Google ships Gemini 3.6 Flash, 3.5 Flash-Lite, and a restricted Flash Cyber
Google released two new fast-tier Gemini models, both usable today. Gemini 3.6 Flash ($1.50/$7.50 per 1M tokens) succeeds 3.5 Flash for coding and knowledge work, posting big agentic jumps — DeepSWE 49% vs 37%, MLE Bench 63.9% vs 49.7%, OSWorld-Verified 83.0% vs 78.4% — while emitting 17% fewer output tokens; it is live in the Gemini API, AI Studio, Android Studio and the Gemini app. Gemini 3.5 Flash-Lite ($0.30/$2.50) targets high-throughput serving at 350 output tokens/second and is rolling into Google Search. The third model, Gemini 3.5 Flash Cyber, is a vulnerability-detection-and-patching specialist gated to a government/trusted-partner pilot over dual-use concerns. Google also confirmed Gemini 3.5 Pro is with testing partners and Gemini 4 pre-training is underway.
Sources & depth
Judge approves $1.5B Anthropic settlement over pirated training books
District Judge Araceli Martínez-Olguín gave final approval to the $1.5 billion class-action settlement in Bartz v. Anthropic, the suit brought in 2024 over pirated books used to train Claude. Authors and publishers receive roughly $3,000 per book; about 91% of the more-than-482,000 covered works have been claimed. Plaintiff counsel calls it the largest known copyright recovery in history, and it is the first major settlement among the dozens of AI copyright cases still moving through US courts — a reference price the rest of the docket will now negotiate against.
Sources & depth
Fireworks: Kimi K3 near-parity with Claude Fable 5 — and routing between them beats either
Fireworks AI published a ~1,030-task evaluation (460 real-repo bug fixes, 89 long terminal-agent runs, 100 algorithmic problems, 225 cross-language implementations, 120 lawyer-graded legal tasks) putting the open-weights Kimi K3 at 92.4% on its SWE set vs Claude Fable 5's 92.6%. The strengths split — K3 wins terminal work and symbolic math, Fable wins web and visualization — and an "oracle router" picking the cheapest correct model per task selects K3 72–96% of the time, which Fireworks translates into headline claims of 93% accuracy and up to ~50× better cost-efficiency than Fable alone on extended agentic loops. Read with the obvious caveat: Fireworks sells K3 inference, and oracle routing assumes you already know which model will be right.
Sources & depth
Jack Dorsey's Block launches Buzz — a workspace where AI agents are members, not bots
Block released Buzz, an open-source group-chat platform that folds team chat, Git hosting and workflow automation into one system and treats AI agents as first-class workspace members. Identity runs on Nostr keypairs: every human and every agent signs its actions, and an agent's signature chains back to its human owner, giving a verifiable custody trail for agent-submitted patches, reviews and workflow runs. It's positioned as a Slack-plus-GitHub replacement (Block uses it internally), ships as a free desktop app for macOS, Windows and Linux, is self-hostable and model-agnostic — and at v0.4.21 is unambiguously early software.
Sources & depth
Qwen-Image-3.0 arrives — without weights or benchmarks
Alibaba's Qwen team announced Qwen-Image-3.0, the third generation of its image model, pitched on "rich content, authentic details, deep knowledge" — image generation practical enough to work as a tool, not a demo. The notable part is what's missing: Qwen-Image 1.0 shipped Apache-2.0 open weights in August 2025, but 3.0 launches with no weights, no license and no benchmark table — only the vendor's own output reel (the Qwen blog page renders empty to non-JS readers; details per Unite.AI's write-up). For the most-watched open-weights image lineage, that's a real strategy shift worth confirming against the weights page in the coming days.
Sources & depth
Nikkei: five US tech giants carry $1.65T of AI debt off their balance sheets
A Nikkei analysis finds Alphabet, Microsoft, Amazon, Meta and Oracle hold $1.65 trillion in off-balance-sheet obligations tied to AI build-out — more than the $1.35 trillion they report, and roughly eight times the hidden figure of four years ago. The mechanism is special-purpose joint ventures: Meta's Hyperion data center with Blue Owl carries $27B that Meta argues isn't its debt; Meta's total off-book obligations near $420B, Oracle's have grown ~30× with $260B in future lease commitments. Commentators note the structures are legal but rhyme with Enron's — the sector plans over $3 trillion of AI data-center spend through 2028 largely financed this way.
Sources & depth