AI Roundup: Gemini 3.1 Flash-Lite, Donald Knuth Praises Claude, and Qwen Drama Continues

A few things worth knowing about this week in AI. Nothing earth-shattering, but three stories that are actually worth your time.

TL;DR: Google released Gemini 3.1 Flash-Lite, a cheap fast model that beats 2.5 Flash on most benchmarks. Donald Knuth called out Claude Opus 4.6 for solving an open math problem he had been working on for weeks. And the Qwen leadership fallout is still unresolved.

Gemini 3.1 Flash-Lite Is Quietly Good

Google released Gemini 3.1 Flash-Lite on March 3rd. It is available now in Google AI Studio and Vertex AI.

The headline numbers: $0.25 per million input tokens, $1.50 per million output tokens. That makes it one of the cheapest capable models available right now, about one-eighth the price of Gemini 3.1 Pro.

What makes it interesting is the benchmark position. It outperforms 2.5 Flash with 2.5x faster time-to-first-token and a 45% increase in output speed. It also scores 86.9% on GPQA Diamond and 76.8% on MMMU Pro, which puts it above older Gemini flagship models from prior generations.

It comes with four thinking levels built in (minimal, low, medium, high), giving you cost control on a per-request basis. For high-volume tasks like content moderation, translation, or UI generation at scale, this is the kind of model that changes the cost math on a whole category of products.

Available in preview via Google AI Studio and Vertex AI.

Donald Knuth: Claude Solved My Open Problem

This one is worth reading twice. Donald Knuth, who wrote The Art of Computer Programming and is essentially the godfather of algorithmic computer science, posted that Claude Opus 4.6 solved an open problem he had been working on for several weeks.

His exact words: "Shock! Shock! I learned yesterday that an open problem I'd been working on for several weeks had just been solved by Claude Opus 4.6... What a joy it is to learn not only that my conjecture has a nice solution but also to celebrate this dramatic advance in automatic deduction and creative problem solving."

Knuth has historically been skeptical of AI hype, which makes this statement land differently than the usual benchmark announcements. He is not the kind of person who says things like this casually. The problem was related to Claude's Cycles, a mathematical structure explored in a paper he co-authored.

This is the kind of story that matters more than leaderboard scores.

Qwen Drama: Still Unresolved

Covered this yesterday but it is worth a follow-up note. As of March 5th, there is no confirmation that Junyang Lin or the other key Qwen team members have returned. Alibaba's CEO held an emergency all-hands, Lin posted a vague reassurance on WeChat, and that is where things stand.

The Qwen 3.5 models are still out there and they are still exceptional. The 27B and 35B are getting strong community feedback for coding tasks. But the team that built them is scattered, and the roadmap for what comes next is unclear.

If you are evaluating open-weight models for production use right now, Qwen 3.5 is still worth considering. Just keep an eye on how this situation resolves before committing to it as a long-term dependency.

Sources: Google Blog, Simon Willison's Weblog (simonwillison.net), Hacker News