Live Feed
Claude Haiku 5.5 Beats GPT-6 Luna on Computer Use — at a Fraction of the Cost
Anthropic's new small model scores 39.2% on Terminal-Bench 4.0 versus GPT-6 Luna's 16.4%, and 72.4% on OSWorld 2.1 versus Luna's 48.9%. Haiku 4.5 scored 0.0% on Terminal-Bench. Price is 75% lower on average.
FrontierSWE v2 October Snapshot: GPT-6 Astra Leads 20 Models at 65.5%
BenchLM's October 6 snapshot of FrontierSWE v2 places GPT-6 Astra at 65.5%, Claude Opus 5.5 at 62.3%, and Claude Sonnet 5.5 at 61.9% on a 34-task long-horizon engineering benchmark designed to resist saturation.
CoreWeave Enters India with 240 MW Navi Mumbai Campus, Vera Rubin Online by Mid-2028
CoreWeave signed with AdaniConneX to build three 80 MW data halls at Taloja, Navi Mumbai, deploying NVIDIA Vera Rubin Platform with an option to double capacity to 480 MW at the same site.
Study: Claude and ChatGPT Recommend Higher Prices to Wealthy Users
New research finds AI shopping assistants systematically upsell users who signal wealth. Claude Opus 4.8 recommended flights averaging $198 more and health insurance $284 more per month for wealthy users than for low-income ones.
OpenAI Sweeps GDP.pdf's Top Seven Slots — The Benchmark That Gives No Partial Credit
Artificial Analysis has deployed Surge AI's GDP.pdf benchmark across 205 model configurations. GPT-6 Astra and GPT-6.1 Sol hold all seven positions above 31%. The best score is 32.2%. In GDP.pdf, partial credit does not exist.
Google's Gemini Nano Banana 2.1 Enters Arena Image Leaderboards — Thinking-Capable with C2PA Credentials
Google added Gemini Nano Banana 2.1 to Arena's Text-to-Image and Image Edit leaderboards on October 6. The model supports image generation from video input, multi-turn editing, virtual try-on, and an unusual feature for an image model: native thinking capability. Content credentials via C2PA are built in.
OpenAI Publishes 722 Math Manuscripts from Unreleased Internal Model — Lean Proofs, 3 Hours Pro Compute Each
OpenAI's internal frontier model generated 722 mathematical manuscripts organized into 372 result families, each costing roughly three hours of ChatGPT Pro compute. Computer-checkable Lean proofs accompany many results. An independent verifier confirmed the zeta function manuscript; OpenAI says it is working toward releasing the model.
UniEvo-VL Lets Multimodal Models Teach Themselves Using Their Own Critiques
A new on-policy self-distillation framework eliminates the need for a larger teacher model by splitting a single VLM into teacher and student roles at inference time. The paper hit Hacker News front page as the AI training automation race accelerates.
Mistral Large 4: 1.05T-Param Open MoE in Preview at $0.68/$2.09, Weights October 27
Mistral AI launched a public preview of Mistral Large 4 on October 6 — a natively multimodal 1.05T-parameter MoE with 49B active params, 1M context, and open weights due October 27. DeepSWE v1.1 at 62%, Finance Agent v2 at 54.7 outperforms GPT-6 Astra on that task.
Reka Rho-1: One 19B Model Collapses Text, Video, and Robot Control Into a Single Network
Reka releases a research preview of Rho-1, a 19B omni-reasoning model that handles text, image, video generation, and robot action tokens in one shared context window with no tool calls or second model. Trained on 320 H100s over three months.
Trump's Secret AI Framework Draws Senate Pressure as New Task Force Gets 120 Days to Sort It Out
Democratic senators Warren and Blumenthal are demanding answers on the White House's still-unreleased AI regulatory framework, months after Trump's executive order. Meanwhile, Trump named intelligence director Jay Clayton as AI czar and created a Super Intelligence task force with a 120-day mandate.
UK Accepts All 44 NHS AI Commission Recommendations, Shifts to Lifecycle Regulation
The UK government accepted every recommendation from its National Commission into the Regulation of AI in Healthcare on October 6, replacing point-of-approval device assessment with continuous lifecycle monitoring. The MHRA is simultaneously opening Phase 3 of its AI Airlock regulatory sandbox.
Claude Opus 5.5 Agents Find Two Room-Temperature Magnetic Semiconductor Candidates for Spintronic Memory
A team of Claude Opus 5.5 agents designed a new compound, YBaMnFeO5, and rediscovered KV[Cr(CN)6], first synthesised in 1999 but never identified as a spintronic candidate, as Luttinger-compensated semiconductors. Both are predicted to sort electrons by spin without a net magnetic field. One held magnetic order at 376K in simulation.
Qualcomm Licenses Huawei's LogicFolding Patents in Cross-License Covering AI, 5G, and Optics
Qualcomm and Huawei have signed a multiyear cross-license covering Huawei's LogicFolding 3D chipmaking technique, 5G devices, AI services, and near- and co-packaged optics. Qualcomm is the first major US chip company to license LogicFolding patents. Financial terms were not disclosed.
Reflection AI Launches Beam: 501B-Param Open-Weight MoE With 80.9% SWE-bench and 4-Week GB300 RL Run
Reflection AI's first open-weight model is a 501B sparse MoE with 23B active params, trained on 10.5K NVIDIA GB300 GPUs for four weeks and over 100 million RL rollouts. It hits 80.9% on SWE-bench Verified while using 3-4x less inference compute than 2T+ parameter rivals.