Vibeleaderboard
Index — Latest Intelligence

Intel

Page 03

Introducing Together Storage Accelerator: fast, free data caching for your GPU cluster

Together AI launches Storage Accelerator, a distributed caching layer for its GPU clusters that avoids repeated reads from object storage and Hugging Face Hub.

Why it mattersTeams training or serving on Together clusters can cut model download time and egress fees by caching checkpoints near the GPUs, and avoid Hugging Face rate limits stalling jobs.

articlewww.together.ai

Stripe evals: Measuring real-world coding agent performance

Stripe describes an eval suite of 69 integration scenarios run on 14 model and harness pairs.

Why it mattersCoding agents favor older, more common APIs and miss newer features, which pass-fail benchmarks do not detect. Stripe's roleplay-driven evals and plugin results show how to measure and steer integration quality.

articleKiran Prasad

Give your agents controlled access to analyze your data and move funds

Stripe extends Treasury for agents from read-only to write access over MCP.

Why it mattersAgents connected through Stripe's MCP server can now transfer funds, convert currency and pay vendors within limits you set, with optional human approval. This is a practical pattern for scoping agent authority over money.

articlestripe.com

Tencent Hy4 is live on Together AI

Together AI now serves Tencent's Hy4, the first Hy model with image input, with a 1M-token context, tool calling and JSON output.

Why it mattersHy4 is now callable on serverless and dedicated endpoints at $0.83 per 1M input and $2.50 per 1M output tokens, with a 1M-token window and image input. Engineers can price it against current models.

articlewww.together.ai

llm-gemini 0.38

Simon Willison's llm-gemini 0.38 adds Gemini 4 Neon. His notes untangle the Gemini 4 pricing: Neon at $3/$15, Carbon at $2/$10 until a December 1 doubling, Hydrogen at $1.50/$7.50.

Why it mattersNeon lists at $3 input and $15 output per million tokens, currently above half-price Carbon. Carbon doubles on December 1, so model choice and cost planning change.

articlesimonwillison.net
Cursor@cursor_ai

Cursor adds Gemini 4 Neon, reporting 63.7% on CursorBench 4.1 at default effort versus 57.3% for Gemini 4 Hydrogen, priced at $3/M input and $15/M output tokens. Benchmark figures are Cursor's own, and the linked details were not read.

Why it mattersGemini 4 Neon is now selectable in Cursor at $3/M input and $15/M output tokens. Cursor reports 63.7% on its own CursorBench 4.1, up from 57.3% for Gemini 4 Hydrogen, which helps when choosing a model for cost.

OpenAIDevs@OpenAIDevs

OpenAI added public-beta discounted windows to Responses API background mode. A completion_window of 2h saves 30% per token and 8h saves 40% versus Standard on GPT-6.2 Astra and Sol. Reusing a continuation ID gives follow-up requests higher queue priority.

Why it mattersSetting completion_window to 2h or 8h in Responses API background mode cuts per-token cost by 30% or 40% versus Standard. Batch-style agent jobs that can wait, such as overnight briefs, get cheaper without a separate batch pipeline.

Post-Training Gemini 4 for Real Engineering Work — Melvin Johnson, Google DeepMind

Google DeepMind's Melvin Johnson covers Gemini 4 post-training for real engineering work.

Why it mattersExplains how a frontier lab grades coding agents without rewarding test gaming and trains them to report honestly, which shows what to expect from and how to evaluate agent output.

videoAI Engineer

LLM assisted reverse engineering

A veteran reverse engineer describes how agents with a Ghidra MCP bridge and dedicated skills analyze binaries.

Why it mattersGives a repeatable workflow for pointing a coding agent at binaries, including notes files that survive context compaction and specific ways to unstick the agent.

articleioma8

Gemini 4 Neon Reaches General Availability in GitHub Copilot

GitHub made Gemini 4 Neon generally available in Copilot across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the CLI, and the coding agent.

Why it mattersA new model option for agentic coding is now selectable in Copilot's IDEs, CLI, and coding agent. Eligible plans and admin policy controls are specified, and usage is billed at provider list price, which affects cost planning.

articleGitHub

Launching Trusted Access Program

Reflection AI opens a Trusted Access Program, starting with Laser, an autonomous pentesting harness for its Mirror terminal agent that verifies findings via attack graphs for vetted security teams.

Why it mattersReflection AI's Laser gives vetted security teams a scalable autonomous penetration-testing harness with attack-graph-verified findings, a concrete new access tier for defenders using agentic pentesting.

articleReflection AI editorial sitemap

Astra vs Carbon vs Fable: five-task frontier model comparison

Peter Yang tests GPT 6.2 Astra, Gemini 4 Carbon and Claude Fable 5.2 on a 3D video, a game, a site redesign, blind writing and pushback.

Why it mattersCompares three frontier models on identical build tasks with generation times, plus pricing: Carbon is $2/$10 per million tokens until Dec 1. It also shows which models hold their position under pushback and which give ground.

An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team

Cloudflare's new Workers AI superintelligence team releases an updated Clairic extraction model with PDF and image support, plus Clairic-flash, a 9B variant.

Why it mattersCloudflare released Clairic and the 9B Clairic-flash as open weights, with PDF and image extraction. You can self-host structured extraction or call it on Workers AI.

blogblog.cloudflare.com
Artificial Analysis@ArtificialAnlys

Artificial Analysis benchmarks Google's Gemini 4 Neon: 61 on its Intelligence Index, 3rd overall, 25% lower token prices than Carbon, about 40% faster per task, and a 59 on the Coding Agent Index in Antigravity CLI at lower cost than Claude Code setups.

Why it mattersGemini 4 Neon scores 61 on the Intelligence Index at $6.44 per task and matches Claude Code with Opus 5.5 on coding agents for roughly half the cost. This gives agent builders a concrete cost and speed comparison when picking a model.

OpenRouter@OpenRouter

OpenRouter announced that Google DeepMind's Gemini 4 Neon is live on its platform. The model is described as built for complex software engineering, knowledge work and science, and sits between the Carbon and Hydrogen tiers. The post gives no pricing, context or benchmark data.

Why it mattersGemini 4 Neon can now be called through OpenRouter, so you can test it for software engineering and knowledge-work tasks without a separate Google integration. Pricing and benchmarks are not in the post.

Introducing Gemini 4 Neon

Google DeepMind releases Gemini 4 Neon, a frontier model aimed at software engineering and knowledge work.

Why it mattersGemini 4 Neon ranks #3 on the Artificial Analysis index at $3 per million input tokens and focuses on coding agents and steerability. Engineers choosing a model can compare cost and capability.

articledeepmind.google
GoogleAI@GoogleAI

Google AI launched Gemini 4 Neon, a model aimed at coding and knowledge work with a 1M-token context window and four thinking levels. It rolls out in the Gemini API, AI Studio, Antigravity and the Gemini app. Fabrication-rate claims come from internal testing; the demo video was not transcribed.

Why it mattersGemini 4 Neon is available in the Gemini API paid tier and Antigravity with a 1M-token context and four adjustable thinking levels, so engineers can tune speed against depth on coding work. Fabrication claims are internal Google testing.

GoogleDeepMind@GoogleDeepMind

Google DeepMind announces Gemini 4 Neon is rolling out now. It sits between the Carbon and Hydrogen tiers and targets complex software engineering, knowledge work, and science. The post is brief and gives no benchmarks, pricing, or availability details.

Why it mattersA new Gemini tier targeting complex software engineering is now rolling out. Engineers choosing coding models should evaluate it against their current stack once benchmarks and pricing are available.

Elevenagents For Healthcare

ElevenLabs adds Scribe v3 Live Medical, real-time clinical transcription for ElevenAgents voice agents, with deployments at Banner Health, Cadence and Angle Health.

Why it mattersVoice agents in clinical settings can now use real-time transcription tuned for drug names and clinical terms. Teams get control over what agents say and when calls hand off to a person.

NVIDIA's Nemotron-Labs-Sigma: continuous diffusion LMs for faster decoding

NVIDIA's Nemotron-Labs-Sigma is a 3B/8B family of continuous diffusion language models that decode blocks of tokens in parallel.

Why it mattersNemotron-Labs-Sigma shows continuous diffusion LMs can reach competitive coding accuracy while decoding 4-8 tokens per forward pass, with open 3B/8B weights.

⚡️Anthropic: Know Before You IPO — Tanay Jaipuria, Wing VC

Latent Space and Wing VC's Tanay Jaipuria read Anthropic's S-1.

Why it mattersShows how training, free-tier inference, and distribution channels shape a frontier lab's costs and margins. That informs how practitioners should think about future model pricing and availability.

videoLatent Space

DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

RadixArk releases DaisyRose-27B, an Apache 2.0 post-train of Qwen3.8-27B using RL on 1,200 long-horizon agent tasks, trained in 18 days on 144 H200 GPUs.

Why it mattersAn open Apache 2.0 27B model trained with RL on long-horizon agent sessions reports Terminal-Bench 4.1 pass@1 rising from 5.6% to 12.0%. It shows what a small curated task set and modest GPU budget can do for agent post-training.

articlehuggingface.co

DaisyRose-27B

RadixArk releases DaisyRose-27B, a Qwen-based model post-trained with RL on 1,200 curated long-horizon agent tasks.

Why it mattersShows that a small set of carefully designed long-horizon tasks can lift a 27B model on a hard agent benchmark, with open checkpoints and an RL framework you can reuse.

articlede_sty

ai ai@7.0.177

Vercel's AI SDK ai@7.0.177 adds an interrupt option to steering, letting a steer cancel a model call that has not produced output.

Why it mattersAI SDK 7.0.177 adds an interrupt option so a steer can cancel a model call that has not yet produced output. It also ships the AI SDK agent skill inside the package, so agents get SDK guidance through pnpm.

repogithub-actions[bot]

Mathematical proof, powered by AI. And no hallucinations!

Amazon Science explains connecting AI agents to proof assistants such as Isabelle via MCP, interleaving proof checking with generation.

Why it mattersWiring an agent to a proof assistant over MCP lets the checker reject invalid steps during generation, instead of verifying after the fact. The pattern applies to any agent with a formal verifier.

articlewww.amazon.science

Claude Black Hole Physics

Anthropic describes Claude Fable 5.2 solving two open black hole physics problems.

Why it mattersAnthropic reports Claude solved two open theoretical physics problems about black holes and quantum gravity, with preprints for checking. It is a concrete data point on AI doing original research-level work.

articleAnthropic research
Perplexity@perplexity_ai

Perplexity brought Teammates and Multiagent to its iOS and Android apps and added voice control for Teammates. The company says about a third of its most demanding Computer work already starts on iPhone. Rollout is available now for Max subscribers only.

Why it mattersPerplexity's agent features (Teammates and Multiagent) are now usable from mobile, and Teammates can be driven by voice. Access is limited to Max subscribers, so check your plan before relying on it.

Advancing Open Model Safety Gray Swan

Reflection AI describes its safety partnership with Gray Swan for the open Beam model.

Why it mattersReflection AI partners with Gray Swan to adversarially test its open-weight Beam model, separating pre-release jailbreak resistance testing from deployment-time agentic security like indirect prompt injection defense and reward-hacking checks.

articleReflection AI editorial sitemap
Google Labs@GoogleLabs

Google Labs moves Project Genie onto DeepMind's Genie 4 world model. Users can prompt changes to a world mid-session, build worlds with a conversational agent, and remix others' creations in a gallery. Available to eligible Google AI Ultra subscribers.

Why it mattersGenie 4 lets you edit a generated interactive world while inside it and build worlds through an agent. Eligible Google AI Ultra subscribers can try it now, which shows where world-model tooling stands.

GoogleAI@GoogleAI

Google ships Genie 4 in Project Genie with native audio and real-time world editing, plus a conversational agent, a remix gallery, and Genie 3 upgrades (3-minute sessions, chained worlds). Available to AI Ultra subscribers now.

Why it mattersGenie 4 lets you edit an interactive generated world while inside it, and a new agent orchestrates multimodal models to build worlds. Access is gated to Google AI Ultra, with a Pro waitlist.

Notion CEO on context, Legos versus clay, and agent oversight

Notion CEO Ivan Zhao tells Dan Shipper that agents commoditize UI and workflows, so context is the lasting layer.

Why it mattersArgues that agents make UI and workflow logic cheap, so durable value shifts to a team's data and context. Notion's response is open UI pattern libraries for agents and Kanban boards for orchestrating them.

Software factories and code reviews with Harjot Gill

The Pragmatic Engineer talks with CodeRabbit founder Harjot Gill about how an AI-pilled startup operates, from software factories to AI code review.

Why it mattersA practitioner view from an AI code-review CEO on running an engineering org with heavy agent usage, arguing against cutting AI spend. Useful for teams weighing review workflows and token budgets. Based on show notes only.

articleGergely Orosz

Genie 4 is now available

Google DeepMind ships Genie 4 to eligible AI Ultra subscribers.

Why it mattersGenie 4 lets you edit generated interactive worlds in real time with text prompts and adds native audio. A Project Genie agent builds worlds from conversation. Access is limited to Google AI Ultra, with a waitlist for Pro.

blogblog.google
OpenAI@OpenAI

OpenAI reports its models found up to 14% PPA improvement on AMD flagship GPU IP in under five months, improved its Jalapeño inference chip with gains verified on silicon, and is bringing the capability to chip designers via Synopsys EDA tools. Full article was not readable; figures come from the an

Why it mattersOpenAI reports its models found up to 14% PPA gains on AMD GPU IP that engineers had already tuned for years, plus silicon-verified gains on its own inference chip.

Passing benchmark tests isn't the same as being correct.

Amazon's BenchProofer generates formal specifications, reference implementations and Lean proofs for coding-benchmark tasks.

Why it mattersTest-passing patches on SWE-Bench Verified can hide real bugs: about half of human-reviewed flagged Opus 4.8 patches had defects.

articleAmazon Science

Build Training-Ready Physical AI Datasets with Two Prompts in NVIDIA Cosmos

NVIDIA describes Cosmos Data Factory, with five blueprints and six composable agent skills that let a coding agent curate, augment, generate and label physical AI data.

articleNakul Arora

How Meta Uses AI Agents to Scale ML Innovations Across Ads Ranking Models

Meta describes using AI agents to propagate ranking-model techniques across its ads fleet.

Why it mattersMeta reports that agents handle over 60% of technique propagation across its ads ranking models, with 3x engineering productivity and an estimated $45M in compute savings. It is a production data point on agents in ML engineering.

article

zed v1.29.1

Zed v1.29.1 adds Claude Fable 5.2 for Anthropic BYOK users, with Mythos 5.2 where an organization has access.

Why it mattersZed now supports Claude Fable 5.2 for Anthropic BYOK users and lets you turn off auto-generated agent thread titles. A new Git Panel branch mode helps review everything changed since a base branch.

repozed-zippy[bot]

crewAI 1.15.42

crewAI 1.15.42 makes ChromaDB an optional extra and defaults Knowledge to local Qdrant.

Why it matterscrewAI 1.15.42 makes ChromaDB optional and defaults Knowledge storage to local Qdrant, which changes install footprint and defaults. It also fixes execution-time limits and raises a dependency floor for a CVE.

repoVidit-Ostwal

Seven new models at scalable price-performance

Microsoft AI launches seven models across code, cyber, image, voice, and reasoning, positioned on price-performance.

Why it mattersMicrosoft AI released seven models, including MAI Code 2, claimed at 98% of GPT-6.1 Sol coding quality at about 40% of the price. Cheaper coding models change the cost of running agents continuously. The figures are vendor-reported.

articlegmays

Comparing pricing across 100 web search APIs

Voxoria compares entry-plan prices per 1,000 searches across 100 web search APIs, checked against each pricing page on 17 November 2026.

Why it mattersSearch is a recurring cost for agents, and entry-plan prices span 90x. A checked per-1,000 comparison with free tiers and full-page-content flags helps when picking or switching a search provider.

articlesanj001

AssemblyAI-STT-3.7 splits transcription into three models

AssemblyAI ships three speech-to-text models.

Why it mattersVoice-agent builders get a streaming model with final transcripts 86 ms after speech ends, plus live speaker labels and 32 languages, at $0.30 to $0.45/hr. Migration details and the Universal model retirement dates are included.

articleAssemblyAI

Stratumus – Qwen3 32B on a 16 GB and a 24 GB Mac over home Wi-Fi

Stratumus explains how Mac Pool runs a 20 GB Qwen3 32B quantization across a 16 GB and a 24 GB Mac using pipeline parallelism, why layer splits tolerate Wi-Fi.

Why it mattersSplitting a model by whole layers sends only boundary activations per token, which makes a model too large for one Mac feasible over ordinary Wi-Fi. Useful for judging what local multi-machine inference costs in latency.

articlejakebowser

langchain langchain-anthropic==1.7.11

langchain-anthropic 1.7.11 adds Claude Fable and Mythos 5.2 support, raises ModelRefusalError when tool calls are not forced, fixes empty thinking block streaming.

Why it mattersIf you call Claude through LangChain, this release adds Fable and Mythos 5.2 support and raises ModelRefusalError on refusals, so refusals no longer pass silently as empty tool-call results.

repogithub-actions[bot]
MistralAI@MistralAI

Mistral released its Vibe coding agent as a desktop app for MacOS and Windows, with agents that run locally or remotely. A free Pro plan trial is offered to builders in Europe. The demo video was not readable, so feature details beyond the announcement are unconfirmed.

Why it mattersMistral's Vibe coding agent now ships as a MacOS and Windows desktop app that can run agents locally or remotely, giving engineers another non-IDE agent option. European builders get a free Pro plan trial.

How To Build The Future: Eric Glyman, Co-CEO of Ramp

Y Combinator interviews Ramp co-CEO Eric Glyman on how the company reorganized around AI.

Why it mattersRamp's CEO describes building in-house agent harnesses, skills, and a shared company brain, and pausing hiring for a quarter to retool. It shows how a large fintech operationalizes agents. Based on the episode description.

videoY Combinator

An audit log for what our open-source agent memory learns offline

Honcho 3.6.0 adds Dream Histories, letting teams audit how offline reasoning changes an agent's memory.

Why it mattersHoncho's background memory reasoning was opaque. You can now list and inspect each dream run, see which conclusions changed and why, and debug questionable inferences in the dashboard or CLI.

articlevvoruganti

40x More Merged PRs With 2.7x the Engineers — Walden Yan, Cognition

Cognition's Walden Yan describes how merged PRs grew 40x against 2.7x engineers.

Why it mattersDescribes concrete practices for agent-heavy teams: every service runnable locally so agents can test, rebuilt VMs, and small owning pods. Figures are self-reported by Cognition.

videoAI Engineer

Tencent Q3: 100T tokens a day and a compute spending surge

Tencent's Q3 2026 results quantify AI demand and spend.

Why it mattersTencent reports WorkBuddy consuming over 100 trillion tokens per day and RMB 65.1B capex (up 225%), with compute prepayments driving negative free cash flow. It is a concrete data point on token demand and compute spend.

Making OpenClaw CI faster and 95% cheaper · Vincent Koc

OpenClaw's CI postmortem after agents flooded it with tests and jobs.

Why it mattersAgents generate more tests, jobs and retries than CI was built for. OpenClaw reports runner-hours down 86.5% using test pruning, diff-based test selection and shard balancing, with tradeoffs and caveats stated.

articlevincent.build

An index of the vibe-coding frontier. Corrections welcome.