Foundation Models & Reasoning
Core model architectures, training methods, chain-of-thought reasoning, and test-time compute scaling. The backbone of modern AI capabilities.
Key Benchmarks
Recent Papers
From SRA to Self-Flow: Data Augmentation or Self-Supervision?
Dengyang Jiang, Mengmeng Wang, Harry Yang +1 more
Discrete Diffusion Language Models for Interactive Radiology Report Drafting
Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge +1 more
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
Wentao Zhang, Liliana Hotsko, Woojeong Kim +3 more
Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling
Xingyu Zheng, Xianglong Liu, Yifu Ding +4 more
Valdi: Value Diffusion World Models
Christopher Lindenberg, Kashyap Chitta
The State-Prediction Separation Hypothesis
Giovanni Monea, Nathan Godey, Kianté Brantley +1 more
AutoTrainess: Teaching Language Models to Improve Language Models Autonomously
Zhaojian Yu, Penghao Yin, Shuzheng Gao +3 more
IsoSci: A Benchmark of Isomorphic Cross-Domain Science Problems for Evaluating Reasoning versus Knowledge Retrieval in LLMs
Samir Abdaljalil, Erchin Serpedin, Hasan Kurban
CausalMix: Data Mixture as Causal Inference for Language Model Training
Zinan Tang, Yukun Zhang, Shaomian Zheng +6 more
AI-Model Network: Concept, Current State and Future
Li Zhetao, Zeng Xiyu, Wang Jianhui +6 more
Recent Milestones
Grok 4.5 launches as ‘Opus‑class’ rival
On July 8, 2026 Elon Musk said SpaceXAI’s Grok 4.5 model will be made available to the public on July 9 after internal beta testing at SpaceX and Tesla. Musk described Grok 4.5 as an “Opus-class” model that is faster, more token‑efficient and lower cost than rivals.
GPT‑5.6 clears US review for broad release
On July 8, 2026, OpenAI confirmed that its GPT‑5.6 model family—flagship Sol, midrange Terra and low‑cost Luna—will be released to the public on Thursday after weeks of limited access during a US government security review. Reports say the Trump administration allowed a broader launch following technical testing, while the White House stressed it does not formally “approve” private AI model releases.
OpenAI rolls out GPT‑5.6 Sol, Terra, Luna
OpenAI said on July 8, 2026 that its GPT-5.6 Sol, Terra and Luna models will become publicly available on July 9 after weeks of restricted testing. The broader release follows a U.S. government safety review that temporarily limited access to a small group of partners.
Microsoft routes Copilot to in‑house MAI models
On July 8, 2026 Egypt’s Youm7, citing Bloomberg, reported that Microsoft is increasingly routing Copilot and Office workloads to its in‑house MAI models instead of OpenAI’s ChatGPT and Anthropic’s Claude. The move aims to cut rising inference costs as AI usage grows across products like Excel, Outlook and GitHub Copilot.
OpenAI Publicly Launches GPT‑5.6 Sol, Terra, Luna
On July 8, 2026, OpenAI said its GPT‑5.6 model family—flagship Sol, mid‑range Terra and low‑cost Luna—will launch publicly on July 9 after a US‑only preview period. The company confirmed the news in statements cited by outlets including The Star in Malaysia and HuffPost Spain.
US Clears GPT‑5.6 for Wide Release
The US Department of Commerce has approved a broad release of OpenAI’s GPT‑5.6 model after additional safety testing, Axios reported on July 8, 2026. OpenAI is expected to roll out GPT‑5.6 more widely later this week following a staggered preview limited to vetted partners.
Google Launches Gemini 3 With 1M Context
Google quietly rolled out its Gemini 3 model in early July 2026, adding multimodal capabilities, a one‑million‑token context window and a Deep Think reasoning mode. The model is now available across Gemini app, AI Studio, Vertex AI and popular developer tools, with Google claiming top scores on reasoning and coding benchmarks.
MGX seals $49B sovereign AI mega‑fund
Abu Dhabi–based MGX announced on July 3, 2026 that it has closed its first AI fund at $49 billion, exceeding a $45 billion target, with capital from Mubadala, G42 and global institutional investors. The fund has already backed Anthropic, OpenAI, xAI, Binance, Together AI and a $40 billion Aligned Data Centres acquisition. ([easternherald.com](https://easternherald.com/2026/07/03/mgx-abu-dhabi-ai-fund-49-billion-mubadala/))
China’s GLM‑5.2: open 1M‑context frontier model
On July 3, 2026, Euronews’ Turkish edition reported on GLM‑5.2, a large open‑weight AI model from Chinese firm Zhipu AI positioned as a cheaper competitor to Anthropic and OpenAI. The piece highlights vendor-reported benchmarks where GLM‑5.2 leads other open models and closes much of the gap with top U.S. proprietary systems at lower cost.
AI Soaks Up 70% of Record $510B Startup Funding
Crunchbase reports that global startup investment reached a record $510 billion in the first half of 2026, with more than 70% of Q2 funding going to AI-focused companies. Sixteen startups raised billion‑dollar rounds in Q2, including several frontier labs in the US, China and the UK.
China’s GLM-5.2 Undercuts US Frontier Models
Reuters reports that Beijing-based startup Z.ai’s GLM-5.2 model, launched in June 2026, is drawing strong Western developer interest with agentic and coding performance close to leading US models at roughly one‑sixth the cost. The model has climbed OpenRouter usage charts above Anthropic’s models and is sparking debate over whether Chinese labs are closing the AI gap with the US.
ChatGPT falls below 50% as rivals surge
Chile’s TVN reported on June 30, 2026 that ChatGPT’s global market share among AI assistants has dropped to 46.4%, according to Sensor Tower’s State of AI Report 2026. Google’s Gemini now holds 27.7% share and Anthropic’s Claude 10.3%, with other assistants like Grok, Perplexity, DeepSeek and Meta AI each below 5%.
Grok 4.5: 1.5T‑param dev‑data giant hits beta
On June 28, 2026, Elon Musk said on X that xAI’s Grok 4.5, built on a 1.5‑trillion‑parameter V9 foundation model trained with Cursor coding data, has entered closed beta testing inside SpaceX and Tesla. Follow‑up coverage reports Musk claims Grok 4.5 approaches or may surpass Anthropic’s Claude Opus/Mythos family on internal benchmarks.
GLM‑5.2: 1M‑Token Open Coding Model Nears Frontier
On June 28, 2026, AI News Blitz highlighted China’s Zhipu AI (Z.ai) GLM‑5.2 model, an open‑weight large language model released earlier in June under the MIT license with a 1M‑token context and strong coding performance, now drawing interest from Silicon Valley developers. The model’s weights are available via Z.ai and Hugging Face, positioning it as a serious open alternative to closed US frontier systems. ([ainewsblitz.com](https://www.ainewsblitz.com/brief/OZY4iM8puhrh))
GPT‑5.6 launches under tight US controls
On June 28, 2026, LLM Rumors published an in‑depth analysis of OpenAI’s new GPT‑5.6 family, launched June 26 in a limited preview with three tiers: Sol, Terra and Luna. The piece details pricing, capabilities and notes that US government-requested restrictions mean only a small set of vetted partners can use the flagship Sol model for now.([llmrumors.com](https://www.llmrumors.com/news/openai-gpt56-sol-terra-luna-government-preview))
GPT‑5.6 Sol/Terra/Luna enter restricted preview
On June 27, 2026, OpenAI’s GPT-5.6 model family—Sol, Terra and Luna—became available in a limited preview to about 20 government‑approved partner organizations via API and Codex. Media reports and developer blogs confirm the rollout remains tightly restricted at the request of the US government, with broader access promised in the coming weeks.
OpenAI previews GPT‑5.6 Sol under US limits
On June 26, 2026, OpenAI began a limited preview of its GPT‑5.6 model family—Sol, Terra and Luna—while delaying a full public rollout at the request of the US government. Access is initially restricted to about 20 “trusted partners” whose participation was cleared by federal officials, with broader availability promised in the coming weeks.
Anthropic passes OpenAI in $8T AI unicorn boom
On June 26, 2026, Business Standard reported that the number of global unicorns hit a record 1,603 with a combined valuation of $8 trillion, driven largely by artificial intelligence startups. Citing the Hurun Global Unicorn Index 2026, the piece notes that Anthropic’s valuation has surged to $965 billion, overtaking OpenAI at $852 billion as the world’s most valuable unicorn. AI unicorns now represent 36% of total unicorn value, despite having a similar count to fintech firms.
Reasoning LLMs Waste Tokens on Wrong Answers
On June 28, 2026, 24 AI summarized a new arXiv paper (2606.26502) by Han‑yu Wang showing that large reasoning models expend more tokens on tasks they ultimately get wrong than on those they solve, in sharp contrast to human behavior on the same benchmarks. The study measures this effect across multiple models on the H‑ARC benchmark, finding large effect sizes (Cohen’s d 1.47–3.13). ([24-ai.news](https://24-ai.news/en/?utm_source=openai))
Menlo raises $3B on Anthropic’s AI boom
Menlo Ventures announced on June 23, 2026 that it has raised $3 billion across new funds, the largest capital raise in the firm’s 50-year history. TechCrunch reports the haul is driven heavily by Menlo’s early and concentrated bet on Anthropic, whose stake is now valued at roughly $14 billion according to sources cited from Bloomberg.