AI news

Released: BU Ultrafast is live for everyone⚡ Fastest and cheapest web agent on the planet. > tasks in under 3s > median task takes 14.4s > $0.126 total for 25 tasks in the video Real time, not sped up. Try it on our cloud today 🔥

We’ve released AA-Omniscience v1.1, updating our knowledge and reasoning benchmark with a refreshed public set, removal of saturated questions, and ambiguity fixes AA-Omniscience is our proprietary benchmark measuring model knowledge and hallucinations, including whether models admit when they don’t know an answer. Our primary metric, Net Score, rewards correct answers, penalizes incorrect answers, and treats abstentions neutrally. AA-Omniscience was updated to v1.1 as part of the Artificial Analysis Intelligence Index v5 release. Key changes in AA-Omniscience v1.1: ➤ Updated 3,000-question dataset: Down from 6,000 questions, with previously public and saturated questions removed, errors and ambiguities corrected, and new questions added ➤ Newly calibrated 300-question public set: AA-Omniscience v1.1 has a new public set on @huggingface calibrated to match the difficulty of the private set. At release, public and private scores have a 0.99 Pearson correlation and 0.99 Spearman rank correlation for AA-Omniscience Net Score, with a median difference of 0.0 points. Accuracy and hallucination rate also have 0.99 correlations, with median differences of -0.07 and -0.29 percentage points, respectively ➤ Slightly updated domain distribution: Questions span the same 6 domains and 42 topics as version 1, with a revised distribution: Science, Engineering and Mathematics (600), Software Engineering (600), Finance (450), Health (450), Humanities and Social Sciences (450), and Law (450) Key results for AA-Omniscience v1.1: ➤ Rank order largely preserved: AA-Omniscience v1.1 has a 0.98 Spearman rank correlation with v1. The recently released Gemini 4 Carbon (high) leads with a Net Score of 54, followed by Claude Opus 5.5 (48, max with fallback) and Claude Fable 5.1 (46, max with fallback). Muse Spark Open (xhigh) is the leading open weights model, 28 points behind Gemini 4 Carbon ➤ Scores average ~2 points lower on the refreshed set: Models score 1.9 points lower on Net Score on average compared to v1, with a median decrease of 0.6 points. Some models see much larger declines, particularly MiniMax M3.1 Flash, down 22 points (16 to -6), and MiMo V2.6 Pro, down 13 points (8 to -5) ➤ Slightly lower accuracy: AA-Omniscience Accuracy measures the percentage of questions the model answers correctly. Models answer a median of 1.5 percentage points fewer questions correctly on average compared to v1. Claude Fable 5.1 (max with fallback) has the highest accuracy at 68%, followed by Claude Opus 5.5 (max with fallback) at 66% ➤ Hallucination rates decrease: AA-Hallucination measures how often models choose to answer when they do not know the answer. The median Hallucination Rate falls 1.4 percentage points, with Gemini 4 Argon (high) recording the lowest hallucination rate at 4%

I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.

1/ Excited to announce our new company Particle This is our first step in a series of models focused on developers. With particle-1, we focused on token efficiency and managed to keep parity in performance while using ~45% fewer tokens. Live now: particle.ai

Starting today, eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in API credits to build apps and agents.

By popular request, you can now text your dot. Give it a task to take off your plate or something to keep an eye on, using text, photos, or even voice notes. It will text you proactively with relevant updates. Safe, secure texting via iMessage or RCS. You’re the only one who can text your dot, and your dot can only text you. Rolling out now as a preview for Pro users in the US.

Muse Spark Open from @AIatMeta on ARC-AGI (Verified): - ARC-AGI-3: 0.4%, $3.0K (provider adapter harness), 0.1%, $3.9K (standard harness) - ARC-AGI-2: 57.5%, $0.73/task - ARC-AGI-1: 89.0%, $0.29/task Its scores on v1/2 are similar to Haiku 5.5 & GPT-6 Luna, but ~9x the cost.

together gpu clusters can now take a node from gpu fault to healthy again automatically health checks identify the failure, the platform chooses the lightest repair, and you decide what can run unattended, down to individual xid codes here’s how it works

Cartesia Managed Agents scores an Elo of 1115 on the Artificial Analysis Speech Agent Arena, 98 points above Cartesia's earlier voice agent product, Cartesia Line, with Task Success Rate up from 77.5% to 91.7% Managed Agents is @CartesiaAI's fully managed platform for building and deploying voice agents. It is a cascaded system: Cartesia's own Ink Speech to Text and Sonic Text to Speech models run around an LLM of the developer's choice. We evaluated Cartesia Managed Agents with GPT-5.5 as the LLM and the Skylar and Daniel voices. Key takeaways: ➤ Speech Agent Arena: Cartesia Managed Agents scores an Elo of 1115 (95% CI 1086-1144, 685 samples), 98 points above Cartesia Line at 1017 (95% CI 985-1049) ➤ Agentic tasks: On tasks that need tool calls, Cartesia Managed Agents scores an Elo of 1112 ➤ Task Success Rate: 91.7% of eligible tool-calling conversations end with the correct final task-completing tool call Congrats to the @CartesiaAI team! See below for more details ⬇️

GIF

Much requested, and now yours to connect: iCloud Mail and Calendar plugins are available in ChatGPT. Ask ChatGPT to find that email, catch up on your inbox, check what’s next, or update a calendar event. Phew. Available on all ChatGPT plans on web, desktop, and iOS.

Claude Code is now multiplayer. Run `/delta share` in Claude Code, and your teammate joins in their own session with your full context and uncommitted work. You're both working on the same thread, backed by DeltaDB. delta.dev/blog/claude-code-mult…

Today we're coming out of stealth, introducing Neywa Labs 🚀 We built the infrastructure to build and deploy AI that runs on the devices people already own. Private, offline, unlimited. One mission: make AI available to everyone in the world. neywalabs.ai

Judgment Labs is now GA 👀 It’s a “continuous-improvement stack” for AI agents. Agent swarms can go through production traces to spot failures, triage them, and test fixes. The Judgment MCP works with Claude, Codex, Cursor, or any MCP client. Curious to see how much time it saves when you run agents at scale.

Introducing Supabase Compute, in private alpha. You don’t need a second platform for your agents and long-running jobs. Run them in your Supabase project, next to your data in Postgres. See how it works → supabase.com/blog/supabase-comp…

JetBrains IDEs are now powered by libghostty! This brings huge feature and performance improvements to JB terminals! Based on their own annual reports, that's over 12.5 million active users across all platforms about to get a libghostty-powered terminal. blog.jetbrains.com/platform/202…

RAG is so back! Another paper showing that plain RAG keeps winning the agent memory comparison. They show that many dedicated memory systems score below naive retrieval over the raw history on LongMemEval, PersonaMem, and MINT-Eval, while costing much more. They split a memory

Great paper from Alibaba and Cornell on verification in coding-agent harnesses. (bookmark it) If you build agent harnesses, this one is worth your time. They find that a separate reviewer agent adds 5.9 to 9.5 points on Terminal-Bench at 1.13 to 1.25x the baseline cost. They

Meet Prism, a specialized 8B model trained to route coding tasks to the right model in real time. Prism chooses fast cost-efficient models for the easy stuff and bigger, smarter models for the hard stuff. We're seeing upwards 45% lower inference costs. Now live on Charm Hyper.

we've extended the same pattern as our tui plugins to support this so basically one package can ship server, tui, and desktop plugins that right panel is a great place to put all kinds of useful things x.com/lukeparkerdev/…

Faster streaming speech recognition on @Arm edge devices: see how ExecuTorch, SME2, FP16 and QC8W balance accuracy and latency—delivering up to 2.12× faster processing on Pixel 11. Read our latest blog: pytorch.org/blog/optimizin… ✍ Josiah Davis, Kshitij Sisodia

We made Unsloth UD3 MiniMax-M3.1-Flash quants! We tested across multiple domains to ensure our quants work great for them as well! By default, MiniMax-M3.1-Flash uses NVFP4 for MoE layers and BF16 / MXFP8 for the rest. UD-IQ1_M retains ~89.9% top-1%. Works in Unsloth Desktop! x.com/UnslothAI/stat…

You can now use AsyncGRPO with tree packing on your favorite hybrid models such as @Alibaba_Qwen 3.5 (GDN), @liquidai LFM 2.5, and @NVIDIAAI Nemotron 3.5 (Mamba2). Built as a standalone component that hooks into @huggingface Transformers and leverages Hub kernels. x.com/DirhousssiAmin…

Introducing *HackBench* 🤡💀⏱️📊 ⏲️ 💪🔧 🔥 - Detecting & localizing reward hacking in autoresearch agents - Frontier models fail - Dataset on HuggingFace - Recipe to collect natural & synthetic hacks Blog post: facebookresearch.github.io/RAM/blogs/hack… 🧵1/5

Kimi K3.1 Pro is now the best open-weight model at training other AI models. But it’s still far behind Opus 5.5 at every budget. New FrontierML results, our benchmark for end-to-end ML work ⬇️

Your Agent can generate videos now?? 🎬 Yep. Video generation just landed in LobeHub. Tell your Agent what you want to see, and it'll make it happen. Text to video. Multiple models. All in one workspace. No switching apps. No extra steps. Go give it a spin: lobehub.com

Introducing Jev + Firecrawl company data Pull company records from Alexandria and let Jev classify them. We sorted 146 startups by who buys from them in 2.2 seconds. Perfect for lead scoring, account research & market maps at lightning speed! firecrawl.dev/alexandria/cat…

New in Cline Desktop v0.0.49 🧵 1) Install and manage local models through Ollama or LM Studio. Cline checks your hardware and recommends the best local models available. Try via Settings → Provider → Local

Bun AOT also introduces `--trim` which deletes unused parts of Bun from the compiled binary On macOS, a binary that uses Bun.serve’s HTTP server goes from 89 MB to 25 MB

GPT-6.1 Terra is also coming to the API and Codex. Use Terra to build apps and agents that interpret images, navigate software, and take action. You get fast computer use at a lower cost, ideal for high-volume workflows. x.com/4398626122/sta…