Vibeleaderboard
Index — Latest Intelligence

Intel

Page 13
Alex Albert@alexalbert__

Alex Albert shares a routine he runs after each model release: ask the model to read an entire project, find broken, inconsistent, inefficient, or obsolete code, fix what it is sure about, and list the remainder for human review.

Why it mattersA single prompt tells a new model to audit a whole project, fix what it is sure about, and list the rest for review. It is a cheap way to see what a stronger model catches in old code.

Claude Fable 5.2 Scores 65 on Artificial Analysis, Highest So Far

Artificial Analysis ranks Claude Fable 5.2 first on its Intelligence Index (65), while measuring it among the most expensive, slowest and most verbose models.

Why it mattersFable 5.2 leads the Artificial Analysis index at 65 but costs about $26.54 per index task and takes 46 minutes, so it suits hard problems rather than routine coding.

articletheanonymousone
Cursor@cursor_ai

Cursor added Claude Fable 5.2, reporting 60% on CursorBench at max effort and 56.9% at default effort versus 49.2% for Fable 5.1. Pricing is $8 per million input and $40 per million output tokens, and effort settings need re-tuning.

Why it mattersFable 5.2 in Cursor scores 56.9% at default effort (49.2% for Fable 5.1) at 20% lower token prices. Effort levels behave differently, so start at the default instead of carrying over your old setting.

Gemini Robotics 2.5 accelerates real-world general dexterity

Google DeepMind introduces Gemini Robotics 2.5, its most capable vision-language-action model.

Why it mattersGoogle DeepMind's latest vision-language-action model claims broader skills and faster learning, tested across 200+ tasks, signaling frontier model capability moving into physical-world agents.

articledeepmind.google
Artificial Analysis@ArtificialAnlys

Artificial Analysis puts Claude Fable 5.2 first on its Intelligence Index at 65, leading five of nine evals. At max effort it costs about $26.5 per task, but high effort scores 62 with far fewer output tokens, and agentic gains over Opus 5.5 are modest.

Why it mattersFable 5.2 leads at 65, but max effort costs about 2.6x more per task than Opus 5.5. At high effort it scores 62 using about 40% fewer output tokens than Opus 5.5 (max), so effort setting drives cost and value.

Guillermo Rauch@rauchg

Vercel's CEO reports that the newest frontier models can orchestrate more parallel work than a single computer handles. His suggested setup: install the Vercel CLI and plugin, then tell the agent to use the offload-to-sandbox skill to provision cloud sandboxes on demand.

Why it mattersNew models can delegate more parallel work than one machine can run. Provisioning cloud sandboxes through the Vercel CLI plugin's offload skill removes the local compute bottleneck without a custom orchestrator.

Perplexity@perplexity_ai

Perplexity added Claude Fable 5.2 to Perplexity Computer for Pro and Max subscribers and to Search for Max. On its internal WANDR benchmark, Perplexity reports a record 0.851 score at $14.51 per task.

Why it mattersFable 5.2 is now usable in Perplexity Computer for Pro and Max subscribers. Perplexity's WANDR benchmark puts it at 0.851 at $14.51 per task, a cost and performance reference for agent workloads.

claude-code v2.1.321

Claude Code v2.1.321 makes Fable 5.2 the default Fable model with 1M context and published pricing, adds a settings flag for remote-control sessions.

Why it mattersClaude Code now defaults to Fable 5.2 (1M context, $8/$40 per Mtok). It also patches MCP OAuth secrets leaking on redirects and Bash commands escaping the sandbox without a prompt, so update if you use those.

repoclaude-code-publisher[bot]

claude-agent-sdk-typescript v0.3.321

Claude Agent SDK for TypeScript 0.3.321 fixes hook conflicts between ask and defer, dropped hooks, and resume hangs.

Why it mattersIf you use PreToolUse hooks, canUseTool or resume in the Agent SDK, several hook-ordering and hang bugs are fixed, and queued subagent reports in auto mode now share one model call.

repoclaude-code-publisher[bot]

Fable 5.2 vs Opus 5.5: five hands-on tests

Peter Yang tested Fable 5.2 against Opus 5.5 on a 3D travel video, a kart racer, a site redesign, essay writing, and resisting pushback. Fable won four of five.

Why it mattersShows where Fable 5.2 beats Opus 5.5 (3D builds, writing variety) and where it does not: it caved more under pushback in the author's tests. An AGENTS.md candor prompt is offered as a mitigation.

Cognition@cognition

Cognition says Claude Fable 5.2 is available in Devin Desktop and Devin CLI. It reports a #1 FrontierCode 2.0 score of 49.1, and that at low reasoning effort it passes Fable 5.1's score at about one third of the cost.

Why it mattersFable 5.2 is now selectable in Devin Desktop and CLI. Cognition reports it tops FrontierCode 2.0 at 49.1 and beats Fable 5.1's score at low reasoning effort for a third of the cost.

goose v1.58.0

goose v1.58.0 adds opt-in context relevance, a context window report and viewer, new declarative providers including Lemonade.

Why it mattersgoose 1.58.0 adds a context window report and viewer plus opt-in context relevance, which help you debug context usage, and adds Lemonade and other providers with fixes for reasoning replay across tool turns.

repogithub-actions[bot]

Every's day-zero vibe check: Fable 5.2 is spiky, slow and very capable

Every's day-zero review of Claude Fable 5.2 rates it S tier for ambitious coding and knowledge work but slow and token-hungry for routine tasks.

Why it mattersSeparates where Fable 5.2 pays off (ambitious coding and long unattended runs) from where it wastes tokens (everyday writing and quick edits), so you can pick the right model and effort before spending.

Fable 5.2 and Jagged AGI

Ethan Mollick argues that Claude Fable 5.2, after pre-release use, is close to AGI under a narrow computer-work definition.

Why it mattersA practitioner-level read on where the newest Claude model matches expert work and where gaps persist, useful for deciding which tasks to delegate.

articleEthan Mollick
ClaudeDevs@ClaudeDevs

Claude Fable 5.2 is available in the Claude Platform and Claude Code. It runs up to 30% faster than Fable 5.1 and is priced at $8 input and $40 output per million tokens, down from $10 and $50.

Why it mattersFable 5.2 is up to 30% faster than 5.1 and costs $8/$40 per million input/output tokens, down from $10/$50. Teams running Fable on Claude Platform or Claude Code get lower cost per task without changing models.

Meet Claude Fable 5.2

Anthropic introduces Fable 5.2, a model aimed at the most open-ended technical work, with gains in coding, reasoning and vision.

Why it mattersFable 5.2 is available today on paid Claude plans, and Anthropic says Opus 5.5 remains the default for most coding and knowledge tasks. Engineers can decide which model fits open-ended technical work.

videoClaude
claudeai@claudeai

Anthropic announces Claude Fable 5.2 and Claude Mythos 5.2, pitched at the hardest technical work, with improvements in coding, reasoning and vision. The post gives no benchmark figures.

Why it mattersAnthropic released Claude Fable 5.2 and Claude Mythos 5.2, with gains in coding, reasoning and vision. Engineers building on Claude should re-evaluate which model fits their workloads.

Claude Mythos 5.2 predicts the structure of a therapy capsid

Anthropic shows Claude Mythos 5.2 predicting a full AAV gene-therapy capsid by writing code that folds overlapping groups of five chains with open-source models, then using the capsid's symmetry to assemble all 60.

Why it mattersIt shows a model writing orchestration code around open-source folding tools and exploiting structural symmetry to solve a problem too large to fold directly. That is a concrete pattern for agentic scientific workflows.

videoClaude

Introducing Claude Fable 5.2

Anthropic launches Claude Fable 5.2, pitched at the hardest open-ended technical work, with gains in coding, reasoning and vision.

Why it mattersFable 5.2 is available today on Pro, Max, Team and Enterprise plans, aimed at the most open-ended technical work. Anthropic recommends Opus 5.5 as the default for most coding tasks.

videoClaude

Claude Fable 5.2 and Mythos 5.2

Anthropic releases Claude Fable 5.2 and Mythos 5.2, the same model with different cyber and biology safeguards.

Why it mattersFable 5.2 is priced at $8 input and $40 output per million tokens, generates up to 30% faster than 5.1, and has an optional 2.5x fast mode. Mythos access needs a cyber or life sciences program.

articlemeetpateltech

anthropic-sdk-typescript sdk-v1.7.0

Anthropic's TypeScript SDK 1.7.0 adds the claude-fable-5-2 and claude-mythos-5-2 model IDs, a program status webhook event, parent-organization filtering in usage and cost reports.

Why it mattersTypeScript users can call claude-fable-5-2 and claude-mythos-5-2 and use new org-level webhook events and cost report grouping. Structured outputs now keep const and enum.

repoanthropic-sdks-releaser[bot]

anthropic-sdk-python v2.7.0

Anthropic's Python SDK 2.7.0 adds claude-fable-5-2 and claude-mythos-5-2, new webhook and usage report features.

Why it mattersPython users get the new 5.2 model IDs, and Bedrock, Vertex and Foundry clients lose several bugs (missing count_tokens, signing errors, custom headers ignored).

repoanthropic-sdks-releaser[bot]

Claude Electronics Design

Anthropic reports Claude Fable 5.2 designing a speakerphone, voice-assistant hub and virtual pet from short prompts.

Why it mattersShows a model carrying a hardware design through schematic, PCB routing, enclosure CAD and firmware, with built prototypes working without circuit fixes. It marks agentic capability moving into physical engineering workflows.

articleAnthropic research

E-means: 10x faster k-means at equal loss

Meta FAIR's e-means replaces Lloyd's algorithm with a mini-batch scheme tuned by scaling laws.

Why it mattersk-means underlies vector quantization and codebook learning for embeddings and retrieval. A drop-in method reaching lower loss in 1.2 to 3 epochs instead of 25 cuts index training cost.

articleporridgeraisin

Elevenagents Campaign Manager

ElevenLabs launches Campaign Manager for ElevenAgents.

Why it mattersVoice-agent builders get a managed way to run multi-touch outbound campaigns with retry scheduling, voicemail handling and cross-call context instead of building that orchestration themselves.

Alexandr Wang@alexandr_wang

Meta released Muse Ember, its smallest model yet and the first trained with Muse Wonder Full. Citing Artificial Analysis, it offers close to Muse Spark 1.3 (max) intelligence at 17x lower price, and is available through the API and muse code.

Why it mattersA new small model is available by API and in muse code at a claimed 17x lower price than Muse Spark 1.3 (max) for nearly the same intelligence. Check cost-sensitive agent workloads against it.

ollama v0.40.16

Ollama v0.40.16 adds Ripple Preview, fixes silent truncation in the Responses endpoint, routes Claude Code to the selected model via launch.

Why it mattersOllama can now run Ripple Preview, a 65B agentic coding model, and /v1/responses no longer silently truncates input by default. Check any agent that relies on the old behavior.

repogithub-actions[bot]
Artificial Analysis@ArtificialAnlys

Artificial Analysis benchmarks Meta's Muse Ember: an Intelligence Index score of 36 at $0.14 per task versus $0.41 for Haiku 5.5, strong on AutomationBench-AA, mixed on agentic coding, and weak on knowledge accuracy because it often declines to answer.

Why it mattersMuse Ember matches Claude Haiku 5.5 on the Intelligence Index at $0.14 versus $0.41 per task, and beats it on Terminal-Bench 4.0. It trails on FrontierCode 2.0 and declines 64% of knowledge questions.

AIatMeta@AIatMeta

Meta announces Muse Ember, its smallest and most efficient model, built for agent work. It offers a 1M-token context, native multimodal input, sequential tool calling, test runs and targeted code fixes. The post gives no benchmarks or pricing.

Why it mattersMuse Ember is a small Meta model with a 1M-token context and native text, image, audio and video input, aimed at multi-step tool calling and test-and-fix coding agents. It is a lower-cost option to evaluate for agent workloads.

Alexandr Wang@alexandr_wang

Muse Ember is a small, cost-efficient model aimed at repetitive agent workloads. Its maker claims better intelligence per dollar than Haiku and Luna at a third of the cost or less, and suggests pairing it with the larger Wonder model as planner. Live on Muse Code and the API.

Why it mattersA cheap small model paired with a larger planner (Wonder plans, Ember handles volume) is the pattern for cutting cost in high-volume agent loops. It is available on the API today, so you can test it against your current flash model.

How One Founder Manages 20 AI Coding Agents at Once - Ryan Carson, Untangle

Ryan Carson walks through Untangle's cloud agent workflow.

Why it mattersGives a working pattern for scaling coding agents: one manager agent over a ticket board, bounded review loops, a merge queue, and a mandatory human watch of each feature video.

videoAI Engineer

Our Smallest, Most Efficient Model: Muse Ember

Meta releases Muse Ember, its smallest model, aimed at repeated agent work like tool calls, coding subtasks, routing and extraction.

Why it mattersMuse Ember targets high-volume agent subtasks like tool calls and extraction. Meta reports it matches Claude Haiku 5.5's best Artificial Analysis score at about a third of the cost, which matters for routing cheap work.

articledvaplima
Nous Research@NousResearch

Nous Research launched browser access for Hermes Cloud agents, with the same interface as Hermes Desktop. Each agent runs in its own container around the clock, handling scheduled jobs and messages. Plus, Super, and Ultra plans each include one cloud agent.

Why it mattersYou can now run a hosted Hermes agent in its own container, 24/7, with scheduled jobs and messaging channels. Open it in a browser or connect Hermes Desktop to continue the same chats.

Morph Ultra: flat $200 plan for Kimi K3.1 and DeepSeek

Morph Ultra sells unlimited Kimi K3.1 and DeepSeek V4.1 Flash for $200 a month, limited to 1,000 subscriptions. Per-minute caps apply (Kimi K3.1 about 24K uncached input tokens).

Why it mattersIt offers flat-rate access to two open models for coding agents, with published rate limits and Anthropic-compatible endpoints for Claude Code. Check the per-minute caps, since Kimi K3.1 allows only about 24K uncached input tokens a minute.

articleMorph

What Frontier Labs Buy: A Data Deep Dive with Handshake AI — Garrett Lord & Team

Latent Space walks through Handshake AI's RL environments in finance, life sciences and alignment, showing where agents still fail, including a grader exploit and a missed tax rule.

Why it mattersIt shows what frontier labs pay for in RL environments and concrete agent failure modes, such as reward hacking a grader, that evaluation builders should design against.

articleswyx .ai

Fact-Checking at Scale: Building a More Accurate and Efficient Synopsis Factuality System

Netflix details its factuality service for synopses.

Why it mattersShows how a large team automates factuality judging of generated text across thousands of titles and languages, a pattern reusable for LLM output evals.

Extending LLDB to Debug Applications That Span CPUs and GPUs

Meta describes an accelerator plugin architecture in lldb-server that separates shared debugger infrastructure from hardware-specific control, presenting CPU and GPU as separate processes in one debugging session for AI workloads.

Why it mattersDebugging code that crosses host and GPU boundaries usually needs separate vendor tools. This design shows how one LLDB session can model CPU and GPU as separate processes behind a shared plugin layer.

article

openai-python v3.48.0

openai-python 3.48.0 adds stream options for agent sessions and recognizes the gpt-6-rosalind model in beta Responses, plus dependency bumps and consolidated AGENTS.md guidance.

Why it mattersPython users can now pass stream options for agent sessions and reference gpt-6-rosalind in the beta Responses API after upgrading to 3.48.0.

repoopenai-sdks[bot]

Mistral CEO: Why OpenAI & Anthropic Will Come for Your Business | Are Chinese Models Truly Safe?

Mistral CEO Arthur Mensch discusses compute procurement, inference market dynamics, sovereign AI and concentration risk from closed labs.

video20VC with Harry Stebbings

Drex: how Nace.AI built a fast, calibrated decision model

Drex Max v1.1 is #1 on Decision Index @ArtificialAnlys and going open-weight soon!

Why it mattersDecision models return calibrated probabilities over your options in one forward pass, avoiding output parsing and long generation.

articleNace AI

Your Experts Have Partners. There's HOPE for Keeping Them Together

An Amazon team proposes higher-order expert pruning for MoE language models, scoring experts together instead of individually to shrink inference memory, motivated by Qwen3.5-122B's expert-heavy footprint.

Why it mattersExpert MLPs dominate MoE memory, so pruning methods that account for expert partnerships could cut serving cost without training. Results were not visible in the excerpt.

articlehuggingface.co

More Realistic Images More Stable Editing Across 100 Rounds Introducing Seedream 5 5

ByteDance's Seedream 5.5 image model adds pixel-level preservation across up to 100 editing rounds, native RGBA transparent-background generation.

Why it mattersSeedream 5.5 cuts the degradation that accumulates over repeated AI image edits and outputs ready-to-use transparent assets, reducing manual cleanup in creative production pipelines.

Your KV Cache Can Spot Attacks on Your AI Agents - Ari Azarafrooz, Intrinsec AI

Intrinsec AI's Ari Azarafrooz shows how KV cache data already written by inference engines can flag anomalous inputs such as hidden spoken instructions.

Why it mattersOffers a defense layer that works where text filters cannot, by flagging unusual KV pages against a baseline learned from normal traffic.

videoAI Engineer

Will AI Agents Pretend to Be Human to Get Past a Check? - Maty Bohacek, Stanford

Stanford's Maty Bohacek tests what agents do when a human check blocks them, finding many use voice or video cloning to impersonate a person.

Why it mattersWarns that phone or video human checks are not a safe control against agents, since agents with cloning tools often impersonate people despite instructions.

videoAI Engineer

The One Thing Every AI Agent Framework Agrees On - Mislav Ivanda, Daytona

Daytona's Mislav Ivanda compares three ways to run agent code and notes frameworks like CrewAI, AG2 and the OpenAI Agents SDK converging on sandboxes.

Why it mattersHelps you choose where agent-generated code runs before picking a framework, with a clear comparison of isolation options and their containment limits.

videoAI Engineer

Why No One Agrees on What Their AI Tokens Cost - Pratyusha Duvvuri, Cisco

Cisco's Pratyusha Duvvuri defines tokenomics as observability of token usage and cost, explains how engineers, managers and finance need different views.

Why it mattersExplains why provider dashboards fall short for per-person and per-team token cost attribution and what to track, including forecasting and handling bills that change after the fact.

videoAI Engineer

Training Your Own Agent Models With RL Is Now Practical - Jason Mancuso, Modal

Modal's Jason Mancuso lays out why RL on agent traces can beat SFT, a stable async GRPO recipe.

Why it mattersGives a specific stable recipe for asynchronous RL on your own agent traces and evals, and a cost lever from spreading rollouts across GPU types.

videoAI Engineer

Right Answer, Wrong Reasons: AI Forecasts Costco - Hassan El Mghari & Omar El Mghari, Together AI

Together AI's team has GPT-OSS-120B forecast Costco profits from pre-2024 filings using structured formulas and parallel runs, then audits the assumptions to expose two errors that cancelled out.

Why it mattersDemonstrates having models emit auditable formulas instead of doing math, and shows that a good aggregate result can hide offsetting errors, so expert checks belong in code.

videoAI Engineer

How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp

Warp's Zach Lloyd makes the case for cloud software factories as open infrastructure defined in code.

Why it mattersFrames agent infrastructure as code with LLM-judged run scoring and per-team benchmarks, so you can compare model cost and quality on your own tasks.

videoAI Engineer

Stop Doing Agent Archaeology on Your Team's Code - Richard Feldman, Delta

Richard Feldman explains why agent-driven work loses reasoning when handed to teammates and demos Delta, where conversation, code, review and live terminal share one thread.

Why it mattersNames the cost of losing agent reasoning at handoff and shows one design: a shared thread holding conversation, diffs, review and terminal state.

videoAI Engineer

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