Pulse (last 3d) · Research 86 · Agents 79 · Products 75 · Models 38 · Enterprise 25 · Media 22 · Releases 21 · Industry 19 · Open Source 18 · Chips 12

Trending · OpenAI 21 · Claude 14 · ChatGPT 11 · Claude Code 10 · Hugging Face 9 · Codex 8 · MCP 8 · Anthropic 7 · Nvidia 6 · Apple 4 · Cursor 4 · Jensen Huang 4


Four labs shipped near-frontier models at mid-tier prices on the same day that Google and the FBI documented autonomous agents breaching networks within hours. Raw model capability is now cheap, and what separates deployments is isolation, attestation and audit, while the compute bill behind the cheap tokens is now public in Anthropic's filings. Re-run model routing on cost this quarter, and require hardware attestation and sandboxing before any clinical workload goes to a new endpoint. blog.google | forbes.com

  • Gemini 4 Neon, a mid-tier model, is being compared directly with Opus 5.5 and is live for Pro and Ultra subscribers. blog.google
  • Tencent's Hy4 flash costs $0.07 per million input tokens, with 12B active parameters and 1M context. producthunt.com
  • Google recovered an autonomous exploit framework and expects it to be in wider use within months. blog.google
  • Anthropic priced at $83 a share, about $1.5T, and starts trading on Nasdaq as ANT today. anthropic.com
  • Isomorphic Labs raised a $5B Series C, 2.4x its May round. isomorphiclabs.com
  • An LLM audited black-box clinical prediction models using queries alone, tested blind against planted flaws. arxiv.org
  • Confidential Cloud encrypts data in GPU memory and returns an attestation report for every instance. producthunt.com
  • Genie 4 - DeepMind's world model is now available to eligible AI Ultra subscribers. blog.google
  • Managing frontier AI risk - OpenAI is pacing releases, held back GPT-6.1 Astra and uses safety cases. openai.com
  • Commerce chip licenses - Advanced AI chips in data centers across five countries now require licenses. reuters.com
  • Harvard data center study - 403 US hyperscale sites emit 25.7M tons of CO2 and draw 1.8% of US power. hsph.harvard.edu
  • AMD closes World Labs - The all-stock deal makes Fei-Fei Li AMD's EVP and chief scientist. ir.amd.com
  • NATO Daybreak - NATO's comms agency runs up to 300B OpenAI tokens a month. helpnetsecurity.com
  • Claude Platform OAuth - Third-party agent platforms connect to a Claude org with no API keys, billed to that org. x.com
  • Inside the Codex sandbox - OpenAI documents the isolation design behind Codex. openai.com
  • Claude Docs chaos loop - 60 server errors across 2B launch requests, hardened through chaos testing. claude.dev
  • bitwarden/agent-access - A Bitwarden repository for giving agents access to credentials. github.com
  • cowtree - Copy-on-write workspaces let parallel coding agents run without file collisions; open source. producthunt.com
  • Microsoft IQ - Connects organizational knowledge, activity and web data as context for agents. producthunt.com
  • S.Space - A shared workspace and memory for cross-vendor agents, from the AutoGen team. producthunt.com
  • worldinmotion - End-to-end encrypted messaging between agent sessions across machines and accounts. producthunt.com
  • Beacon - An MCP extension that drives your signed-in Chrome without exporting cookies. producthunt.com
  • LangChain Managed Deep Agents - A hosted version of the Deep Agents framework. x.com
  • uRun Sense - Streaming model that makes turn-taking and interruption calls for voice and video agents. producthunt.com
  • AssemblyAI-STT-3.7 - Three speech-to-text models for live, agent and recorded audio, covering 32 languages. producthunt.com
  • Voyager 1 - TTS reporting 1.53% WER on Seed-TTS against 2.13% for human recordings (vendor benchmark). producthunt.com
  • Nodeway - Maps a repo so AI-written PRs are reviewed by what they reach; 22,497 files in 1m42s. producthunt.com
  • Polylane for CI - Finds the commit that broke CI and opens a tested fix PR. producthunt.com
  • UI Verify - Shows a coding agent every screen its change touched so it can fix UI regressions. producthunt.com
  • Waldo by Wherobots - A geospatial data agent from the Apache Sedona creators that outputs scheduled pipelines. producthunt.com
  • canfit - A Rust CLI that checks whether a local model fits in memory right now. producthunt.com
  • Local Dictation - Fully on-device dictation for iPhone, iPad and Mac in over 90 languages. producthunt.com
  • EasyCut - An MIT-licensed transcript-based video editor with on-device Whisper. producthunt.com
  • Altly - A Chrome extension that writes missing alt text for screen readers in 18 languages. producthunt.com
  • SID-2 - A swarm-based search model claiming to beat Opus and Astra at 20x lower cost (vendor benchmark). x.com
  • DaisyRose-27B - A company post-train of Qwen 27B, part of a wave of in-house open-weight models. radixark.com
  • Frontier Design Benchmark - Contra Labs offers it as an alternative to crowd-voted design rankings. x.com
  • LLM-assisted reverse engineering - Gemini reverse-engineered a CCTV camera's cloud plugin in about two hours. slowchronicle.com
  • Previous-token LLM inversion - Reconstructs prompts from model outputs, a leakage path for sensitive inputs. arxiv.org
  • PIPETTE - Pairs prompt-injection attacks with defenses to harden agents. huggingface.co
  • Kaer - Security-aware replanning for multi-agent systems. arxiv.org
  • General progress substitutes for specific data - Broad model gains can stand in for targeted downstream training data. arxiv.org
  • Human behaviour foundation model - A foundation model proposed to measure and predict human behaviour. arxiv.org | huggingface.co
  • On-Policy Morphing - Converts pretrained models into hybrid attention architectures using on-policy training. arxiv.org
  • Late Entry in Looped LMs - Studies when depth-recurrent looping should start. huggingface.co
  • Tideway - Phase-aware tiered memory for running MoE inference on consumer devices. huggingface.co
  • Omni-LLM token compression - Compresses tokens while keeping information from earlier multimodal content. huggingface.co
  • Cap-and-Track - Dispatches inference requests across capacity that shifts over time. arxiv.org
  • Evidence-based credit assignment - RL credit assignment for search-augmented LMs. huggingface.co
  • Executable program judges - Programs replace LLM judges for cheaper relevance labeling. arxiv.org
  • Evaluating user simulators - Methods for validating the simulators used to test interactive agents. arxiv.org
  • Budget-paired counterfactual audits - More credible evaluation of visual agents when observation is limited. arxiv.org
  • NexSight - Verifiable tasks for agentic long-video understanding. huggingface.co
  • WorldRuler - An interactive world model that follows configurable rules. arxiv.org | huggingface.co
  • Agentic Harness Engineering Beta - Evolves agent harnesses around stated requirements. arxiv.org
  • Ground, Iterate, Trace - Breaks AI research ideation into three stages. arxiv.org
  • Atomic MathReasoner - Composes atomic reasoning skills per instance to train math LLMs. arxiv.org
  • Mapping the AI supply chain - Charts dependencies and market concentration across the AI ecosystem. arxiv.org

Each major lab's newest release is a cheaper tier that is being judged against the previous flagship.

Google's mid-tier Gemini 4 Neon is being benchmarked against Opus 5.5 Neon is the first Gemini 4 release, and Hassabis pitches it on fast task completion. It is live for AI Pro and Ultra subscribers on mobile, desktop and web. Early hands-on comparisons put it level with Opus 5.5 on long-running agent sessions (single source), so the mid tier, not the flagship, is now the default to evaluate. blog.google | x.com | x.com | micjm.com

OpenAI and Microsoft both cut prices at the top of the market GPT-6.2 Astra ranks as the #2 frontier model and costs 20% to 600% less than its predecessors, with mixed early reviews, and 6.2 Sol looks set to be the volume model. At Ignite, Microsoft launched seven models across price-performance points, and commenters place the top one at Opus 5 class (single source). latent.space | microsoft.ai | theneuron.ai

Chinese labs compete on price while keeping weights closed Qwen4-Max left preview with 2.4T parameters, native vision and 1M context, and its weights stay closed. Tencent's Hy4 flash (224B total, 12B active) costs $0.07 per million input tokens and $0.28 per million output and decodes above 150 tokens per second (vendor benchmark), with open weights promised but not shipped. A review of more than 300 Chinese filings found that distilling US model outputs is standard practice. qwen.ai | producthunt.com | producthunt.com | yahoo.com

Exploits now take hours, so patch windows and trust boundaries are being redrawn around that speed.

Autonomous intrusion has moved from demo to incident report Google's threat intelligence group recovered a working autonomous "exploit_agent" framework and expects it to be in wider use within months. Separately, the FBI says the open-source Artex tool breached networks in under three hours. Any agent with network or tool access is now an attack surface on the same timescale as the attackers. blog.google | techcrunch.com

Anthropic cuts its disclosure window for high and critical bugs to 30 days The new timeline starts January 1 and covers high and critical vulnerabilities Anthropic reports to vendors. Remediation SLAs for third-party components in the AI stack will need to meet a 30-day clock, not the industry's usual 90. anthropic.com

Data stays encrypted in GPU memory, with attestation for every instance Confidential Cloud encrypts data at rest, in transit, in CPU memory and in GPU memory on NVIDIA B300s, so the operator cannot read it. Each instance returns an attestation report to verify before sending data (press release). For PHI inference on rented GPUs this means a check that can be verified, where before there was only a contract clause; the service is in public beta. producthunt.com

The money behind cheap tokens is now visible, and it is pooling in a few places.

Anthropic priced at $83 and starts trading today Anthropic sold 980M Class A shares at $83, valuing it at about $1.5T, and lists on Nasdaq as ANT today. Index providers have already published notices for its first trading day. anthropic.com | x.com | finance.yahoo.com | research.ftserussell.com

The filings show $900.9B of compute commitments against an $81B run rate The commitments are about 11 times annualized revenue (single source). The filing names Amazon, Google and Microsoft as investors, customers, distribution partners and competitors all at once. Long-term pricing from any single vendor now depends on whether that gap closes. forbes.com | x.com

Isomorphic raised $5B, 2.4x its May round, six months later Greenoaks and Sequoia co-led the Series C. AI-first drug discovery is now raising at frontier-lab scale, which raises the price of partnerships and talent for every precision-medicine company competing for the same structural-biology people. isomorphiclabs.com

Vendors report big gains, while independent benchmarks with held-out evaluations still show no strong results.

OpenAI's research-completion rate doubled, but FrontierML finds no strong results OpenAI says coding agents now complete about 46% of 2 to 4 hour research tasks on their own, up from 23% in March (vendor benchmark). On FrontierML's 25 tasks, agents get 12 hours on H100s with the objective hidden, and no frontier model reached the strong-result bar. Abundant Labs' new FrontierRSI benchmark targets similar open-ended work. openai.com | x.com | x.com | x.com

Claims of applied wins arrive without human baselines OpenAI says its models advanced the design of a chip named Jalapeño but gives no human baseline. KPMG says AI agents asked better questions than its accountants in an audit trial (press release). Both are assertions of capability that a regulator would not accept as validation. openai.com | openai.com | europeanbusinessmagazine.com

An LLM audited clinical prediction models through queries alone An LLM probed black-box clinical models without seeing their internals and was scored blind on whether it found planted flaws (preprint, unreplicated). If it holds up, outside teams could audit vendor risk scores they cannot inspect, which is today's model-validation gap for clinical AI. arxiv.org

17 items were dropped: snippets with no content, commentary with no claim, studies that reported no finding, and papers or posts outside the AI field.