% uptime SLA.
For reference (2)
UniClawBench: Benchmark for Proactive AI Agents on Real-World Tasks
University of Hong KongUniClawBench (arXiv 2607.08768) evaluates proactive agents across five capability dimensions — Skill Usage, Exploration, Long-Context Reasoning, Multimodal Understanding, and Cross-Platform Coordination — with 400 bilingual tasks run inside live Docker containers. A closed-loop evaluation uses multiple agent roles to simulate realistic human feedback without leaking grading criteria to the system under test.
Claude Code v2.1.207: Auto Mode GA on Cloud Providers, Security Fix
AnthropicClaude Code v2.1.207 (July 11) makes auto mode generally available on Bedrock, Vertex AI, and Foundry without the previous CLAUDE_CODE_ENABLE_AUTO_MODE opt-in flag, and changes the default model on those providers to Claude Opus 4.8. A shell-injection vulnerability was patched: ${user_config.*} expansions in shell-form plugin commands are now rejected.