Large Language Models

Today's Stories

1. 1

   DeepSeek R1 becomes #2 reasoning model, rivals OpenAI o1

   DeepSeek released R1, an open-weights reasoning model that has become the #2 open-weights reasoning model. It matches or exceeds the performance of OpenAI's o1 on several benchmarks while using significantly less computing power—27% of the FLOPs compared with DeepSeek-V3.2. DeepSeek R1's efficiency and strong performance signal that high-capability AI reasoning is becoming more accessible and affordable to build. For businesses and developers, this may lower the barriers to deploying advanced reasoning capabilities without relying solely on proprietary, expensive models like o1.

   DeepSeek R1 is available as open-weights code, meaning researchers and developers can download and run it themselves. The model demonstrates that competitive reasoning performance does not necessarily require vastly more computational resources, potentially reshaping how organizations approach AI deployment.

2. 2

   What to know about Anthropic's new Claude 'Sonnet 5' AI model

   What to know about Anthropic's new Claude 'Sonnet 5' AI model

3. 3

   Announcing Anthropic Claude Sonnet 5 on Snowflake Cortex AI

   Announcing Anthropic Claude Sonnet 5 on Snowflake Cortex AI

4. 4

   Google launches Nano Banana 2 Lite and Gemini Omni Flash for faster AI images and video

   Google released two new AI models—Nano Banana 2 Lite for fast image generation at $0.034 per image in four seconds, and Gemini Omni Flash for video generation and editing via API at $0.10 per second of output. Both are now available to developers through the Gemini API and Google AI Studio. Nano Banana 2 Lite replaces Google's older image model and offers a lower-cost, faster option for developers who prioritize speed over quality. Gemini Omni Flash opens video generation to the API for the first time, allowing developers to combine text, images, and video in a single workflow—and Google recommends chaining both models together to quickly generate images and animate them into video.

   Gemini Omni Flash currently generates only ten-second clips, and audio references and scene extensions are not yet supported in the API. Character consistency across scene changes remains limited. Both models apply SynthID watermarks to tag AI-generated content, with verification available through the Gemini app, Gemini in Chrome, and Google Search.

5. 5

   Anthropic launches Claude Sonnet 5, narrowing gap to pricier Opus model

   Anthropic released Claude Sonnet 5, a new AI model that matches or beats its larger Opus 4.8 sibling on several benchmarks. On knowledge work (GDPval-AA v2), Sonnet 5 scores 1,618 points versus Opus 4.8's 1,615. The model is available now across all Anthropic plans, with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026. Anthropic is positioning Sonnet 5 as capable of autonomous work—planning, using tools like browsers and terminals—that previously required more expensive models. The launch comes as the US government blocks Anthropic's two most capable models (Mythos 5 and Fable 5) over cybersecurity concerns, so Sonnet 5 sidesteps that issue: it scores far below those models on risky capabilities like writing software exploits, and Anthropic has enabled cyber safeguards by default. For businesses and developers, this suggests a cheaper way to access agent-like capabilities without premium pricing.

   Sonnet 5 does chew through more tokens per task due to its agentic behavior—meaning real-world costs could exceed the per-token savings. The model has a one-million-token context window and a training cutoff of January 2026. Prices rise to $3 per million input tokens and $15 per million output tokens after August 31, 2026.

6. 6

   OpenAI cuts AI inference costs by more than half

   According to The Information, OpenAI has reduced inference costs for its AI models by more than half. The company applied these optimizations to ChatGPT, cutting the number of Nvidia GPUs needed to just a few hundred at times. Lower inference costs—the expense of running an AI model to produce an answer—could make AI services more affordable to operate and potentially more accessible to users, though the business implications depend on how OpenAI passes these savings along.

   The report does not specify when these optimizations took effect, what specific techniques drove the cost reduction, or how the savings will be reflected in ChatGPT pricing.