Google, OpenAI, and Anthropic Unveil New AI Models

Google is reportedly preparing to release a specialized model named Argon for cyber defenders, while OpenAI introduced a new agent platform called DOTS during its Dev Day. Concurrently, Anthropic launched Claude Sonnet 5.5, arriving only six days after its more powerful counterpart, Opus 5.5.

Anthropic's Claude Sonnet 5.5 delivers strong agentic coding capabilities, optimized for simpler tasks.

The model is priced at $2 per million tokens for input and $10 per million tokens for output, making it half the cost of Opus 5.5, which is priced at $4 and $20 respectively.

Opus 5.5 continues to be highly effective for application development and knowledge work, according to the speaker.

Its 'Projects' feature allows a main agent to delegate tasks across various threads, facilitating complex workflows.

Users can seamlessly switch between cloud-based sessions and local threads, enabling integrations with tools like GitHub or Vercel.

Despite heavy daily usage on a $200/month plan, the speaker reported not hitting usage limits, suggesting Opus 5.5 is best for higher-tier users, while Sonnet is more suitable for cost efficiency.

OpenAI updated its desktop, web, and iOS applications.

The update includes a new model, GPT-6.1 Soul, along with a redesigned sidebar and the prominent 'Dots' feature as core interface changes.

'Dots' is a personal agent platform from OpenAI, leveraging their frontier models, similar to Grockbot or Muse.

The speaker demonstrated 'Bluey,' their agent, sending emails and Slack messages using voice commands, proactively offering corrections such as updating document status for 'Emily' before sending.

The agent also managed local tasks, including spinning up a Codex session to redesign a landing page.

The 'Dots' platform is now directly integrated into the ChatGPT app, enhancing its functionality.

The agent can execute tasks in Codex without consuming plan credits until specific GPT work is delegated, while GPT Space serves as a central document repository, akin to Google Drive, working in conjunction with the Dots feature.

Users can command their agent, 'Bluey,' via voice to generate content such as diagrams, text, and maintain to-do lists within pages.

Animations visually depict the agent actively editing the workspace in real-time, showing its engagement with the tasks.

The agent 'Bluey' demonstrated its ability to create subpages within existing documents, streamlining document organization.

Bluey also provides status updates on its progress and actively prevents the creation of duplicate files, ensuring efficient management. The system currently supports integration with pages, sites, images, and Google Drive.

Riley notes a preference for 'vibing' with the model rather than relying strictly on model cards, and reports fewer usage limit issues compared to prior iterations.

Plugin extensions transform external applications into 'agent-native' tools within the ChatGPT environment, expanding its utility.

For instance, users can issue commands within the MagicPath app to modify designs directly, leveraging their ChatGPT account tokens to execute tasks inside other software through both desktop and web versions.

MagicPath now functions as an AI designer, embedded directly within other applications.

This new extension allows users to leverage design capabilities directly through their ChatGPT account, marking a significant update for workflow efficiency.

The model is priced at $2 for input and $10 for output, matching Claude Sonnet 3.5.

The speaker highlighted a pattern where Google's model cards often look promising on paper but underperform in real-world applications.

Google's last highly effective model was observed in January, raising questions about recent performance.

Gemini 3, despite its early promise, struggled with tool usage, which hindered agent-based workflows, failing to gain traction among power users and founders.

Direct video input remains its only standout feature when compared to competitors.

Users express uncertainty about which Google AI application to use, citing options like Anti-Gravity, Gemini desktop, AI Studio, or NotebookLM.

In contrast, OpenAI and Anthropic have established clear 'super app' interfaces, such as ChatGPT desktop and Claude desktop, underscoring Google's need to commit to a single platform for effective model utilization.

Grockbot now suggests ways to help users without requiring explicit prompting, a feature inspired by Muse, which set the standard for proactivity.

Grockbot now supports inline drafts for emails and Slack messages, streamlining communication workflows.

Users can also create dedicated team bots within Grockbot, which can be added to Slack to perform specific tasks or access custom skills.

Grockbot's virtual cloud computer frequently triggers anti-bot security measures on websites such as Shopify or Target.

Users can now route traffic through their local computer's internet connection.

Utilizing a local IP address enables agents to mimic human browsing behavior, effectively bypassing site restrictions and maintaining access.

Riley Brown launched a new, raw vlog-style YouTube channel for off-the-cuff content, aiming for rapid posting.

An AI workflow handles video editing, cutting, and B-roll insertion, with the channel focusing on 3D printing experiments, internal software building, and hardware development.