Pulse (last 3d) · Research 88 · Agents 64 · Products 62 · Models 42 · Open Source 25 · Media 25 · Enterprise 22 · Releases 17 · Industry 15 · Chips 15

Trending · OpenAI 17 · Claude 11 · Codex 10 · Nvidia 10 · Claude Code 9 · ChatGPT 8 · Hugging Face 7 · MCP 6 · Anthropic 5 · Apple 5 · California 3 · Cursor 3


Frontier output got cheaper again today: GPT-6.2 and Microsoft's seven MAI models cut the price of a capable answer, and Cerebras raised throughput at the same speed. Two independent evals show that benchmark score and harness choice do not predict finished, verified work, and most of the agent launches compete on knowing when to stop and hand off to a person. The real constraints are now task-completion evals on your own workloads, calibrated abstention, and the fact that frontier chat models already contain protein-sequence data. openai.com | arxiv.org

  • GPT-6.2 ships as Astra (frontier) and Sol (efficient), with better quality and lower cost than 6.1. openai.com
  • Microsoft launched seven MAI models at Ignite, including MAI Code 2, at lower prices. microsoft.ai
  • Coding models with higher benchmark scores did not get more actual work done. arxiv.org
  • 64% of agent spend on Terminal-Bench went to tasks that failed verification. runta.com
  • Claude models predict amino acids in obscure proteins, which points to protein-sequence training data. xinmingtu.cn
  • Liquid d1.2 returns calibrated probabilities so low-confidence answers can go to human review. producthunt.com
  • Gemini 4 Carbon takes up to an hour on tasks Opus and Astra finish in minutes. kernel.dev
  • Agents API by Recall.ai - one API call sends a talking, on-camera agent into Zoom, Meet, Teams or Webex. producthunt.com
  • BrowserPod - a free Linux microVM in the user's browser for running agent-written code, with no sign-up. producthunt.com
  • OnlineCompiler.io - a sandbox API with no network, hidden tests and signed receipts, in 46 languages. producthunt.com
  • Lantern - an MIT local-first code graph and PR reviewer that also monitors Claude Code and Codex sessions. producthunt.com
  • Faros Desktop - searches sessions across agents, links them to outputs and stretches AI quota. producthunt.com
  • superlog - finds production errors nobody set alerts for and proposes fixes; claims 87% get merged (vendor claim). producthunt.com
  • StarSling - YC P25 self-improving CI for repos with heavy agent-driven PR volume. starsling.dev
  • Omnitome - compresses books 7:1 offline on a 4GB laptop GPU, citing every sentence back to the source (vendor benchmark). producthunt.com
  • deeepSpace - syncs Figma and code design systems, traces drift to commits and exposes both over MCP. producthunt.com
  • Pangular Inspector - inspects Angular components, signals and routes from the page, the CLI or MCP. producthunt.com
  • daegun - a no_std, zero-dependency Rust font parser, shaper, layout engine and rasterizer with a C API. producthunt.com
  • Claude Code Mode - experimental mode where Claude writes scripts that chain tools and keeps responding during long commands. x.com
  • My Voice Harness Is My New Operating System - a voice-driven agent harness became one developer's primary interface. aifoc.us
  • OpenAI model finds 35 Apple vulnerabilities - including flaws that allow iPhone and Mac takeover (single source). x.com
  • Strengthening NATO and allied cyber defenses with AI - OpenAI's cyber-defense work with NATO and allies. openai.com
  • TeleChat to Xing4 - frontier agentic models built on domestic Chinese chips and frameworks. huggingface.co
  • AscendGYM - a cost-aware benchmark for LLM-driven kernel optimization on Huawei Ascend NPUs. huggingface.co
  • Agentic On-Policy Distillation - distillation implicitly reshapes how models follow ReAct patterns. huggingface.co
  • Do LLM Agents Still Need to Be Told How to Self-Improve? - tests whether agents still need explicit self-improvement instructions. arxiv.org | huggingface.co
  • Self-Improvement Memory - a memory mechanism for continual recursive self-improvement. huggingface.co
  • RSI-Prime - training models to train models. arxiv.org
  • Agent Trust Infrastructure (ATI) - federated identity, attestation and accountability for agents that work across organizations. arxiv.org
  • Mind the Gap - a deployed multi-agent system that keeps software documentation current. arxiv.org
  • Formal Verification Is Not Verifiable - argues formal verification cannot itself be verified, which limits trust in AI-written code. alperenkeles.com
  • Glyph - an interpretability engine for AI-driven scientific discovery in clinical data. huggingface.co
  • Breath-Dx - a time-series benchmark for disease screening from electronic-nose breath data. huggingface.co
  • Retrieval-Augmented Correction for Generalist Medical Image Segmentation - retrieval corrects the output of generalist segmentation models. arxiv.org
  • Say Less, Hallucinate Less? - preference optimization shrinks VLM object coverage; the paper restores it. arxiv.org
  • Reading Scene States - VLM safety failures that arise from reading scene states. arxiv.org
  • What Does a Bias Head Measure? - bias attribution at the level of attention heads is non-identifiable. arxiv.org
  • Attribution Through the Decoder - explanations in prediction space for structured outputs. arxiv.org
  • Polarity Composition Hypothesis - causal evidence from negation in NLI. arxiv.org
  • Weight-Space Programming Doesn't Need Transformers - weight-space programming works without transformers. huggingface.co
  • Distortion-Free Semantic Watermarks - watermarks black-box generative models without degrading output. arxiv.org
  • Spool - low-latency streaming inference for hybrid vision-language models. arxiv.org
  • Ambient LLMs - data-efficient supervised fine-tuning by matching temperature. arxiv.org
  • Fine-Tuning Gemma 4 Edge Models - SFT and QLoRA on a laptop iGPU for MATLAB signal-processing code. arxiv.org
  • MoE-LoRA Methods for Multilingual Adaptation - a systematic comparison of MoE-LoRA variants. arxiv.org
  • Structure-Residual Inductive Biases - inductive biases for language model pretraining. arxiv.org
  • Learning Stratified Policies for Inference-Time Objectives - policies trained for objectives applied at inference time. arxiv.org
  • Correcting Flows with Marginal Matching - corrects flow models using marginal matching. arxiv.org
  • Déjà View - a shared working-memory model for visual tasks. arxiv.org
  • π^4 - a foundation model for 4D scene geometry and motion. huggingface.co
  • FastFrames - video generation at interactive latency. huggingface.co
  • MPEG-Diffusion - mixed-resolution video generation. huggingface.co
  • TextSphere - efficient text-to-image generation through hyperspherical alignment. arxiv.org
  • TCP4D - unified tool-center-point tracking and generation for robot manipulation. huggingface.co
  • Heave-Ho - decentralized long-horizon collaboration among multiple humanoids. huggingface.co
  • ResWAM - world-action models that learn only what changes. huggingface.co
  • Capability Composition in Heterogeneous Teams - composes team capabilities with barrier-guided diffusion sampling. arxiv.org
  • WoRD - diffusion-based posterior inference for online goal recognition. arxiv.org
  • StringQFT - a textbook resource for AI-assisted theoretical physics. arxiv.org
  • Where Should Physics Enter? - identifiability and calibration in probabilistic sea surface temperature forecasting. arxiv.org
  • Multimodal Masked Diffusion Autoencoder for Modeling Impaired Driving - a diffusion autoencoder that models impaired driving. arxiv.org
  • A Neural Envelope Method - fast gradient-based design of optimal takeoff trajectories. arxiv.org
  • Google's AI answers reducing search traffic by 40% - AI answers cut click-through to publishers by about 40%. revscope.co
  • Perplexity hits a $1B revenue run rate - up 4x in 2026, after a $20B valuation round. latent.space | theinformation.com
  • Cloudflare Labels Itself a Neolab - Cloudflare repositions itself as an AI lab. latent.space
  • Satya Nadella says he's AGI-pilled now - Nadella now expects AGI in the near term. theverge.com
  • Google's Spark becomes Tasks - Google renamed its Spark agent Tasks and made it free. theverge.com
  • Nvidia and Jensen Huang - Q3 met expectations amid China chip-sales criticism, a DOJ Megaspeed probe and Huang's reported ouster (single source). theverge.com | thezvi.substack.com | theverge.com | x.com
  • Anthropic lobbying claim - says Anthropic paid a firm tied to Rubio's former chief of staff $740K (single source). x.com
  • ChatGPT Ads - now runs in 44 countries. openai.com
  • Introducing ChatGPT for Sales - a vertical ChatGPT product for sales teams. openai.com
  • Introducing OpenAI for Chile - an OpenAI country initiative for Chile. openai.com
  • Not even Sam Altman is safe from AI deepfakes - deepfakes of Altman are circulating on Instagram. theverge.com
  • Tesla Cybercab factory - a €3B Giga Berlin plant with capacity for up to 1 million vehicles a year. electrek.co
  • Introducing Spatial CSS - a WebKit developer preview of spatial CSS. webkit.org

Three vendors cut the price of frontier output this week, so how long a task takes is now the clearest difference between them.

GPT-6.2 splits into Astra for frontier work and Sol for cheap volume OpenAI shipped GPT-6.2 as Astra (frontier) and Sol (efficient), with better results on fewer output tokens and sharply lower cost than 6.1 in early third-party evals (single source). A mathematician who tested Astra in agentic CLI mode for two weeks calls it very strong on computation-heavy math, and OpenAI claims it leads on legal evals (vendor benchmark). openai.com | openai.com | x.com | openai.com

Microsoft's seven MAI models give Azure a second first-party frontier option Microsoft AI launched seven models at Ignite, including MAI Code 2 and Code 2 Flash, pitched as frontier performance at a fraction of current prices and built for 24/7 operation (press release). Early Copilot users report Opus-level coding (single source), which matters to teams whose clinical data contracts already sit on Azure. microsoft.ai | x.com | x.com

Cerebras and a Gemini latency teardown both make speed the axis that matters Cerebras runs prefill and decode on separate hardware for trillion-parameter models and reports about 5x throughput at the same token speed (vendor benchmark). An independent eval finds Gemini 4 Carbon taking up to an hour on tasks Opus and Astra finish in minutes, so price per token no longer captures what an agentic run costs. cerebras.ai | kernel.dev

Three evaluations measured completed, verified work instead of leaderboard points, and leaderboard rank did not predict the outcome.

Higher benchmark scores did not produce more finished coding work An arXiv study finds coding models with higher benchmark scores did not get more actual work done (preprint, unreplicated). FrontierHarness Eval v1.1 compared harnesses for each model on Terminal-Bench and found 64% of spend went to tasks that failed verification, so the harness and a verification gate affect cost more than swapping models. arxiv.org | runta.com

Agents controlling real lab instruments top out at 36.7% FrontierEE Bench has frontier models operate real lab instruments and hardware on electrical engineering tasks, and the best model scored 36.7% (single source). It is the closest public proxy for agents running wet-lab or sequencing equipment, and it shows that unattended instrument control is not ready. x.com

Judges that read whole trajectories beat chunked LLM judges on accuracy and cost Judgment Labs' Agent-as-a-Judge evaluates agent trajectories longer than 1M tokens and reports +0.21 F1 at 62% lower cost than chunked LLM judges (vendor benchmark). Chunked judges miss errors that span several steps, which is the failure that matters most in long agent runs over patient records. x.com

Most of today's launches differentiate on knowing when to stop and ask, not on raw capability.

Liquid d1.2 returns calibrated probabilities so low-confidence calls go to humans Liquid AI's omnimodal decision models d1.2 and d1.2-ultra add audio input and five tasks, including numeric estimates with a range and multi-select. d1.2-ultra scores 75 on the Artificial Analysis Decision Index, two points above the previous best (vendor benchmark). Calibrated confidence is the primitive that clinical abstraction review queues have been missing. producthunt.com

infaiw lets the parent agent, not the model, decide what "done" means The Apache-2.0 Go harness makes sidecar agents pass an acceptance script before they report completion. It adds durable sessions you can branch, approvals that survive a closed terminal, and switching between DeepSeek, Codex and Claude Code mid-session. An acceptance gate is exactly the control FrontierHarness found missing when most spend went to failed tasks. producthunt.com

Voice and vision agents now ship with abstention and audit logs built in Dialect turns a camera into a typed REST and MCP endpoint that says when it is unsure, at $99 a month per monitored thing. Pinecall is an Apache-2.0 voice platform built on LiveKit, with simulated callers, judged call logs and a human one tap away. Both treat "I don't know" and replayable logs as features, which is the posture regulated deployments need. producthunt.com | producthunt.com

The human-approval queue is becoming its own app category Waggle shows each agent question with options and a recommendation over MCP on the Mac. Hako answers agents from the Lock Screen or Watch with end-to-end encrypted notifications, and polyroom puts agents and a team in one room with recorded decisions. Three launches on one day means the approval step now gets its own product. producthunt.com | producthunt.com | producthunt.com

Protein data inside frontier LLMs, and coding agents running the analysis, both change what a result on biomedical data proves.

Claude models predict next amino acids in obscure proteins Researchers show Claude Fable 5.2 and Opus 5.5 predict the next amino acid in obscure protein sequences, which points to training on protein-sequence data at scale (preprint, unreplicated). Any benchmark that compares a general LLM with a protein model on public sequences may be contaminated, and questions about the provenance of biological training data now apply to frontier chat models. xinmingtu.cn

Statistical inference has to change when coding agents run the analysis A new paper examines how statistical inference changes when coding agents carry out the data analysis (preprint, unreplicated). When an agent explores cohort data, every analysis it tries and discards becomes part of what a reported p-value has to account for. arxiv.org

5 items: no discernible content, or no AI relevance.