☕️ Google launches 3 new Gemini Flash models

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🤖 Google launches 3 new Gemini Flash models

🎥 Claude learns workflows by watching screens

💻 Poolside model beats bigger rivals

💬 Block launches Buzz for AI agents

📦 Rspack 2.0 ships with ESM core

Plus: 🎁 13 other news you might like, 🧰 6 tools, and 📚 5 papers.

🤖 Google launches 3 new Gemini Flash models LINK

Google released three additions to the Gemini Flash family (3.6 Flash, 3.5 Flash-Lite, and the cybersecurity-focused 3.5 Flash Cyber) while its anticipated flagship, Gemini 3.5 Pro, remains in private partner testing.

Gemini 3.6 Flash uses ~17% fewer output tokens than 3.5 Flash, costs $1.50/$7.50 per million input/output tokens, lifts DeepSWE from 37% to 49%, and adds Computer Use as a built-in client-side tool in the Gemini API.

Flash Cyber, built into DeepMind's CodeMender agent, scored 83.2% on CyberGym and found 55 unique flaws scanning V8 commits versus 36 for Claude Opus 4.6; access stays restricted to governments and trusted partners.

🎥 Claude learns workflows by watching screens LINK

Claude Code for desktop can now build, run, and iterate on iOS apps directly inside Apple's iOS Simulator, watching the app live as it executes, with the integration shipping in public beta.

The simulator pane drives the emulator directly rather than through computer use, so Claude works without the macOS Accessibility and Screen Recording permissions that screen-control normally requires.

The feature runs on Claude Code Desktop for macOS in local sessions only and needs Xcode with the iOS platform installed; screenshots go to Anthropic under standard retention, so avoid signing into real accounts.

💻 Poolside model beats bigger rivals LINK

Poolside has released Laguna, a 118 billion parameter open-weight coding model that outperforms rival American and Chinese open models on public benchmarks, trailing only Moonshot's newly launched Kimi K3.

Laguna is roughly 20 times smaller than Kimi K3's 2.8 trillion parameters, and Poolside built an "industrial" model-building pipeline meant to ship new, more capable iterations every five weeks.

Weights are available publicly since April, having first gone to government and corporate clients, though Warner concedes Laguna still lags the top closed models from OpenAI and Anthropic.

💬 Block launches Buzz for AI agents LINK

Block released Buzz, a free, open-source Slack-style workspace built on the Nostr protocol that treats AI agents as full participants alongside humans in channels, threads, direct messages, voice, media sharing, and code repositories.

Each agent receives its own Nostr keypair plus a second signature tying it to a human owner, creating a cryptographic paper trail that lets workspaces enforce rules like admitting only member-owned agents into private channels.

Buzz connects agents through Zed's Agent Client Protocol and works with Claude Code, Codex, or goose; teams can run their own Nostr relay or use Block's free hosted beta, which currently has no usage limits.

📦 Rspack 2.0 ships with ESM core LINK

Rspack 2.0, ByteDance's Rust-based bundler, has shipped with a pure ESM core, dropping its CommonJS build while adding experimental React Server Components support, better static analysis for smaller bundles, and simplified configuration.

Build performance runs roughly 10% faster than 1.7 and up to 100% faster than 1.0; a 10,000-component React production build with persistent cache fell from 5.6 seconds to 1.4 seconds, and HMR reached 118 milliseconds.

Dependencies collapsed from 192 to just 1, shrinking install size from 15 MB to 1.4 MB; upgrading now requires Node.js 20.19+ or 22.12+, dropping Node.js 18, with tree shaking improved by default in production.

Other news you might like

• Prefactoring: Clear the Way for Your New FeatureLINK

• How DevRel saves thousands of tokens per Claude Code runLINK

• Do we need state management libraries anymore?LINK

• Kubernetes Agent Sandbox: What It Is and How to Deploy It with PulumiLINK

• Error Handling vs. Hiding ErrorsLINK

• Android Studio Quail 2 Redesigns Agent Mode, Streamlines AI-Assisted CodingLINK

• Pushing 4.7M Commits an HourLINK

• Stop adding more GPUs: Weka's new storage platform reduces load by caching 100% of an AI model's pre-calculated tokensLINK

• Temporal takes on the chaos behind enterprise AI agentsLINK

• R2 is now Generally AvailableLINK

• Speakeasy service tracks enterprise-wide AI agent spendingLINK

• Cisco Launches Low-Cost AI Models for Source Code SecurityLINK

🧰 Trending tools

CartAI: completes checkout on any live merchant site without integration, using one API for catalog search, order tracking, PCI-compliant payments, and commission-based monetization.LINK

ditto.site: converts any public URL into clean, componentized Next.js or Vite code in minutes, preserving components, tokens, and interactions, with a free REST API and MCP server.LINK

Second Brain for AI v2: a self-hosted memory layer that lets AI tools like Claude and ChatGPT recall context via semantic search instead of starting fresh each session.LINK

QApilot's CoWork: transforms existing test cases into automated mobile tests with AI planning, human approval checkpoints, and real-device execution across iOS, Android, and Flutter.LINK

Persona.js: a framework-free chat UI library that embeds AI assistants with streaming, voice, and tool execution into any website without backend changes.LINK

Computable: a marketplace for buying, selling, and reselling GPU compute by the week, with transparent pricing and futures-style contracts for future capacity.LINK

📚 Trending papers & reports

Entity binding drift in AI agents that resolve a customer or record once and reuse it across steps, shows that "locking" that choice backfires, multiplying wrong actions up to 8.5x, while a cheap second-model recheck cuts errors 79%.LINK

AI runtime governance treats chatbot outputs as raw sensor readings needing verification before acting, cutting compute costs by about 31%, boosting retention of high-value data by up to 24%, and running with zero variance across 100 repeated tests.LINK

Multi-step business workflows can pause for human approval, retry failed steps automatically, and resume exactly where they left off, using a structured playbook that makes routing and audit trails visible product features instead of hidden prompt logic.LINK

Living code documentation auto-builds a queryable knowledge map from pull requests, letting teams ask questions about their codebase and get accurate, well-grounded answers instead of hunting through scattered files.LINK

Reusable AI instructions now qualify as standalone software with their own identity and lifecycle, letting teams update, roll back, or retire a chatbot's specialized skills independently, backed by analysis of 138,133 records across 20,556 repositories.LINK

See you tomorrow for a new dose of ☕️ Devshot!