Good morning. It’s Monday, August 24th.

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

An anonymous reasoning model called Ox Alpha surfaced on OpenRouter, drawing attention for its 1M-token context window, 131K-token max output, tool calling, structured outputs, and multimodal support. Early buzz accelerated when a viral subset test put it at an 80% pass rate across 10 DeepSWE tasks, outperforming Claude Fable 5 and GPT-5.6 Sol, though broader evaluations on the full 113-task suite paint a more modest 63% score that roughly matches GPT-5.6.

A cryptic post from DeepMind researcher Evan Otero sparked rumors that it might be a stealth Gemini Pro release. But when you look at the actual empirical evidence, tokenizer behavior, error signatures, and system artifacts align directly with Zhipu AI's GLM-5.3, making a Chinese lab origin the most parameter-consistent explanation. Nothing has been confirmed.

Google DeepMind is also pushing AI beyond traditional benchmarks, training agents in an isolated EVE Online server to test long-term memory, continual learning, planning and multi-agent behavior. The research could eventually move into EVE Frontier, where agents would face changing rules and persistent economies. Google says the goal is to develop AI that can enrich games rather than replace players.

Google Antigravity is adding Remote Control, letting users access active sessions from supported browsers and iOS and Android devices.

Meanwhile, in an 𝕏 post sharing what he called an excellent Economist article, DeepMind VP Zoubin Ghahramani argues that AI consciousness may need to be judged without assuming intelligence has to look human.

Anthropic is testing two early-access Claude models under the codenames Marshmallow and Melon, according to developer sightings. The models, identified as claude-marshmallow-eap and claude-melon-eap, could eventually become Opus 5.1 and Sonnet 5.1. Marshmallow reportedly beats the current Opus 5, but neither model is said to match Fable 5. Anthropic hasn’t confirmed the models or their release plans.

At the same time, Anthropic is expanding Claude Mythos 5 for cybersecurity. Claude Security can now use Mythos 5 to scan Enterprise codebases, identify vulnerabilities and suggest patches for human review. Anthropic is also working with security vendors to integrate Mythos into their tools and has committed $35 million in Claude credits to help open-source projects find and fix vulnerabilities.

Anthropic is also addressing complaints about Opus 5 being too verbose. Engineer Boris Cherny says the model is strong at optimization tasks such as reducing CPU usage, CI times and latency. For now, Claude Code users can use claude /config outputStyle=concise to shorten responses.

Nvidia’s AVO coding agent scored 100% on the ARC-AGI-3 benchmark, completing all 183 levels across 25 public environments without instructions, explicit rules or stated goals. The result suggests the agent can infer objectives and adapt its approach in unfamiliar interactive tasks.

Nvidia is also warning customers that AI servers using Vera Rubin and Grace Blackwell chips could cost more than 15% extra early next year. Rising DRAM prices from Samsung, SK Hynix and Micron are driving the increases as AI data center demand continues to outpace memory supply. Microsoft, Google and Oracle have already been notified, while Amazon, Meta, OpenAI and Anthropic remain heavily dependent on Nvidia despite developing their own chips.

Nvidia is investing in the infrastructure needed to keep that demand growing, taking a minority stake in data center developer Cloverleaf Infrastructure in a deal potentially worth several hundred million dollars. It also committed $1.5 billion to SB Energy’s OpenAI-linked Ohio project.

Separately, Nvidia is reportedly considering a multibillion-dollar investment in Perplexity at a valuation above $30 billion. Perplexity’s annualized revenue has passed $750 million, helped by its Computer AI agent.

fx (by Vercel) is a lightweight open-source coding agent designed to run fast and give models room to work.

HyNote for Mac provides free, fully local transcription that captures system audio across meeting apps without using bots.

FetchSandbox MCP reproduces real integration failures on your code, fixes them, and proves the fix held with 70+ sandboxes.

Epho runs cloud agents on your repo with a single API call while managing sandboxes and streaming automatically.

Mindcase provides web data APIs that supply structured data across multiple platforms directly to AI agents.

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.