Google's Nano Banana 2.1 ships 4K images at half the price
Google quietly turned its popular image model into a cheaper, sharper product this week — while the receipts show the update is real and the pricing math cuts both ways. Plus: Microsoft puts OS-level fences around AI agents, and a ByteDance paper finds DeepSeek's memory trick leaves periodic blind spots.
Google released Nano Banana 2.1, and the API bill for image generation just got cut roughly in half. The new model — available as gemini-nano-banana-2.1 in the Gemini app, AI Studio, and the Gemini API — replaces Nano Banana 2, adds direct 4K output, and improves mask-based editing, subject consistency, and text rendering enough that Chinese-language designers are calling the typo and layout problems largely gone. Google's own pricing page confirms the headline numbers: image output at 1K and 2K tiers roughly halved (about $0.067 to $0.034 and $0.101 to $0.050 per image), 4K down about 25%. The counterpoint the launch coverage skips: input tokens tripled from $0.50 to $1.50 per million, and text-output pricing jumped as well — so prompt-heavy workflows may not save anything. On LMArena the model sits in the top six for text-to-image with roughly a 60-point average gain over its predecessor across image and editing boards, which is a real jump even if the exact rank drifts daily. Our take: this is Google doing what it does best right now — competing on price and iteration speed rather than spectacle — and the missing piece is that English-language coverage remains thin; the sharpest user reports are coming from Chinese design communities, which is where the model's biggest practical improvements are being documented.
Microsoft's Execution Containers hit general availability — the first OS-level containment layer built for AI agents. Announced on the Windows Developer Blog October 7, MXC lets developers declare exactly which files, network destinations, and UI an agent's workload may touch, then enforces that boundary with the platform's own containers — process sandboxes on Windows, macOS, and Linux, a separate Windows session for long-running agents, WSL for Linux toolchains, and an experimental microVM tier. The crucial design choice: the policy lives outside the agent, so generated code cannot grant itself access. GitHub Copilot, OpenClaw, OpenAI Codex, Replit, LM Studio, and Unsloth already ship with support; Claude Code, Manus, Perplexity, and Hermes Agent are on the announced list, and NVIDIA has folded its OpenShell controls into the same framework. After months of agent incidents where permission systems were the weak link, an OS-enforced boundary that vendors actually implement is a more meaningful security step than another round of policy promises.
A ByteDance Seed paper says DeepSeek's own memory optimization is what makes its answers flip. In "Periodic Weak Spots," researchers show that DeepSeek V4's chunked KV-cache compression — the trick that makes long contexts affordable — assigns tokens fixed positions inside each chunk, and tokens that land in weak slots get near-ignored. Their 128K needle test found accuracy gaps of up to 40 percentage points between token positions in DeepSeek-V4-Flash-Base and nearly 35 points in V4-Pro-Base; post-training narrows the gap but does not close it, and DeepSeek's newer V4.1-Flash, which halves the compression stride, cuts it to about 6 points. A simple demo makes it tangible: pad a code-completion prompt with decorative characters and the model's top prediction flips depending on how many you add. The authors reproduced the pattern from scratch in small Qwen models, with the error period exactly matching the compression stride — evidence the flaw is baked into fixed-length chunking, not a tuning mistake. It's a pointed finding because every lab chasing cheap long-context inference uses some version of this technique: the savings are structural, and so are the blind spots.
What to watch: whether Google acknowledges the user-reported watermark accumulation bug that makes repeated edits progressively degrade — right now it's documented only in community tests.
Would you ship an agent inside Microsoft's container boundary — or is OS-level sandboxing still a promise until it survives a real break? Tell us in the comments.
