DeepSeek's V4.1-Flash Beats Claude Opus 5 On Agent Benchmarks At A Fraction Of The Price!

Today's AI news: DeepSeek released V4.1-Flash, the smallest model in its new causal encoder-decoder family, with native vision, a one-million-token context, MIT-licensed weights and scores ahead of Claude Opus 5 on DeepSWE and CyberGym at a fraction of the price. OpenAI put GPT-Live-1, its full-duplex voice model, into the API at five cents a minute. OpenAI paused new $200 ChatGPT Pro subscriptions under unprecedented Astra demand. Suno launched v6, turning images, videos and voice memos into songs. And Pamir announced Lapis One, an $880 Linux mini PC built as a dedicated machine for AI agents. Chapters: 0:00 Today's AI News 0:30 DeepSeek V4.1 Flash 1:27 GPT-Live-1 API 1:58 OpenAI Pro Pause 2:40 Suno v6 3:25 Lapis One Blend Roundup 2026-09-10 https://x.com/deepseek_ai/status/2097930608790167907 https://x.com/OpenAIDevs/status/2098099269551149398 https://x.com/thsottiaux/status/2098113585683808624 https://x.com/suno/status/2097846245540888664 https://x.com/pamir_ai/status/2097502085269131407 DeepSeek's V4.1 Flash beats Claude Opus 5 on two agent benchmarks at a fraction of the price; OpenAI puts GPT-Live-1 in the API but Astra demand forces them to pause new two-hundred-dollar Pro subscriptions; Suno v6 turns photos, videos and voice memos into songs; And Pamir's Lapis One gives your agent a Linux box of its own. Here's today's AI news. DeepSeek has released V4.1 Flash, the smallest model in its new causal encoder-decoder architecture family. It has five hundred and fifty two billion parameters, but only 8 billion activate per token during prefill and 16 billion during decode, with a one-million-token context and native image understanding built in rather than bolted on. The weights are MIT licensed on Hugging Face. On DeepSeek's own chart it scores 74.2 on DeepSWE, just ahead of Claude Opus 5 at 74.0, and 88.1 on CyberGym against 84.5 for both Opus 5 and GPT 5.6 Sol. Opus still wins Terminal-Bench, 43.3 to 30.0. Off-peak API pricing lands at fifteen cents per million uncached input tokens and sixty cents per million output. DeepSeek says it beats its own larger V4-Pro on several benchmarks, which raises the question of what Pro is now for. OpenAI has put GPT-Live-1, the full-duplex voice model behind ChatGPT's voice mode, into the API. It listens and speaks at the same time, so it can handle interruptions and the little acknowledgements as they happen rather than waiting for a turn. The voice layer is a single model, and deeper reasoning and tool calls are delegated to whichever backend model and harness you choose. Pricing is five cents a minute for the voice layer plus whatever your reasoning model costs. OpenAI has paused new subscriptions to its two-hundred-dollar ChatGPT Pro plan. Tibo Sottiaux, who runs Codex and ChatGPT, said Pro accounts put the most strain on the system and this was the smallest step that keeps the broadest access open. Existing Pro subscribers are unaffected, every other plan stays on sale, and the API is untouched. The day before he had called Astra demand unprecedented, steeper than anything OpenAI had seen through years of very steep growth. So the company that was paying people to wait for Astra last week is now turning away its highest-paying customers. Capacity, not demand, is the constraint. Suno has launched v6, a new generation of music models built in partnership with Warner Music Group, BMG and Believe. It comes in three tiers: v6 for predictable results, v6-wild for experimental output, both for Pro and Premier subscribers, and a faster v6-mini available to everyone. The headline feature is multimodal input: turn an image, a video clip or a voice memo into a finished song. Editing is more surgical too, with changes to a specific section, vocal, arrangement or lyric without regenerating the rest, plus mashups that lift vocals or drums from one song into another. Every previous Suno model is being retired. Pamir has announced Lapis One, a small Linux computer designed as a dedicated machine for your AI agents. It is an 85-millimetre square weighing 227 grams, running a Rockchip RK3576 with a 6 TOPS neural unit, 8 gigabytes of memory and 64 gigabytes of storage, plus a monochrome memory LCD and a battery. Lapis OS is Debian-based with a read-only verified core, root access, apt, Nix and a Python SDK. Plug it into another computer over USB-C and it emulates a keyboard and mouse while capturing that screen, so the agent can drive your main PC. It is eight hundred and eighty dollars.