Qwen3.8-Omni-Flash: omnimodal, 1M context, 90% cheaper than Gemini
Alibaba's API-only omni model tops Gemini on audio at a fraction of the cost — plus Gemini Live hits GA, Claude Code syncs your skills, and Jev skips text.

Copy markdown
Omnimodal in, cheap tokens out
Qwen3.8-Omni-Flash takes text, image, audio and video across a 1M-token window and answers at roughly $0.15 per million input tokens and $0.47 per million output — about 90% and 95% below Gemini 3.8 Flash, with audio-input pricing cut over 98%. Alibaba says it beats Gemini 3.8 Flash on overall audio and gains about 25% across 29 benchmarks versus the prior Omni-Plus. It's API-only for now, with no open weights.
The model it's undercutting just shipped
Google made Gemini 3.8 Live and Live Extended Thinking generally available this week — native audio-to-audio models for real-time voice agents, with interleaved reasoning and asynchronous function calling. If you were building voice on the Live API, you now have a same-week price war: Qwen's omni model claims stronger audio at roughly a tenth of the token cost.
Your claude.ai skills now follow you into the terminal
Claude Code 2.1.275 syncs the skills and plugins you've enabled on your claude.ai account down into terminal sessions (opt out with syncClaudeAiSkills: false), and adds a send-now key (Ctrl+Enter) to fire queued messages mid-turn. The very next build, 2.1.276, patched a regression that 400'd every request routed through a proxy or gateway — update now if you run a custom ANTHROPIC_BASE_URL.
A model that decides instead of writing
TypeSafe's Jev is a "System One" model that emits a typed decision — a class, a route, a struct — instead of generating text, which the company says makes hallucination structurally impossible and runs 40–200x faster than a frontier LLM at about $0.042 per million input tokens. At 3B parameters it targets edge and real-time classification, where you'd otherwise burn a big model on a yes/no.
Let an agent drive your real, logged-in browser
Tencent open-sourced BrowserSkill, a CLI-plus-extension that lets any shell-capable agent (Claude Code, Codex, Cursor) act inside your actual logged-in browser through a separate Agent Window — so it can work on sites you're authenticated to without hijacking the tab in front of you. It was one of the week's fastest-climbing repos on GitHub.