In partnership with
Good morning. It’s Friday, April 24th.
One of the most interesting dynamics in AI right now is China’s strategy for competing with the U.S.
Instead of trying to win by building the biggest closed company, they’re releasing frontier-level models open source and for free. It feels like a deliberate attempt to undercut the valuation of the leading American AI companies by giving the competition away.
DeepSeek V4 is a good example. It dropped yesterday, and early reactions suggest it rivals a lot of the top frontier models. Trust is still a major issue. Most people in the West aren’t eager to hand their data to Chinese-owned platforms, but that may not matter if the strategy is simply to give away the secret recipe itself.
If the model is free, the moat gets a lot smaller. It’s a fascinating approach, and it’s working well enough that it’s worth paying attention to.
-Jeff
AI Breakfast
You read. We listen. Let us know what you think by replying to this email.
Speak your PR description, bug reproduction, or Cursor prompt. Wispr Flow auto-tags file names, preserves variable names, and formats everything for immediate paste into GitHub, Jira, or your editor.
No re-typing. No context gaps. No mangled syntax. Works natively inside Cursor, Warp, and every IDE at the system level.
4x faster than typing. 89% of messages sent with zero edits. Used by engineering teams at OpenAI, Vercel, and Clay.
OpenAI releases GPT-5.5 as a "new class of intelligence" at double the API price
OpenAI just launched GPT-5.5, and it changes the interaction model from single prompts to agent-style execution, where the system can plan steps, use tools, and carry out full workflows across software. Scoring 82.7% on Terminal-Bench 2.0, it edges out Anthropic’s Claude Mythos Preview. Built on NVIDIA GB200 systems, it runs about 20% faster while keeping latency stable. Altman framed the update as a move toward iterative deployment, arguing that frequent, smaller releases are the best way to build global AI safety.
This agentic focus is most visible in the Codex app, which now features an in-app browser for navigating local dev servers and testing web flows. Codex also adds auto-reviewers for long-running automations and an official Google Sheets add-on that handles real-time spreadsheet edits and natural language formulas. For enterprise teams, OpenAI introduced workspace agents, an evolution of custom GPTs with persistent memory and dedicated cloud environments, allowing AI to work on projects even when the user is offline.
To secure deep-access workflows, OpenAI released Privacy Filter, a 1.5B-parameter open-weight model that locally redacts sensitive data using bidirectional token classification, achieving 97.43% F1 accuracy before cloud processing. New ChatGPT for Clinicians targets medical paperwork reduction, while a GPT-5.5 Bio Bug Bounty invites experts to find a universal jailbreak in its biological safety checks. Read more.
Anthropic patches Claude Code bugs and resets user limits
Anthropic has resolved the “AI shrinkflation” affecting Claude Code and the Agent SDK, confirming that performance issues were caused by product-layer bugs rather than model degradation. A technical post-mortem found that reasoning effort was lowered to reduce latency, a caching bug repeatedly cleared short-term context, and strict system prompts limited response detail. The issues affected Claude Code and the SDK (not the API). All have been fixed in v2.1.116, with usage limits reset, improved evaluations, and expanded dogfooding.
Claude now adds connectors to 200+ apps. The system can chain actions across apps in one chat, suggest tools dynamically, and requires user confirmation for purchases. Data is isolated, not used for training, with no ads or sponsored ranking.
The company’s latest survey of 81,000 users shows 48% value new capabilities over 40% speed gains, based mainly on personal users. Highest and lowest income groups report strongest gains. Creatives report limited usefulness and higher job loss anxiety; 20% overall worry about displacement.
Claude Managed Agents now feature a public beta for memory, enabling them to learn and persist data across multiple sessions. These workspace-scoped memories are stored as portable text files that developers can manage directly using the /claude-api command.
Anthropic’s implied valuation hit $1T on secondary markets, above OpenAI’s $880B, driven by rapid revenue growth from $9B to $30B and supply scarcity. A Freshfields partnership targets legal AI tools. A company manager also admitted that $20/$100 plans are “outgrown,” pushing a shift toward usage-based pricing. Legacy Grep/Glob tools were retired in favor of a faster Bash-based system. Read more.
Google debuts million-chip TPU clusters for massive AI training
Google is scaling its stack with 8th-gen TPUs, using the TPU 8t for million-chip training via the Virgo optical network and TPU 8i for inference. This hardware powers the Gemini Enterprise Agent Platform, adding persistent memory, sandboxed execution, and cryptographic identities to agents. New Workspace Intelligence now links Gmail, Docs, and Drive into a shared context layer for automated project history retrieval.
Google and Thinking Machines Lab signed a multibillion-dollar deal for infrastructure powered by Nvidia GB300 chips. This hardware supports Google's internal operations, where AI now generates 75% of new code. Additionally, Google expanded its capabilities with Gemini Embedding 2 for multimodal reasoning, Maps Imagery Grounding for Street View, and Earth AI for satellite-based infrastructure detection.
DeepMind’s latest drop, Vision Banana, unifies generation and perception through generative pretraining, handling segmentation, depth, and surface normals in one system. It beats benchmarks in zero-shot 2D and 3D tasks. Google also open-sourced DESIGN.md to help agents build brand-consistent, accessible UIs via CLI and Tailwind exports. Finally, a Gemma 4 demo showed 100% offline development using LM Studio and OpenCode to build sites and parse PDFs locally. Read more.
Text to 360 creates immersive panoramas from text or images, optimized for new ChatGPT Images 2.0.
Kollab is a shared workspace where teams and agents work together across Slack, apps, and projects.
Monid provides a Web3 wallet for pay-per-call access to cross-platform data through a unified agentic API.
SpeakON is a MagSafe wearable providing instant AI dictation and app-aware text input without unlocking your phone.
Magic Patterns Agent 2.0 creates high-fidelity prototypes that match existing design systems for rapid engineering handoff.
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.
