TechCapsules

At Sarmam AI’s Epoch event, the startup unveiled a suite of language, vision and voice models and announced its move toward a stack AI platform. Highlights included the Epoch Builder Edition, enabling developers to fine‑tune and deploy models in ten Indian languages using 7‑billion and 70‑billion‑parameter systems trained on two trillion tokens, as well as Sarvam Vision 2.0, Bulbul V4 speech tools, Content Studio for dubbing, a retrieval layer for enterprise knowledge and Sarmam Inference on Indian infrastructure. Sarmam [...]

LG AI Research has released K-EXAONE 2.0, Korea’s largest AI foundation model, on the open-source platform Hugging Face. The model features 750 billion parameters, tripling the scale of its predecessor. It achieved a score of 94.4 on the long-text comprehension index, significantly outperforming China’s leading Zhipu AI GLM-5.1, which scored 71.5. Licensed under Apache 2.0, the model permits free commercial use. LG plans to unveil an industry-specialized AI model next week. Concurrently, DGIST Technology Holdings launched its [...]

A recent study analyzing over 680 million citations across major AI models, including ChatGPT, Google AI Overviews, Perplexity, Gemini, and Claude, reveals that Reddit dominates search results, accounting for approximately 40% of all citations. This trend is particularly pronounced in B2B technology, software, and industrial sectors, where community posts frequently outrank official vendor websites. Reddit’s citation share surged by at least 73% between October 2025 and January 2026, with Perplexity alone sourcing nearly a quarter of its [...]

xAI launched Grok 4.5 on July 8, 2026, entering the competitive AI coding assistant market with a model claiming "Opus-class" performance at significantly reduced costs. The new frontier model is integrated directly into Cursor, the coding tool credited with aiding its training, signaling a strategic focus on developer workflows rather than general consumer chat. Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens, undercutting Anthropic’s Claude Opus 4.8 by 76%. [...]

Economist Callum Williams estimates the global large language model market has reached an annualized revenue run rate of approximately $120 billion, marking a significant transition from experimental research to a major commercial industry. Anthropic currently leads in LLM revenue, leveraging backing from Amazon and Google to position its Claude models as enterprise favorites for reasoning and coding. This dominance occurs within a fiercely competitive landscape where OpenAI, Google, and Meta pursue distinct strategies, ranging from API services [...]

Anthropic has officially launched Claude Opus 5, a new artificial intelligence model designed to enhance capabilities in coding, reasoning, and scientific research. The company states that Opus 5 delivers performance comparable to its flagship Claude Fable 5 but at half the cost, making it a more practical solution for daily use. Available via the Claude API at the same price as Opus 4.8, the model introduces new developer features including Fast mode and automatic API fallbacks. Anthropic [...]

Anthropic has launched Claude Opus 5, positioning it as a cost-effective alternative to the heavily restricted Fable 5 model. The new release delivers performance comparable to its competitor while reportedly costing half the price. Benchmarks indicate that while Fable 5 retains slight advantages in specific areas, Claude Opus 5 outperforms in business workflows, computer use, and agentic research. It also achieves state-of-the-art results in several coding and knowledge work evaluations. Anthropic emphasizes that Opus 5 requires less [...]

Google has officially launched three new AI models: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. These updates aim to enhance speed, intelligence, and cost-efficiency for scalable AI agents. The flagship 3.6 Flash model reduces output token usage by 17% compared to its predecessor, lowering costs to $1.50 per million input tokens. It also features improved safety protocols against misuse in sensitive fields like biology and cybersecurity. The 3.5 Flash-Lite offers rapid performance at 350 tokens [...]