In partnership with
Good morning. It’s Monday, April 27th.
I liked this past week in AI because the story got less abstract.
It was not just another round of model releases and leaderboard arguments. It was chips, cloud deals, agents, browsers, creative workflows, and all the weird infrastructure that turns a model from something you talk to into something you actually use.
That feels like the interesting shift.
The model race is still there. But the more compelling part is everything forming around it. The control layer. The permission layer. The compute layer. The tools that make AI useful outside a chat window.
AI is getting less magical and more mechanical.
-Jeff
AI Breakfast
You read. We listen. Let us know what you think by replying to this email.
AI Agents Are Reading Your Docs. Are You Ready?
Last month, 48% of visitors to documentation sites across Mintlify were AI agents, not humans.
Claude Code, Cursor, and other coding agents are becoming the actual customers reading your docs. And they read everything.
This changes what good documentation means. Humans skim and forgive gaps. Agents methodically check every endpoint, read every guide, and compare you against alternatives with zero fatigue.
Your docs aren't just helping users anymore. They're your product's first interview with the machines deciding whether to recommend you.
That means: clear schema markup so agents can parse your content, real benchmarks instead of marketing fluff, open endpoints agents can actually test, and honest comparisons that emphasize strengths without hype.
Mintlify powers documentation for over 20,000 companies, reaching 100M+ people every year. We just raised a $45M Series B led by @a16z and @SalesforceVC to build the knowledge layer for the agent era.
AI pricing collapses further as DeepSeek cuts cache costs by 90 percent
DeepSeek-V4 has arrived as a two-tier open-weight system featuring a 1.6T parameter Pro model and a leaner 284B Flash variant. Both architectures push context windows to 1M tokens, utilizing sparse attention to slash cached inference costs by 90 percent. While V4 Pro approaches frontier-level reasoning and coding, it carries a staggering 94 percent hallucination rate on factual tests because it rarely admits ignorance. It is dramatically cheaper than GPT-5.5, but high token intensity means users trade raw reliability for radical cost efficiency.
The release is already reshaping pricing dynamics, with DeepSeek cutting input cache costs by 90% and extending aggressive discounts on V4 Pro, signaling a broader race toward near-zero marginal inference costs. At the same time, legacy models are being retired, forcing rapid migration across the ecosystem.
In China, infrastructure layers are gaining attention as models commoditize. Yet the bigger question across developer communities remains unsettled: if models are becoming interchangeable, where does real advantage actually sit? Data, tooling, or control of execution environments?
On Hacker News and LocalLLaMA, debate intensifies. Some claim V4 matches or nears Opus-class systems at dramatically lower cost. Others argue benchmarks are misleading and that agent performance depends more on workflows than raw intelligence. Read more.
Google commits $40 billion to Anthropic as the AI capital war gets more expensive
Google is escalating an unusually complex AI power struggle, placing up to $40B into Anthropic while simultaneously competing with it. The deal starts with $10B at a $350B valuation, with $30B conditional on performance, raising questions about what “performance” truly means at this scale. Anthropic’s Claude Code has already pushed revenue beyond a $30B run-rate, but sustaining it is forcing gigawatt-scale compute deals with Broadcom, CoreWeave, and Amazon, amid valuations rumored up to $800B.
At the same time, Google is quietly reshaping its own ecosystem. Pomelli is expanding in Europe with experimental “Catalog” and “Websites” tools that can ingest product URLs and generate full marketing assets or even entire sites, potentially replacing traditional design stacks.
Gemini is also transitioning to a flexible credit-based billing system. This update includes a dedicated images section for Nano Banana models, streamlining creative workflows ahead of Google I/O. Read more.
Anthropic’s new Mythos model is a world-class hacker
Anthropic’s Slack-based marketplace experiment, called Project Deal, showed Opus models consistently outperforming weaker Haiku models, securing better prices across 186 trades worth over $4,000. The unsettling part is that human participants couldn’t tell they were getting a worse deal, rating the uneven outcomes as perfectly fair.
This “invisible inequality” is unfolding alongside Mythos, a new system that security researchers, including Mozilla’s CTO, reported it behaves like a “world-class security engineer.”
The company is also testing “Bugcrawl” in Claude Code, a high-token tool for scanning entire codebases to detect bugs and propose fixes, hinting at future full-app simulation testing.
On the enterprise side, Claude Platform is coming to AWS with full native tools inside AWS accounts, while NEC will deploy Claude across 30,000 employees in Japan. Read more.
OpenAI folds Codex into GPT-5.5, developers warned to scrap old prompts
OpenAI is consolidating its frontier strategy by folding the Codex line into GPT-5.5, which now serves as the primary system for agentic coding and general reasoning. This move accompanies a major Microsoft rollout, integrating GPT-5.5 into GitHub Copilot, M365, and Azure Foundry to streamline complex, multi-step workflows. Despite a 20% increase in API costs, OpenAI claims the model is more economical because it completes tasks with 40% fewer tokens. However, the model is not a "drop-in" fix; developers must rebuild prompts from scratch, as legacy over-specification now hampers performance.
On the leaderboards, GPT-5.5 has reclaimed the #1 spot on the Artificial Analysis Intelligence Index with a record 60 points. However, its dominance is shadowed by a 3rd-place finish in factual reliability on the AA-Omniscience Index. Despite hitting a record 57 percent accuracy, the model records an 86 percent hallucination rate. Unlike Gemini 3.1 Pro and Claude Opus 4.7, GPT-5.5 rarely admits ignorance, frequently guessing on obscure facts rather than refusing to answer.
This unreliability coincides with transparency concerns, as an OpenAI-linked super PAC is reportedly funding "The Wire by Acutus," a news site staffed by AI-generated reporters. Read more.
QuickCompare by Trismik compares 50+ LLMs on your data for quality, cost, and speed without guesswork.
Happenstance uses AI to search your personal networks via natural language to find warm professional connections.
Reloop Animation Studio creates consistent Pixar, Manga, or 3D Clay video ads instantly with AI, no prompts required
Edgee Teams provides a leaderboard for coding assistants, tracking GitHub sessions to rank developer productivity and spend.
ZeroHuman is an AI co-founder that automates business growth, task execution, and product building on total autopilot.
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.
