AI Models Just Got Cheaper?

Anthropic launched Claude Opus 5.5, their latest model since the 'pace the frontier' essay by Dario Amodei. This model emphasizes better safety features to reduce risky or harmful outputs, boasting an 85% reduction in attempts to circumvent boundaries compared to previous models. It performs on par with Fable 5.1 on most tasks, but at around 40% lower operational costs, using fewer tokens and generating responses over 30% faster. Cost metrics indicate it costs $4 per 1 million input tokens and $20 per million output tokens, with total operational costs approximately 40% lower than Opus 5. This reduction is partly driven by its ability to reroute cybersecurity requests to less powerful models and biology-related requests to its Safeguard. Anthropic highlights that the model has been tested by outside partners like Meta and Frontier design and will soon introduce new variants, Claude Sonnet 5.5 and Haiku 5.5.

OpenAI revealed GPT-6 Soul and Luna, claiming that Soul reduces factual errors by about 50% compared to GPT-5.6, making it the more accurate engine in their portfolio. Luna, positioned as a lower-cost option, reportedly performs as well as previous higher-tier models like GPT-5.6 at roughly 1% of the cost, emphasizing the ongoing trend of delivering more affordability. These models are part of OpenAI's rapid iteration cycle, with GPT-6 Luna's performance improving substantially in just three months. The company highlights that the new models deliver equivalent quality at a fraction of previous costs, crucially effectively doubling the factual accuracy of their job-replacement AI models within a short timeframe. The focus remains on achieving better and cheaper AI services.

YouTube introduced a suite of new AI-powered tools aimed at content creators, including Shorts Series, which allows creators to organize short videos into seasons and episodes with features like custom thumbnails and continuous storytelling. This supports the trend of episodic content in Shorts, aligning with viewers’ preference for microdramas. Additionally, YouTube announced an AI agent that scans a creator’s existing videos to identify trending content, provide optimization suggestions, and generate thumbnails and titles automatically, making it easier to improve and repurpose content. For testing, creators can now run multiple versions of videos with different intros or hooks, then automatically select the best performer after testing on small audience segments. Also, YouTube rolled out Custom Feeds, a major innovation that lets users tailor their homepage via natural language prompts, giving them control over what videos appear, beyond traditional thumbs-up or block options. This feature is currently being tested in the US, aiming to make content discovery more personalized and dynamic.

YouTube's new Custom Feeds feature allows users to generate and save personalized homepage feeds based on natural language prompts. Users can describe the vibe, topic, or specific content they want to see, and YouTube’s large language models process their requests to produce customized feeds. The feature aims to give viewers more control over their content, moving beyond generic recommendations and allowing complex, multi-topic queries. It’s integrated into the interface with a button labeled 'Your Custom Feeds,' which can be used to create multiple feeds that blend various interests. YouTube also tested a broader feature that would alter the recommendation algorithm directly, but this was ultimately not launched due to user concerns about messing up their feed. Instead, the focus is on providing a low-stakes environment for experimenting with content preferences, making content discovery more aligned with user moods and specific interests.