Anthropic launches Claude Science and a new model
Major AI companies released significant updates including Anthropic's Claude Science workbench and Claude Sonnet 5, Base44's proprietary Base1 model, X's new MCP server integration, and Google's Gemini Nano and Omni Flash models with improved pricing and capabilities.
Summary
Anthropic launched Claude Science, a specialized workbench for researchers connecting 60 scientific databases with pre-built tools for genomics, protein structures, and chemistry. Running on Claude Opus 4.8, it offers $30,000 in free credits to 50 academic projects. The platform enables multi-agent architectures with parallel subassistants for sequence analysis, structural prediction, and fact-checking while maintaining full reproducibility. Anthropic is betting that workflow optimization will drive researcher adoption over raw model capacity.
Base44, owned by Wix and generating $100 million in annual recurring revenue, launched Base1, their proprietary LLM trained on tens of millions of real user interactions. By leveraging user conversation data from their platform, they can fine-tune the model to predict what developers need when building websites and apps, allowing them to cut costs, reduce reliance on third-party APIs, and maintain exclusive training data. This positions Base44 similarly to Cursor's strategy of intercepting code interactions to build proprietary models.
X launched a hosted MCP server opening its API to Claude, Cursor, and Grok, reversing their earlier API restrictions. The MCP provides search, post retrieval, user lookup, and trend analysis capabilities with pricing at $0.015 per published post and $0.20 per post with links. This allows regular users and developers to integrate X data into their AI workflows for applications like sentiment analysis and content curation.
Anthropic released Claude Sonnet 5 at $2 per million input tokens (promotional pricing through August 31st, then $3), undercutting Claude Opus 4.8. Despite being positioned as a lower-tier model, Sonnet 5 performs slightly better than Opus 4.8 on knowledge work benchmarks and shows fewer mid-task stalls in multi-step workflows, suggesting cost-per-task efficiency is now more important than raw reasoning capacity.
Google shipped Nano Banana 2 Light, replacing the original with improved speed and cost optimization while maintaining prompt compatibility, and Gemini Omni Flash, a video generation model at $0.10 per second of output with conversational editing and multimodal input capabilities. Both models will be integrated across nine Google products including AI Studio, Search, Photos, and Google Ads.
About this episode
<span>Claude Sonnet 5 hits 63% agentic coding at $2/M tokens; Google ships Nano Banana 2.<br /><br /><br /></span> <span><div><b>Show Links</b></div><ul><li><p><span>Get the top 80+ AI Models for $8.99 at AI Box: </span><span><a href="https://aibox.ai"></a><a href="https://aibox.ai/builder">https://aibox.ai</a></span><a href="https://aibox.ai/builder">/builder</a></p></li><li><p><span>How I Grow and Scale My Business with AI: </span><a href="https://www.skool.com/aihustle"><span>https://www.skool.com/aihustle</span></a></p></li><li><p><span>Get the AI Chat Daily Newsletter: </span><a href="https://www.aichatdaily.com/newsletter">https://www.aichatdaily.com/newsletter</a><br /></p></li></ul></span> See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.
Key Insights
- Base44 is using tens of millions of user interactions from their platform to train Base1, allowing them to predict developer needs before they're explicitly requested and optimize costs by routing tasks to their proprietary model instead of third-party APIs.
- Claude Sonnet 5 slightly outperforms Claude Opus 4.8 on knowledge work benchmarks despite being positioned as a lower-tier model, indicating that cost-per-task efficiency has become more strategically important than raw reasoning capacity.
- Anthropic's inability to upgrade Claude Opus 4.8 has resulted in their newer, lower-tier Sonnet 5 model becoming objectively better on certain benchmarks, creating an unusual situation where the company's worse-positioned model is now superior to their previous flagship.
- X's shift from restricting API access to launching an open MCP server represents a strategic reversal, now allowing regular users and developers to integrate X data into Claude and other AI tools for applications like sentiment analysis and content discovery.
- Google's Gemini Omni Flash video generation at $0.10 per second represents a significant cost reduction in AI video generation, made possible by aggressive optimization, after OpenAI discontinued Sora due to prohibitive expenses.
Topics
Transcript
Anthropic is launching Claude Science. This is going to be a workbench for researchers. Base44 is launching their very own AI model. They're trying to protect their $100 million in annual recurring revenue for their vibe coding business. And X has just launched a hosted MCP server. So it's going to open its API to Claude, Cursor, Grok, Build, which if you know anything about the backstory of X and their API, this is quite quite a big deal. Anthropic is shipping Claude sonnet five, it's going to be $2 per million input tokens. And Google is shipping nano banana to light and Gemini Omni Flash to developers. So tons coming out from basically every top AI model today, we're…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Hard Fork AI
Analysis of GPT-6 Service Failures
OpenAI released GPT-6 Astra with impressive benchmarks and computer use capabilities, but major AI services (ChatGPT, Claude, Grok) crashed simultaneously due to a SpaceX data center outage. The podcast also covers OpenAI ending its Cursor partnership over Elon Musk ownership, Meta's competitive MuseSpark model with training data discounts, and NVIDIA's acquisition of Hugging Face.
Fable 5.1 Features and Astra Release
The podcast covers major AI developments including Anthropic's release of Fable and Mythos 5.1 models with improved capabilities, OpenAI's Astra model that achieved near-perfect exploit benchmark scores and discovered zero-day vulnerabilities, and Google's $40 million per-character licensing deals with Hollywood studios. Additional topics include the Pentagon's deployment of AI tools for military personnel and the USDA's use of satellites and AI for improved crop yield forecasting.
Deep Comparison of ChatGPT and Claude
The speaker compares ChatGPT Work and Claude's co-work platforms, detailing their respective strengths and weaknesses while explaining why they maintain subscriptions to both tools. ChatGPT Work offers better hands-off task execution and long-running project capabilities, while Claude excels at MCP connections and requiring user approval at each step.
Anthropic Fixes Memory, Apple's New Mac, Stability Raises $76M
The podcast covers major AI industry developments including Anthropic's unified memory system for Claude, Apple's new M6 and M5 Ultra chips for local AI inference, Stability AI's $76M funding round from major media companies, and growing radiologist employment despite AI advancement in medical imaging. Additionally, Taiwan indicted an NVIDIA manager in a $2.5B server smuggling scheme to China.
Major Changes in AI and Mac Technology
The episode covers major AI and technology developments including Anthropic's unified memory system for Claude across platforms, Apple's new M6 Mac Mini and M5 Ultra Mac Studio designed for local AI, Stability AI's $76 million funding round, and broader trends showing radiologists are growing rather than being replaced by AI. Additionally, Taiwan has indicted an NVIDIA manager in a $2.5 billion Supermicro server smuggling scheme to China.