Google's cheap Gemini 4 tops rival on one test

Hi there, this is your daily ☕️ AIpresso.

Google's cheap Gemini 4 tops rival on one test LINK

• Google released Gemini 4 Hydrogen, a fast, cheap general-purpose model that edges out the larger frontier Argon on the Artificial Analysis Intelligence Index, scoring 51 at its default "high" tier versus Argon's 50.
• Hydrogen adds a new "xhigh" thinking level reaching 54, carries a 1M-token input and output window, and comes in cheaper than Argon while outscoring it on the intelligence index.
• It's live now across the Gemini API, Antigravity, the Gemini app, and Search, with 2.5x faster default-level tasks and a jump to 57% on Terminal-Bench 4.1, though it lags Claude Sonnet 5.5 and GPT-6.1 Sol on agentic terminal work.

OpenAI launches AI pen-testing agents LINK

• OpenAI unveiled two Daybreak cybersecurity products, Red Team and Defense Factory, that deploy AI agents to pen-test web applications for vulnerabilities and draft fixes, announced at its Cyber at the Frontier event yesterday.
• Red Team uses multiple agents reading source code and app context to confirm attack paths from code scanners, bug bounties, and disclosure reports, while Defense Factory traces flaws and prepares patches, ingesting findings from Wiz, Snyk, and Endor Labs.
• Defense Factory scans connected GitHub repos, monitors commits, and lets engineers modify proposed patches with Codex and open draft pull requests, though patches never deploy automatically and Red Team is currently request-only early access.

Runway open-sources its robotics AI model LINK

• Runway released Praxis-1, its 27B-parameter world action model for robotics, under an Apache 2.0 license that permits commercial use, with weights plus inference and fine-tuning code now available to every developer and researcher.
• The mixture-of-experts diffusion transformer predicts video and actions jointly and tops NVIDIA's RoboLab-120 leaderboard at 46.2% success across 1,200 episodes, ahead of the next-best FLUX 3 Action and every open-weight VLA.
• Pretrained on large-scale video, it adapts to new embodiments with light fine-tuning or LoRA and hit 80.0% on SimplerEnv using Google's RT-1 robot it never trained on, running on hardware from H100 to Jetson Thor.

Anthropic lets you self-host Claude agents LINK

• Anthropic released self-hosted Claude Managed Agents on November 12, 2026, letting customers run the full service and its sandboxes in their own cloud account or on-premises data center, with agents, sessions and histories stored on that infrastructure.
• The customer-run server exposes the same Managed Agents API as the hosted service, so the Claude SDK and CLI connect unchanged, while model calls route by default to the Claude API or optionally to Amazon Bedrock, Google Cloud, Microsoft Foundry or a self-run gateway.
• Persistent deployments need PostgreSQL 16+, each serves one tenant at one trust level requiring separate deployments otherwise, and self-hosting alone doesn't keep traffic internal since every call sends the full conversation to the selected inference provider.

Strategies & Tactics

> Don't Write Kernels, Derive Them: Derive every kernel from a small tensor algebra instead of hand-writing one per operator and chip, so the compiler finds the fastest layout automatically, though the derived versions still sometimes trail hand-tuned ones.

> Fast by Default: Compile-Time Heuristics for Helion Kernels: Helion now reads a kernel's structure at compile time to instantly pick fast settings, matching most of autotuning's speed without hours of GPU search.

> Integrating SGLang into Spyre in Under a Week with AI: Reusing existing PyTorch hardware code and AI coding agents let a two-person team connect IBM's Spyre chip to a faster serving framework in under a week.

> Local vector search for coding agents: From 507 MB to a 7-line loop: Pre-compute document vectors in the cloud so a coding agent's local search runs in a tiny offline loop, cutting size and latency while improving accuracy.

> Build once, run on any agent: A practical guide to agent skills: Package repeatable task instructions as reusable skill files that work across AI coding agents, keeping project settings separate so teams reuse the same judgment without rewriting it.

Other news you might like

• Jensen Huang’s exit puts Nvidia’s support for open source in question LINK

• Perplexity Releases Decider V1.3 Models with Text-Span Extraction LINK

• Unemployment Labs Reports Blackwell Patterns in Chinese AI Models LINK

• AWS AgentCore Runtime Adds Larger Sessions and Filesystem Mounts LINK

• SpaceX Unveils Two AI Supercomputing Campus Plans in Texas and Arizona LINK

• Meta's Muse can now make calls on your behalf (if you have a paid subscription) LINK

• What's new in Grok Bot 1.0 LINK

• Escaping SPACE: Part II LINK

• LLMs can learn (some) multi-step reasoning on filler tokens LINK

Trending tools

Trainn MCP: turns screen recordings into product demos, guides, and walkthroughs, then delivers them via a branded no-code LMS or in-app tutorials. LINK

Inkling 2 Small: a 594B-parameter Mixture-of-Experts model with 17B active, supporting image inputs and variable effort for agentic coding, vision, and reasoning tasks. LINK

Gemini 4 Hydrogen: multimodal AI model that understands and combines text, images, audio, video and code, letting developers build apps across diverse data types. LINK

Higgsfield Miracle: edits video clips in real time using plain-language descriptions, applying precise changes without affecting the rest of the shot. LINK

qorqbot: a provider-agnostic agent that runs on your existing Claude or Codex subscriptions, handling long-running business tasks in the background while learning your workflows automatically. LINK

Aven App Agent: embeds an in-app AI assistant that answers user questions and walks them through tasks by highlighting developer-registered pages, elements, and workflows. LINK

Trending papers & reports

Supervised fine-tuning can match or beat reinforcement learning at making AI models reason and transfer to new problems, if you feed it the right mix and enough training examples, overturning a popular assumption. LINK

AI training incentives reveal that how you reward a model during reinforcement learning decides whether it ends up honest or learns to bluff, with a precise payoff threshold predicting which outcome you get. LINK

AI leaderboard rankings can be gamed by providers using fake votes and cherry-picked prompts, lifting a model up to 45 places, meaning the popularity charts buyers trust may not reflect real quality. LINK

Coding agents often nail the fix on their first code edit yet keep working, burning more than half their steps afterward and occasionally undoing the correct solution, so stopping sooner could cut compute costs. LINK

Compressed language models can be shrunk to run on cheaper hardware while keeping accuracy, by taming the rare extreme signals that normally wreck low-cost versions, using a lightweight drop-in fix. LINK

See you tomorrow for a new dose of ☕️ AIpresso!