AI News — Wednesday, July 22, 2026

$ cat top-story.md

▸ Google Ships Gemini 3.6 Flash and Two Cheaper Variants — Still No 3.5 Pro

Google released Gemini 3.6 Flash (~17% more token-efficient, output pricing cut from $9 to $7.50/M tokens), a cheaper 3.5 Flash-Lite tier, and a restricted 3.5 Flash Cyber model for vulnerability research — but Gemini 3.5 Pro remains stuck behind internal quality bars, while Google's product lead confirmed Gemini 4 pretraining has begun. Why it matters: Flash is the tier you'd actually route cheap agentic or RAG workloads through if you ever add a hosted fallback next to your local Gemma stack. // TechCrunch / MarkTechPost

▸ xAI's 2-Trillion-Parameter Grok 4.6 Finishes Training "Next Week," Aimed Squarely at Kimi K3

Musk says Grok 4.6 (2T params, up from Grok 4.5's 1.5T) is entering final training, explicitly framed as xAI's answer to Moonshot's Kimi K3 topping the Artificial Analysis leaderboard; public launch is estimated for late August to mid-September. Why it matters: another sign the open/frontier-weight race just accelerated again — worth tracking if you're deciding what to route through or eventually self-host next. // Dataconomy / TechBriefly

▸ Microsoft Reportedly Weighs Swapping Copilot's Backend to Kimi K3 to Save $600M

Beyond Kimi K3's well-known subscription pause, TrendForce reports Microsoft is evaluating replacing OpenAI/Anthropic models in Copilot with Moonshot's open-weight Kimi K3, potentially cutting annual inference costs by roughly $600M now that Copilot bills on usage. Why it matters: if a hyperscaler actually puts a Chinese open-weight model into production, that's a real enterprise-trust signal for anyone evaluating open models for their own stack. // TrendForce

▸ Rumor Mill: Qwen 4.0 Leak Chatter Points to 3D/Spatial Coding Focus

Unconfirmed leak chatter about two Alibaba stealth test models ("Caleb" and "Terrania Alpha") suggests Qwen 4.0 will push into 3D and spatial-design generation, expected around September, following next month's Qwen 3.8 release. Why it matters: pure speculation for now, but a genuinely new capability axis to watch in the open-weight scene if it pans out. // Geeky Gadgets (unconfirmed)

▸ OmniRoute Unifies 270+ AI Providers Behind One Gateway With Token Compression

MIT-licensed AI gateway (23.9k stars) that fronts 500+ models across 271+ providers (90+ free tiers) with automatic fallback and claimed 15-95% token-compression savings; plugs straight into Claude Code, Cursor and Cline. Why it's interesting: a ready-made way to add cheap hosted fallback and redundancy alongside your local vLLM stack without juggling a dozen API keys. // GitHub

▸ showreel Turns Claude Code Into a Screenshot and Demo-GIF Generator

Claude Code plugin that points a self-contained headless Chromium at URLs and CSS selectors to produce annotated screenshots, flow GIFs/MP4s and pixel-verified before/after composites — no browser MCP required, and a bad capture errors out instead of shipping. Why it's interesting: handy for documenting your own dashboards and tools without manual screenshotting. // GitHub

▸ Claude Code v2.1.217 Caps Subagent Concurrency and Fixes a Budget-Limit Bypass

Shipped July 21 with a default 20-subagent concurrency cap (override via CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS), subagents no longer spawning nested subagents by default, a fix for --max-budget-usd not actually halting background subagents once the cap was hit, and a memory leak where truncated MCP tool outputs kept the full untruncated result in memory. Why it matters: directly affects any fan-out or parallel-subagent workflow you run — the new concurrency ceiling can silently throttle heavier multi-agent pipelines unless you raise the env var. // Claude Code changelog

▸ agentcairn Gives Coding Agents Persistent, Ownable Memory via Obsidian + MCP

Stores long-term, cross-project agent memory as plain Markdown in an Obsidian vault you control, backed by a rebuildable DuckDB retrieval cache and exposed to any MCP host (Claude Code, Codex, Antigravity), with a companion Obsidian plugin visualizing the memory graph. Why it matters: a self-hostable, inspectable alternative to opaque agent-memory SaaS that fits naturally next to a personal retrieval setup. // GitHub

▸ llmfit Tells You What Actually Runs on Your GPU Box Before You Download It

Terminal tool (TUI and CLI) that detects your RAM, CPU, GPU/VRAM and backend, then scores hundreds of catalog models on memory fit, speed, quality and context length; v1.1.6 shipped July 21. Why it matters: directly useful for deciding what to load onto your vLLM+Gemma box before committing to a multi-hour download. // GitHub

▸ "Das ist der Deepseek 2.0 Moment" — DeepSeek V4 Goes GA

German-language deep dive on DeepSeek's V4 line reaching general availability — a 1.6T-parameter V4-Pro and a 284B V4-Flash MoE, both with 1M-token context and a strong agentic/coding focus — framed as a repeat of the original 2025 "DeepSeek shock." Why it matters: a serious, cheap, open-weight challenger worth watching if you ever want an alternative to Gemma for heavier reasoning workloads. // The Morpheus Tutorials

▸ The Most Important Conversation in AI Right Now

Matthew Berman on the industry's shift from "tokenmaxxing" (maximizing AI usage) to "valuemaxxing" (maximizing the actual value that usage produces). Why it matters: a useful lens for auditing whether your own heavy Claude Code and agent usage is actually paying off versus just burning tokens. // Matthew Berman

▸ Reverse-Engineering Is Cheap Now

Simon Willison argues coding agents haven't made reverse-engineering undocumented devices and APIs newly possible — they've made it newly cheap, collapsing the effort-to-benefit ratio that used to make such projects not worth starting. Why it matters: a sharp reframe of how agents change the economics of exploratory technical work, not just its speed. // Simon Willison / Hacker News