Good morning. It’s Monday, June 1st.

You read. We listen. Let us know what you think by replying to this email.

A developer update for Anthropic’s Claude Opus 4.8 introduces a change in how system instructions work during active conversations. Developers can now add or modify system-role messages in the middle of a session without invalidating the prompt cache. Earlier, any change to top-level system prompts would reset cached context, increasing compute cost and latency. With the new behavior, updated system instructions are treated as authoritative from the point they are inserted, while the existing cached tokens remain usable.

This change fits right into a leaked rollout plan to turn Claude from a standard chatbot into a multi-agent platform. Code leaks reveal "Conway," a persistent background agent that runs inside isolated containers and triggers via webhooks using a new .cnw.zip extension standard.

To handle specific workloads, this system connects to dedicated tools: "Orbit" proactively runs tasks across apps like Slack and Gmail; "Operon" automates data pipelines in scientific sandbox environments; "BugCrawl" handles autonomous codebase debugging; and a file-based memory system saves state across separate sessions.

Meanwhile, an Anthropic study on how social scientists use tools like Claude Code reveals that researchers with traditionally male names use AI coding agents more than twice as often as peers with traditionally female names across identical career ranks. Economists lead adoption at 39%, while education researchers sit at 4%. Coding agents are mainly used for data analysis (97%), while writing tasks remain less common. Read more.

OpenAI has expanded its Codex system with Computer Use support for Windows 11, enabling the AI to directly operate desktop apps, files, and system tools. The update lets the AI run multi-step tasks like app testing, bug hunting, and workflow validation without active user supervision. Users can trigger precise control via commands like @computer or app-specific targets like @Paint. Previously limited to macOS, this desktop automation is now accessible on mobile via the ChatGPT app for iOS and Android, allowing users to initiate or monitor Windows tasks remotely.

At the same time, OpenAI updated GPT-5.5 Instant to produce clearer, more natural responses with improved structure and fewer long bullet lists. This rollout shifts editing and coding workflows out of the Canvas feature, which is being removed from both GPT-5.5 Instant and GPT-5.5 Thinking in favor of direct, structured response blocks inside the main chat. Alongside these interface tweaks, OpenAI is retiring older models: GPT-4.5 will leave ChatGPT on June 27, 2026, followed by o3 on August 26, 2026, though o3 remains active in the API.

OpenAI also launched Rosalind Biodefense to support trusted developers building tools to detect, prevent, and respond to biological threats. The company is also expanding access to GPT-Rosalind for selected US government agencies and allied partners. Read more.

Grok-Imagine-Video-1.5 Preview (720p) has reached the #1 position on Arena.ai’s image-to-video leaderboard, scoring 1473 points and improving by +52 over its previous version. It surpasses competing systems such as Bytedance’s Seedance-2.0 and Alibaba’s HappyHorse-1.0, as well as earlier Grok releases. The ranking reflects 1.1 million user votes across 40 models, evaluating motion quality and visual realism, with other strong contenders including Google’s Veo series and KlingAI models.

In parallel, xAI has released grok-build-0.1 in public beta through its API. This is the same model used in the Grok Build CLI and is optimized for agentic coding workflows such as automated code generation, tool execution, and software task completion.
The model is priced at $1 per million input tokens and $2 per million output tokens, positioning it as a low-cost option for large-scale use. It is also available through OpenRouter and Vercel AI Gateway and integrated into developer tools such as Cursor and OpenClaw, extending its reach across software engineering environments. Read more.

Second Brain unifies your context, providing a self-hosted semantic memory layer across all your AI tools.

/monitor by Firecrawl triggers webhooks for web changes, cutting AI agent token spend by up to 90%.

Miora by Tencent is an AI creative studio generating images, video, and design on a single, unified canvas.

Wandesk builds an open source AI desktop with connected local apps and persistent memory on your machine.

Ornold is an MCP server giving AI agents vision-first control and parallel automation over antidetect browsers.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.