AI News — September 12, 2026

$ cat top-story.md

▸ Sakana Ships Fugu Max and Fugu Ultra v2 — Orchestrators That Now Snub Claude and GPT-6

Sakana's "route tasks across a model pool" family gets a cost tier (Fugu Max, $2/$6 per M tokens, built on open-weight and Nvidia Nemotron models) and a capability tier (Fugu Ultra v2), both deliberately excluding Claude and GPT-6 Astra from their routing pool to cut vendor lock-in — yet still claiming to beat both on several benchmarks. Why it matters: A live test of whether an orchestrator built only on open/independent models can match frontier labs, useful if you want a fallback that isn't hostage to one vendor. // MarkTechPost

▸ Mistral Raises Europe's Largest-Ever Tech Round: €3B at a €21B Valuation

Samsung led a €3B Series D for Mistral, nearly doubling its valuation to €21B; the company says it will spend the money building and owning its own data centers rather than renting compute. Why it matters: More capital for the main non-US, non-Chinese open-weight lab means a better-funded alternative supply of models you can actually self-host. // TechCrunch

▸ Grok 4.7 Misses Its Fourth Promised Release Date

Musk's September 12 target for the 2.1-trillion-parameter Grok 4.7 came and went; xAI now blames reinforcement-learning issues (the model reportedly "gives up early" on hard tasks) and has given no new date, after missing three earlier estimates since July. Why it matters: If you were waiting to benchmark Grok 4.7 against Claude, GPT-6 Astra or Fugu, keep waiting. // OrcaRouter

▸ OmniRoute Puts 352 AI Providers Behind One Free, Self-Hosted Endpoint

MIT-licensed gateway (65k stars) that gives Claude Code, Cursor, Cline and 30+ other coding tools one OpenAI-compatible endpoint with automatic fallback across 352 providers (150+ free), plus token compression it claims saves 15–95%. Why it matters: A natural, self-hostable companion to a vLLM/OpenWebUI stack when you want cloud models as fallback without vendor lock-in. // GitHub

▸ Pi Bundles a Unified LLM API, Agent Loop and Coding CLI Into One Toolkit

An open agent harness (104k stars) combining a multi-provider LLM API, a tool-calling agent runtime, a terminal-UI library and an interactive coding-agent CLI; it has no built-in permission system, so the docs recommend Docker or its Gondolin sandbox for isolation. Why it matters: A lighter, hackable base if you want to build custom agent tooling on top of your own self-hosted models. // GitHub

▸ Claude Code 2.1.269: Plugin Eval Command, Output-Style Switching, Bash Diffs

Adds claude plugin eval for scored, reproducible plugin test runs, /output-style to switch styles anywhere including Remote Control, a live diff view in the Bash tool, and an OpenTelemetry tag for repository attribution, plus prompt-cache and terminal-rendering fixes. Why it matters: Plugin eval finally gives skill/plugin authors a repeatable way to test changes before shipping them. // Claude Code changelog

▸ obra/superpowers Turns Claude Code Into a Test-Driven, Self-Reviewing Engineer

A composable skills library (285k stars) that forces coding agents through brainstorming, red-green-refactor TDD, a structured four-phase debugging process and subagent code review before a task counts as done; works across Claude Code, Cursor, Devin, Copilot and more. Why it matters: If your agents keep declaring victory on broken code, this is the most-starred fix the community has converged on right now. // GitHub

▸ vLLM Squeezes 3–4x More Throughput Out of MiniMax M3 on AMD MI355X

A methodical pass through tensor-parallel sharding, batching and data-movement fixes lifted MXFP8 output throughput from 109 to 342 tokens/s/GPU (3.14x) at concurrency 32, with MXFP4 configs reaching up to 943 tokens/s/GPU. Why it matters: Concrete, reproducible tuning steps for anyone running MoE models like MiniMax or GLM on AMD instead of Nvidia hardware. // vLLM Blog

▸ Mistral's "Agentic Search" Replaces One-Shot RAG With a Multi-Step Read/Navigate Loop

Instead of answering from one batch of retrieved chunks, the model gets five tools — search, open, navigate, read, grep — to iteratively dig through documents; on FinanceBench that took Mistral Medium 3.5 from 26.7% to 86% accuracy, and it ships as portable tooling for on-premises deployment too. Why it matters: A concrete, benchmarked pattern for upgrading a classic Qdrant-style RAG pipeline into agentic retrieval without being locked to Mistral's cloud. // Mistral AI

▸ Research: Cross-Document Chunking Beats Semantic Chunking by 12% on Multi-Hop Questions

"Cross-Document Topic-Aligned" (CDTA) chunking builds chunk boundaries around topics shared across a whole document collection instead of splitting each document in isolation, reaching 0.93 faithfulness vs. 0.78 for semantic chunking on HotpotQA multi-hop questions. Why it matters: A concrete chunking strategy to try if your RAG pipeline struggles with questions spanning multiple documents. // arXiv

▸ Mistral AI Raises €3 Billion With Samsung Leading the Round

Bloomberg breaks down what Samsung's lead investment and the €21B valuation mean for Europe's push toward sovereign AI compute. Why it matters: Quick context on the biggest non-US AI funding story of the week. // Bloomberg

▸ Boris Cherny: Claude-Written Code Needs More Guardrails, Not Fewer

Anthropic's own production practices for Claude Code lean on heavy lint rules, extensive tests and Claude-driven end-to-end tests specifically because unreviewed agent-written code degrades a codebase faster than human-written code. Why it matters: A reminder, from the team building Claude Code itself, that agent output needs more automated scaffolding, not less. // Simon Willison

▸ So You Want to Use OpenRouter? The Pitfalls Nobody Mentions

Running an assistant handling 18M+ messages through OpenRouter revealed identical model weights scoring 90% on one provider vs. 75% on another, silent image-input failures, ignored reasoning-effort parameters, and fallback chains that can all fail simultaneously. Why it matters: Essential reading before treating any multi-provider gateway — OpenRouter, OmniRoute, or your own — as a drop-in reliability layer. // via Simon Willison