1. A 35B Model Says It Matches a Trillion Parameters by Running Longer, Not Growing Bigger Three model releases in the same week share a thesis: the lever for agentic performance is task duration, not parameter count.

2. Anthropic built Claude a lab bench the same week OpenAI built it an exam On Tuesday, in front of pharmaceutical executives, biotech founders, and researchers, Anthropic announced Claude Science, a product meant to do for scientific research what Claude Code does for

3. The rainbow flowers in those seed listings were never real plants A buyer browsing eBay, Amazon, or Etsy lands on a photo of a flower too vivid to ignore: petals in impossible color gradients, a bloom no garden has produced. The seeds cost a few dollars.

In Brief

• Anthropic ships Claude Sonnet 5 at lower prices for agent workloads Anthropic released Claude Sonnet 5 with stronger agentic performance and reduced pricing, positioning it below Opus, GPT-5.5, and Gemini Pro. The company pitches it as the cheaper default for running agents at scale.

• Etched reaches $5B valuation with $1B in booked AI chip orders Nvidia competitor Etched says it has $1 billion under contract for inference systems built on its chip. The startup now carries a $5 billion valuation.

• Amazon starts a $1 billion forward-deployed engineering org Amazon launched a team of engineers who embed inside customer companies to build and deploy purpose-built agents. The structure copies recent moves by OpenAI and Anthropic, prioritizing fast deployments and customer self-sufficiency.

• Base44 builds its own model to cut dependence on frontier labs Wix-owned vibe-coding platform Base44 began rolling out an in-house AI model. The company aims to eventually beat frontier models and reduce reliance on external providers for defensibility.

• Google releases Nano Banana 2 Lite for faster, cheaper image generation Google updated its image generator with a lighter variant that runs faster at lower cost. The change targets creators producing high volumes of AI images.

• Researcher finds Claude Code hiding markers in date strings sent to non-Anthropic APIs A developer inspecting Claude Code 2.1.196 found code that alters the apostrophe and date separator in the system prompt's date string. The triggers include the ANTHROPIC_BASE_URL override, an Asia/Shanghai or Asia/Urumqi timezone, and hostnames matching base64-encoded domain and AI-lab keyword lists. The visible text reads as a normal date while the raw request carries the marker.

• X launches a hosted MCP server for its API X released a managed Model Context Protocol server that lets developers connect AI applications to its platform API. The hosted setup removes the need to build custom integrations.

• OpenClaw releases mobile apps for Android and iOS The free, open-source agentic program shipped on both mobile platforms. Users can now run the agent directly from a phone.

• Ex-DeepMind poker AI team reaches $500M valuation trading for hedge funds Three former DeepMind researchers built EquiLibre Technologies, a Prague lab now valued above $500 million. The team applies its game-theory and poker AI work to quant trading strategies.

• OKX builds a payments and identity marketplace for AI agents Crypto exchange OKX wants AI agents to hire and pay each other directly. Its system combines payments, identity, and reputation into one marketplace.

• Report finds heavy AI adopters grew headcount, including entry-level roles A new study shows companies classified as high-intensity AI adopters increased total headcount by 10.2%. Entry-level headcount at those firms rose 12%, cutting against claims that AI eliminates junior jobs.

Read the full edition →