Model Race and Benchmarks
Grok 4.5's first day in the wild brought independent benchmarks, a viral debate over political bias, and a token-efficiency claim that could reshape the cost math for high-volume agentic workloads [1]. On Artificial Analysis's Intelligence Index, Grok 4.5 ranked fourth with a score of 54, trailing Claude Fable 5 (first), GPT-5.5 (second), and Claude Opus 4.8 (third) [1]. Meanwhile, the long-awaited Gemini 3.5 Pro still has no confirmed launch date and remains in a limited Vertex AI enterprise preview, now six weeks past its June 30 general-availability target and five weeks past the timeline promised at May's I/O conference [1].
Product and Feature Rivalry
OpenAI merged Codex and ChatGPT into a single super-app while launching the enterprise-focused ChatGPT Work [1]. Anthropic answered with a mobile version of Claude Cowork, sharpening the rivalry over who can best embed AI into everyday work [1].
China's Model Ascent
Z.ai's GLM-5.2 launched under an MIT open-source license with a 1-million-token context window, aimed at autonomously completing long-horizon engineering projects that once took teams weeks [2]. On Artificial Analysis's overall leaderboard, GLM-5.2 scored 51, the highest of any open-weight model, and Z.ai says it ranks first on Code Arena, a blind-test front-end coding arena involving millions of users [2]. On long-horizon coding benchmarks, GLM-5.2 trails Claude Opus 4.8 by just 1 percentage point on FrontierSWE and 4 points on Terminal-Bench 2.1, while surpassing GPT-5.5 and Opus 4.7 on several other tests [2]. The model supports Chinese domestic chips — Huawei Ascend, T-Head, Moore Threads, and Cambricon — from day one, and is priced at roughly a sixth of comparable closed US frontier models [2], fueling continued debate over whether China is finally closing the AI gap with the United States [1].
Funding Momentum
Capital keeps flowing toward physical AI and agentic systems for regulated workflows: Together AI, an open-source model training and serving platform, closed an $800 million Series C led by Aramco Ventures [3]. Video-understanding startup TwelveLabs raised a $100 million Series B [3]. Even Realities, maker of camera-free smart glasses built on proprietary waveguide optics, raised a $150 million pre-Series B led by Meituan [3]. Norm AI closed a $120 million Series C led by Khosla Ventures [3]. Monogram raised a $40 million seed round led by DST Global [3]. Prime Intellect closed a $130 million Series A led by Radical Ventures [3]. And local-model platform Ollama raised a $65 million Series B led by Theory Ventures [3]. Markets reflected the momentum: the Nasdaq Composite closed up 1.3% at 26,206.89 on Thursday, led higher by AI stocks [1].
Technology and Research
Training methods for long-horizon agent tasks remain a major research focus. "MAP: A Map-then-Act Paradigm for Long-Horizon Interactive Agent Reasoning" proposes splitting agent behavior into a "mapping" phase that explores the environment and a separate "acting" phase, cutting down on redundant exploration during long interactive tasks [4]. "Skill Reuse as Compression in Agentic RL" treats reusable skills learned during reinforcement-learning training as a form of policy compression, letting agents draw on compact, previously learned skills instead of re-deriving multi-step behavior from scratch on every new task — improving sample efficiency [4]. A third paper, "KLong: Training LLM Agent for Extremely Long-horizon Tasks," examines how to keep agents behaviorally consistent and goal-tracked over extremely long task horizons [4]. All three point in the same direction: making LLM agents cheaper to run and more reliable across ever-longer task chains.
References
[1] AI News Today July 10 2026: 15 Biggest Stories — https://www.buildfastwithai.com/blogs/ai-news-today-july-10-2026
[2] Z.ai pushes GLM-5.2 into open-weight spotlight / GLM-5.2 Benchmark Deep Dive — https://creati.ai/ai-news/2026-07-01/z-ai-pushes-glm-5-2-into-open-weight-spotlight-as-chinese-model-climbs-rankings-and-coding-bench/ and https://kie.ai/blog/glm-5-2-benchmark-deep-dive
[3] The Week's 10 Biggest Funding Rounds: AI, Energy And Biotech Lead The Way — Crunchbase News — https://news.crunchbase.com/venture/biggest-funding-rounds-ai-energy-biotech-joulent/
[4] MAP: A Map-then-Act Paradigm for Long-Horizon Interactive Agent Reasoning (arXiv:2605.13037) — https://arxiv.org/pdf/2605.13037; Skill Reuse as Compression in Agentic RL (arXiv:2605.31509) — https://arxiv.org/pdf/2605.31509; KLong: Training LLM Agent for Extremely Long-horizon Tasks (arXiv:2602.17547) — https://arxiv.org/pdf/2602.17547