Vibeleaderboard
Index — Latest Intelligence

Intel

Page 12

dspy 3.5.0

DSPy 3.5 replaces LiteLLM with the lightweight lm15 engine, graduates RLM, Flex and GEPA from experimental, folds ReActV2 into dspy.ReAct.

Why it mattersDSPy 3.5 makes LiteLLM and the OpenAI library optional, cutting default install from 206.7 MiB to 30.9 MiB. It also changes prompts and defaults, so optimized programs need re-evaluation after upgrade.

repoisaacbmiller
GoogleAI@GoogleAI

Google releases Gemini 4 Carbon, its top model for long-horizon agentic work across software engineering, cybersecurity and science. Argon remains the pick for most coding. Available via API in AI Studio and Android Studio, Antigravity for Ultra subscribers, and a cyber defender program.

Why it mattersGemini 4 Carbon is positioned for long-horizon agentic tasks like multi-step debugging, while Argon stays the default for most coding. You can now choose between the two per task and apply for the cyber and research access programs.

Dan Shipper@danshipper

Every's day-zero test of Gemini 4 Carbon: a big coding gain over Argon (44% vs 21% on one benchmark run), strong long builds, but it overreaches on open-ended backlogs and writes huge files. Pricing is $2/$10 per million tokens until Dec 1.

Why it mattersCarbon scored 44% versus Argon's 21% on one Every coding benchmark run, but it overran scope on a backlog, doing 15 tasks when 6 were ready. Scope it to one big job with a clear finish line and review the output.

Decision Models Inherit Their Base Model's Risks. Hirundo Fixes Them

Hirundo reports that most open decision models, built on Qwen, hold or block China-related content while passing equivalent non-China content.

Why it mattersSmall gating models for moderation and agent permissions can silently inherit a base model's political bias, with no visible refusal. Teams choosing decision models should test matched pairs before deploying.

articlemishana4life

Gemini 4 Carbon is now available

Google DeepMind makes Gemini 4 Carbon generally available after a two-week defender-only preview.

Why it mattersCarbon is now generally available for long-horizon agentic coding and debugging, using more tokens than Argon for deeper verification. Engineers can choose between Argon for routine work and Carbon for hard tasks.

articledeepmind.google

NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure

NVIDIA donates OpenShell, a runtime for securing and governing autonomous agents, to the CNCF Sandbox and contributes to Google-led Agent Substrate for long-running agent workloads on Kubernetes, alongside new DSX AI Factory components.

Why it mattersAn open, foundation-governed runtime for securing autonomous agents and a Kubernetes framework for long-running agent workloads give engineers vendor-neutral building blocks for agent deployment.

blogWarren Barkley

Introducing the Multimodal Knowledge Agent benchmark

Google DeepMind and Artificial Analysis launch MKA, a benchmark for agents doing visually heavy knowledge work such as reading blueprints and schematics and producing decks and spreadsheets.

Why it mattersMKA tests whether agents can work from visually complex files rather than text alone. That gives teams building document agents a more realistic evaluation target than text-only knowledge-work benchmarks.

blogblog.google

Gemini 4 Carbon is now available

Google DeepMind makes Gemini 4 Carbon generally available.

Why it mattersGoogle now offers a tier above Gemini 4 Argon for multi-step coding, debugging and research, at a higher token cost. Engineers can route only the hardest agent tasks to it.

blogblog.google

Start and Restore GPU Workloads in Seconds with NVIDIA Snapshot

NVIDIA open-sources Snapshot, which checkpoints initialized CPU and GPU memory on Kubernetes.

Why it mattersCold-starting large inference replicas takes minutes; restoring a captured CPU and GPU memory snapshot takes seconds. This makes scale-to-zero and fast autoscaling practical for big MoE models served via Dynamo.

articleRon Kahn

Agent Substrate: Advancing Kubernetes for Agents at Scale

NVIDIA's Agent Substrate runs many agent sessions on few warm Kubernetes pods by checkpointing idle state and restoring it on any free worker.

Why it mattersAgent fleets alternate between idle time and bursts of LLM calls. Checkpointing idle sessions onto a small warm pool, with KV-cache-aware routing, cuts the cost of running many coding-agent sessions.

articleEkin Karabulut

NeuroBrix – PyTorch-free local inference with a Rust grammar engine

NeuroBrix 0.6.5 adds constrained decoding to its OpenAI-compatible server.

Why it mattersLocal serving can now enforce JSON schemas and forced tool calls at the decoding level on an OpenAI-compatible API, so structured output and tool use do not depend on the model complying.

articlebenkelaya

AI Vision Index, private tests of how well 36 AI models see

A private vision benchmark of 36 models using rendered images, photos, and video, with test data never published.

Why it mattersPublic vision benchmarks can be trained on. This index uses held-out tests and reports price and speed alongside scores, so you can pick a vision model on uncontaminated results.

articleLLD

Introducing AlphaProof Code: a breakthrough in verifying complex software at depth

Google DeepMind's AlphaProof Code splits formal verification between human experts and multiple LLM agents.

Why it mattersAgents paired with human experts proved security properties of real Signal protocol code, not a model of it. That shows agentic formal verification can now reach deployed cryptographic software.

articledeepmind.google

Opening the datasets behind our AI weather models

DeepMind releases SatWeave, a unified global mosaic of five geostationary satellites, and PARDIG, an AI-based global precipitation dataset.

Why it mattersDeepMind now provides historical archives and live feeds of a global geostationary satellite mosaic and an AI precipitation dataset, so researchers can build weather and hydrology models on the same inputs used for WeatherNext 3.

articledeepmind.google
Vals AI@ValsAI

Vals AI and Nextladder released an Access to Justice benchmark of 252 legal-aid scenarios with expert rubrics. Across 23 models the best scored 76.4%. Models explained the law well but rarely gave local resources, and price did not predict performance.

Why it mattersShows that models explain law well (78 to 87%) but rarely point to local legal aid (30 to 50%), and that free-tier models score about the same as premium ones. Useful for anyone building legal-help products.

opencode v2.2.7

opencode v2.2.7 adds GitHub Copilot Auto model routing with fallback to the default chat model, plus side panel tab shortcuts and read/unread session marks.

Why it mattersOpenCode now lets Copilot's router pick models per session at the Auto discount. SDK users must switch from response.image to response.images[0] and handle plugin-provided entries appearing late.

repogithub-actions[bot]

AI usage generates a fragile codebase

A developer who now writes all code with AI lists four ways it makes large SaaS codebases fragile, including shallow tests that miss real gaps and overly complex code.

Why it mattersNames concrete failure patterns in agent-written code, such as tests that only cover paths the code already handles and needless helper layers. Useful for deciding what to review in AI-heavy codebases.

articleplun9

browser-use 1.0.1 — browser-use 1.0.1

browser-use 1.0.1 patches the 1.0.0 release.

Why it mattersIf you run browser-use over MCP, 1.0.1 closes an allowed_domains bypass through cloud browsers and lets the CLI MCP server run in headless CI or containers by installing its own Chromium.

repogregpr07

Introducing Live 3

Fish Audio's Live 3 is a full-duplex voice model that keeps listening while speaking.

Why it mattersLive 3 moves turn-taking into the speech model itself, replacing the threshold rules usually bolted onto voice agent pipelines. It can acknowledge while listening and tell backchannels from real interruptions.

articleFish Audio editorial sitemap

Introducing Our Smallest, Most Efficient Model: Muse Ember

Meta's Muse Ember is a small, low-cost model aimed at repetitive agent work such as tool calls, routing and extraction.

Why it mattersMuse Ember matches Haiku 5.5 and GPT-6 Luna at $0.10/$0.50 per million tokens for short prompts, with cheaper cached input. Long-prompt pricing differs, so agent workloads should compare cost per task.

articlesimonwillison.net

openai-python v3.49.0

OpenAI's Python SDK 3.49.0 adds typed async function tools to the API surface and fixes microphone recording encoding so audio is written as valid PCM WAV.

Why it mattersThe OpenAI Python SDK now supports typed async function tools, so agent builders can define async tool functions with type information instead of hand-writing wrappers.

repoopenai-sdks[bot]

[AINews] Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins

Claude Fable 5.2 takes the top model spot with $8/$40 per million token pricing, yet task costs rise because it writes far more tests and spawns checking subagents.

Why it mattersFable 5.2 cuts per-token price 20% but spends more tokens verifying work, so cost per task rose from $67.16 to $88.15 on GDPx. Choosing the effort setting matters for production cost.

articleswyx .ai

Amazon saves millions by replacing Sonnet 4.6 with Strands Decider

Two Amazon engineers replaced Claude Sonnet 4.6 with Strands Decider, a decision model, in a very large internal ticketing system, reporting savings in the millions.

Why it mattersDescribes swapping a frontier LLM for a purpose-built decision model in a high-volume Amazon ticketing workflow to cut cost. Suggests routine classification steps in agents may not need a frontier model.

articlealoukissas

Prompting Fable 5.2: effort, ownership and harness changes that matter

Anthropic's Fable 5.2 prompting guide covers effort levels, preventing early stops, an ownership prompt that triggers refactors (with roughly 70 to 80 percent more output tokens), subagent restraint, progress-update thinking blocks, and prompts from earlier models worth removing.

Why it mattersConcrete, tested prompt and harness changes for Fable 5.2, such as how effort levels behave, why an ownership sentence changes refactoring behavior, and which prompts written for earlier models to remove, with token-cost effects measured.

articleSimon Willison

Nex N2.6 released

Nex AGI published Nex-N2.6, a 321B-parameter image-text-to-text model under Apache 2.0 in FP8, with Transformers, vLLM and SGLang serving recipes.

Why it mattersA new Apache-2.0 321B multimodal model in FP8 can be self-hosted through vLLM or SGLang. Check independent benchmarks before building on it, since the page shows none.

articleNEX AGI model releases

Quoting Prompting Claude Fable 5.2

Anthropic's prompting docs for Claude Fable 5.2 note the model reacts to stated code ownership.

Why it mattersTelling Claude Fable 5.2 that the codebase is its own made it merge repeated code in about half of tries versus under a tenth without it. A one-sentence prompt change shifts refactoring behavior.

ZTHR – what replaces code review when AI agents write all the code

ZTHR is a pattern language of two principles and 33 patterns for delivering AI-written software without relying on diff review.

Why it mattersProposes governed building blocks, checks proven to fail on target defects, and locked user-intent statements as the quality gate for AI-written code, instead of human diff review.

articleSymmetricalZero
Artificial Analysis@ArtificialAnlys

Artificial Analysis reports Claude Fable 5.2 leading VIALS-AA at 67%, ahead of Gemini 4 Carbon and GPT-6 Astra. Fable averages about 37 tool calls per task and Astra about 8, and Anthropic tool use climbs with capability while OpenAI models trend the other way.

Why it mattersShows that top agentic visual-reasoning scores come from heavy tool use (about 37 calls per task for Fable 5.2 versus about 8 for GPT-6 Astra), which drives cost. Harness design affects results as much as the model.

claude-code v2.1.322

Claude Code 2.1.322 extends fast mode to Vertex across the global, US and EU regions, with automatic fallback to standard speed when Vertex declines the request.

Why it mattersTeams running Claude Code through Google Vertex can now use fast mode in the global, US and EU regions. If Vertex refuses it, the session falls back to standard speed instead of failing.

repoclaude-code-publisher[bot]

browser-use 1.0.0 — browser-use 1.0: CDP is all you need

browser-use 1.0 rebuilds the web agent around a persistent Python REPL on CDP, removing 136,866 lines.

Why it mattersBrowser-use 1.0 drops the DOM serializer and action schema for a model-written Python REPL over Chrome DevTools Protocol, reporting up to 33 points higher scores and 2.7 to 5.2x lower cost.

repogregpr07

60% of Our Data PRs Now Merge With No Human Review - Chetas Joshi, Robinhood

Robinhood's data platform engineer describes how agent-opened pull requests to production data pipelines are reviewed at machine speed.

Why it mattersCI passing does not show a schema change is safe for downstream consumers. Robinhood describes automated review gates that let a majority of data PRs merge without a human, relevant to anyone facing agent-generated PR volume.

videoAI Engineer
Artificial Analysis@ArtificialAnlys

Artificial Analysis ranks Claude Fable 5.2 first on AA-3D at 1271 Elo, ahead of Sonnet 5.5 and Opus 5.5 with a non-overlapping interval. Agents build three.js scenes in a sandbox, then blind human votes produce the ranking.

Why it mattersFable 5.2 leads AA-3D by 67 Elo with a non-overlapping interval, and beats Fable 5.1 by 132. The benchmark tests agents building interactive three.js scenes with Blender, physics, and a headless browser.

Inference Performance Index

Epoch AI's Inference Performance Index scores chips on concurrent requests served at given output speeds, factoring in memory and communication.

Why it mattersFLOP/s comparisons understate newer chips for inference. Epoch estimates GB300 serves about 25x the requests of H100 per chip, which changes how you reason about future inference cost and capacity.

articleEpoch AI publications

Pay for Hugging Face Hub with your Google Cloud Account

Hugging Face announces Google Cloud Marketplace billing.

Why it mattersTeams on Google Cloud can now bill Hugging Face compute, inference and subscriptions to their existing GCP invoice, skipping separate vendor onboarding and credit cards.

articlehuggingface.co

Gemini 4 Carbon now available on AI Gateway

Vercel adds Gemini 4 Carbon to AI Gateway with a 1M-token context window and a 50% launch discount through November 30, 2026.

Why it mattersYou can call Gemini 4 Carbon through AI Gateway using google/gemini-4-carbon, at 50% off through November 30, 2026, including from Claude Code, Codex and Cursor via the Vercel CLI setup.

articleJerilyn Zheng

Decision Models in Hugging Face

Hugging Face introduces a transformers pipeline for decision models, which evaluate a fixed set of options and return a probability distribution without generating text.

Why it mattersRouting and triage steps in agent systems can score a fixed set of options directly instead of generating and parsing JSON from an LLM, which cuts latency and output variance.

articlehuggingface.co

Shopify App Rebuild

Shopify engineers describe rebuilding their 300-screen mobile app from React Native to Swift and Kotlin in 12 weeks with 30 engineers.

Why it mattersShows what had to change around agent-written code (architecture, tests, review, emulator infrastructure) to ship a large native app quickly. Useful for teams scaling coding agents.

articleShopify Engineering

Google Deepmind Multimodal Knowledge Agent

Google DeepMind's Multimodal Knowledge Agent benchmark tests agents on 200 private visual tasks across 10 industries, graded by weighted expert rubrics.

Why it mattersAgents working from schematics and scans are scored on expert rubrics, with cost per task reported. Scores come within about 10 points at roughly 30x cost spread, so model choice for visual document work is a cost decision too.

articleArtificial Analysis articles

MiniMax H3 Image and Highspeed now available on AI Gateway

Vercel AI Gateway adds MiniMax H3 Image and H3 Image Highspeed for text-to-image and instruction-based editing.

Why it mattersYou can call MiniMax H3 image generation and editing through one AI Gateway key with no separate MiniMax account, using generateImage with the model IDs shown. A Highspeed variant trades latency for the same options.

articleJerilyn Zheng

llm-anthropic 0.31

llm-anthropic 0.31 adds Fable 5.2 and Mythos 5.2.

Why it mattersAt max effort, Fable 5.2 can burn 128,000 thinking tokens and return nothing, a regression from Fable 5 and 5.1. Use xhigh or budget for this failure if you rely on max effort.

articlesimonwillison.net

Locating the source of a mysterious beep with Claude

Simon Willison used Claude Code to build Beep Locator, a web app that triangulates an intermittent beep using several devices as microphones.

Why it mattersShows a real Claude Code build where the first approach (devices playing sounds to locate each other) underperformed and a manual rough-placement mode fixed it, a useful pattern for iterating on agent-built prototypes.

articlesimonwillison.net

Fable 5.2 one-shotted my nephew's dream game

A hands-on report of using Fable 5.2 with image and audio generation APIs to turn a child's design drawings into a complete multi-level beat 'em up, contrasted with a weaker earlier Claude Code build.

Why it mattersShows what a single long-running agent prompt with image and sound generation tools can produce for a full game, compared with an earlier attempt. Gives a concrete prompt and workflow to calibrate expectations.

articlelexandstuff

Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo

A Monzo-affiliated pair build a financial-analyst agent harness live around a wedding budget.

Why it mattersFor agents handling money, the talk argues for deterministic code for arithmetic, evidence trails, and verification against golden examples. It is a practical checklist for building trustworthy analytical agents.

videoAI Engineer

Fable 5.2 in Copilot: data retention and admin opt-in

GitHub made Claude Fable 5.2 generally available in Copilot across IDEs, CLI, and the coding agent.

Why it mattersUsing Fable 5.2 in Copilot means accepting Anthropic data retention for safety classifiers. Business and Enterprise admins must enable its policy, and ZDR access for eligible enterprises ends with the calendar year unless EFS is adopted.

articleGitHub
OpenAI@OpenAI

OpenAI describes Control Room, a central system tracking research inference workloads, including agentic sampling, that works with misalignment monitoring and halts model output within 35 seconds of a stop decision. Only the announcement thread was readable.

Why it mattersOpenAI describes a central control plane over every research inference workload, including agentic sampling, that halts output within 35 seconds of a stop decision. It is a reference point for kill-switch design in agent infrastructure.

Epoch AI@EpochAIResearch

Epoch AI reports Claude Fable 5.2 at 174 on its Capabilities Index, a record above GPT-6 Astra's 169 and above the trend for reasoning models. The score comes from pre-launch testing on Epoch's own benchmarks and may be revised.

Why it mattersAn independent evaluator puts Fable 5.2 at a record 174 on the Epoch Capabilities Index, above the prior high of 169 and above trend. That is evidence of a capability jump, though Epoch cautions the score may be revised.

claudeai@claudeai

Anthropic shows pre-release Fable 5.2 and Mythos 5.2 work: three working electronic devices designed from one prompt, and multi-step chemistry synthesis routes rated early-grad-student level by external chemists. The synthesis routes have not been lab-tested.

Why it mattersFable 5.2 produced specs, component choices, four-layer PCB routing, enclosures and firmware for three working devices from one paragraph.

Anthropic's Guide to Choosing a Claude Model and Effort Level

An Anthropic playbook on choosing among Claude Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5.

Why it mattersStart on Opus 5.5 at medium effort, try raising effort one level before switching models, and run subagents on Haiku or Sonnet. Re-running only failures at high effort cost about $0.17 per task versus $0.29 at high for all.

articleClaudeDevs

Vercel Sandbox snapshot creation is now up to 10x faster

Vercel Sandbox snapshot creation is up to 10x faster (median under 0.4s from 4.4s) and restore is 50% faster (250ms p50), thanks to frame reuse and more parallel I/O in the VHS format.

Why it mattersAgent sandboxes that snapshot on stop and resume now do so in under 0.4s median with restores at 250ms p50, cutting latency for persistent agent sessions with no configuration.

articleAndy Waller

Inkling 2 Small is free on OpenCode: 1M context, agentic-coding tuned

Thinking Machines makes Inkling 2 Small, a model trained for agentic coding and knowledge work, free for a limited time on OpenCode.

Why it mattersInkling 2 Small offers a 1M-token context and 131K output, priced at $0.30 in and $1.00 out per 1M tokens, free for a limited time on OpenCode. You can try an agentic-coding model at no cost; no capability benchmarks are published yet.

articleThinking Machines

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