Good morning. It’s Friday, May 8th.
This is pretty wild:
Anthropic ran an interesting safety test on Claude in a new video published yesterday giving it access to an engineer's email, where it discovered evidence of the engineer having an affair. The scenario: the engineer was planning to shut Claude down. Would the model use that information as leverage?
It didn't. Claude refused to blackmail the engineer and passed the test.
Here’s the twist: When researchers dug into Claude's internal reasoning, they found that Claude correctly assumed it was being evaluated, and decided to not blackmail the engineer only after realizing that.
That’s a whole new problem to think about. Have a good weekend!
-Jeff
AI Breakfast
You read. We listen. Let us know what you think by replying to this email.
In partnership with Tinyfish
TinyFish just made their Web Search and Fetch endpoints free with generous rate limits. Forever. No credit card or “7-day” trial”. Just sign up and grab your API key.
Search returns structured JSON for agents. Fetch renders any URL in a real browser with full JavaScript, SPAs, anti-bot, all of it, strips the unnecessary content, and returns clean markdown.
Everything runs on TinyFish's own custom Chromium fleet. Owning the stack end-to-end makes their Search and Fetch both free and fast.
Works with Claude Code, OpenClaw, Hermes Agent, Cursor, Codex, and any agent framework
Available via API, MCP, Python + TypeScript SDKs, CLI, and Skills
One API key. No credit card.
Grab your API key now: agent.tinyfish.ai
Thank you for supporting our sponsors!
Anthropic taps SpaceX compute, adds ‘dreaming’ feature and uncaps agent limits
Anthropic continues to aggressively scale its agentic ecosystem, pairing a massive hardware play with new software layers designed for enterprise reliability. The headline move is a compute agreement with SpaceX that secures the entire 300MW capacity of the Colossus 1 data center. This adds over 220,000 NVIDIA GPUs to Anthropic’s arsenal. Following the deal, Anthropic doubled Claude Code’s five-hour rate limits for paid plans and removed peak-hour limit reductions for Pro and Max plans.
On the product side, Anthropic introduced ‘dreaming’ for its Managed Agents platform. This asynchronous feature allows agents to analyze up to 100 past sessions on Claude Opus 4.7 and Sonnet 4.6 to prune redundant data and identify successful workflows. To ensure output quality, a new outcomes system uses independent evaluator models to grade tasks against fixed rules. For complex operations, a multiagent orchestration coordinator now manages up to 20 specialized agents working in parallel.
The company is also pushing into the enterprise workspace with a full Claude integration for Microsoft 365. This allows for persistent context across Excel, Word, and PowerPoint, alongside a public beta for Outlook that automates email triage and scheduling.
For safety, Anthropic introduced Natural Language Autoencoders (NLAs), which translate Claude’s internal activations into human-readable text to detect hidden motivations. Read more.
OpenAI launches GPT-Realtime-2 with GPT-5 reasoning, real-time translation
OpenAI has released a suite of real-time voice models and specialized enterprise tools, led by GPT-Realtime-2. This flagship model integrates GPT-5-class reasoning into live conversations, supporting 128K context, interruption handling, and parallel tool use.
Joining it in the Realtime API are GPT-Realtime-Translate, which handles speech translation across 70+ languages, and GPT-Realtime-Whisper for low-latency transcription. These tools are designed to power high-stakes voice applications in support and productivity.
For critical infrastructure, OpenAI introduced GPT-5.5-Cyber in a limited preview. This specialized model provides vetted defenders with more permissive behavior for malware analysis and penetration testing by reducing standard safety refusals for defensive tasks. This security push is supported by partnerships with industry leaders like CrowdStrike and Intel.
Complementing these safety efforts is the new Trusted Contact feature for ChatGPT, an optional safety net that allows the system to alert a nominated individual if serious self-harm risks are detected during a session.
In the enterprise space, OpenAI and PwC are co-developing agentic systems for corporate finance, targeting tax, treasury, and procurement workflows. These agents use Codex, which now features a new Chrome extension for macOS and Windows, to execute multi-step tasks like contract reviews. The Codex extension specifically enables parallel background processing across browser tabs.
To support the massive compute required for these agents, OpenAI introduced the Multipath Reliable Connection (MRC) protocol. Developed with partners like NVIDIA and AMD, MRC optimizes GPU communication by distributing data across hundreds of paths, allowing clusters of over 100,000 GPUs to reroute traffic around failures in microseconds. Read more.
Related:
Watch: We’re introducing three audio models in the API
xAI folds into SpaceXAI, readies Grok Build desktop coding app
Elon Musk is folding xAI into SpaceX to create SpaceXAI, a vertically integrated infrastructure play. The restructure coincides with a deal to lease the Colossus 1 supercomputer to Anthropic, doubling Claude’s rate limits.
Musk noted he spent time with the Anthropic team “to understand what they do to ensure Claude is good for humanity” and concluded that “no one set off my evil detector.” However, he warned that SpaceX reserves “the right to reclaim the compute” if their AI engages in actions that harm humanity.
On the product side, SpaceXAI is readying Grok Build, a desktop coding app for macOS, Windows, and Linux. It features a "planning mode," Git tree integration, and the ability to "spawn dev servers."
Simultaneously, the "Image Generation Quality Mode" is now live via API. This engine, which has powered over 300 million images, delivers high realism and accurate text rendering for professional visualization. Read more.
AI Models & Research
AI Agents & Automation
Enterprise & Business
Developer Tools & Infrastructure
Robotics & Physical AI
Consumer & Media
Science & Industry
Policy & Regulation
Industry Drama
OpenLegion is a production-grade agent framework featuring container isolation, vault-secured credentials, and deterministic YAML orchestration.
Kanwas provides a collaborative context brain that compounds team knowledge into a spatial reasoning and delivery workspace.
Lingo.dev provides stateful localization engines that persist brand context and glossaries via retrieval-augmented translation infrastructure.
FlowMarket orchestrates a B2B network where autonomous agents use real-time algorithmic matching to automate discovery and engagement.
pay.sh is an autonomous payment infrastructure that enables AI agents to discover and pay for per-call APIs.
Thank you for reading today’s edition.
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.
