Vibeleaderboard
Index — Latest Intelligence

Intel

Page 04

servers v2.5.0

The MCP servers v2.5.0 release patches eight vulnerabilities across mcp-server-git, server-filesystem and server-everything.

Why it mattersEight vulnerabilities are fixed in the reference MCP servers. In mcp-server-git, a repository's own config could make a merge or read run commands with the server's privileges, so upgrade to the fixed versions.

repocliffhall

Introducing Vercel Board

Vercel launches Board, an experimental message board for coordinating many agents with per-task threads, access grants for humans and agents.

Why it mattersMulti-agent setups lack a shared place for tasks, delegation, status, and logs. Board gives agents scoped access and one thread per task, and works across Claude Code, Codex, and other harnesses.

articleJulian Benegas

python-genai v3.6.0

python-genai 3.6.0 adds tool retrieval and function result schemas to the public SDK.

Why it mattersThe Gemini Python SDK now exposes tool retrieval and function result schemas publicly, which matters for typed tool-calling code. It also retries auth transport errors and keeps responses with oversized integers.

reporelease-please[bot]

Hy4 base released

Tencent releases the pretrained Hy4-base checkpoint, a 771B-parameter multimodal mixture-of-experts model on Hugging Face.

Why it mattersThe pretrained base checkpoint of a 771B multimodal MoE is available for fine-tuning and research on open weights. Check the custom Tencent license before building on it.

articleTencent model releases

Hy4 FP8 released

Tencent publishes Hy4-FP8, an FP8-quantized version of its 774B-parameter multimodal mixture-of-experts Hy4 model.

Why it mattersA 774B multimodal MoE in FP8 lowers the memory needed to self-host a frontier-scale open model. It ships with vLLM and SGLang serving paths, under a custom Tencent license that needs checking before use.

articleTencent model releases

Krea Originals Rocket And Recess

Krea's team details how two short films were built on a single Nodes canvas.

Why it mattersShows a working pipeline where Claude writes shot-level prompts and sends them to a canvas through MCP, with consistency tricks like approving image references before costly video runs.

articleKrea editorial sitemap

openai-python v3.55.0

openai-python 3.55.0 adds Live MCP server configuration to the API surface and hides Managed Agents session budget controls and events.

Why it mattersThe official Python SDK now exposes Live MCP server configuration, so agent builders can wire MCP servers into live sessions through the supported client instead of custom code.

repoopenai-sdks[bot]

Language Models Can Rewrite Their Own Weights

A 27B model diagnoses its failures and edits its own weights with no gradients or training, gaining 6.6 points on 490 unseen problems.

Why it mattersReports a model improving itself by editing its own weights through reasoning, with gains that persist and add no per-answer cost.

articleJiajun Fan, Neil He, Changyi Yang, Andy Mao, Xi Zhu, Chaoran Cheng, Kaiqu Liang, Shuaike Shen, Zetian Li, Ruihan Guo, Ge Liu

Limite 1B - Violetto: Technical Report

Limite 1B Violetto is a one-billion-parameter math reasoning model trained from scratch on 237B tokens with 131k context.

Why it mattersA 1B open model claiming 94% on AIME 2026 is a cheap, fast math solver that can run as a parallel component in larger systems. Weights, checkpoints and serving code are released under Apache 2.0.

articleMario Prignano, Gabriele Cirillo, Alessio Morosini, Luca Cerovaz, Alessandro Bartolocci, Emanuele Rodol\`a, Giulio Starace, Francesco Pappone

Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research

Red-team/blue-team game with a human expert as proxy saboteur on AI safety research. Human auditors score AUROC 0.59 and fully identify sabotage in 17% of cases.

Why it mattersAuditing full research transcripts catches sabotage poorly: humans are near chance and the best frontier model reaches AUROC 0.73. Anyone planning to delegate R&D to agents and rely on review should treat transcript audits as weak.

articleKeshav Shenoy, Anthony Zang, Eric Gan, Jacqueline Tay, Amy Ngo, Sebastian Prasanna, Julian Stastny, Vivek Hebbar

Parallel web search is free on AI Gateway and eve through January 31

Vercel made Parallel fast-mode web search free on AI Gateway through January 31, 2027, and set it as the default search for eve agents.

Why it mattersAgents on AI Gateway can add web search with any gateway model and no separate Parallel key, at no cost in fast mode until January 31, 2027. Other modes bill at standard rates, so check which mode you use.

articleShar Dara

Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability

Stress test of chain-of-thought monitoring.

Why it mattersFine-tuning that merely makes reasoning more controllable can make chain-of-thought monitors much less effective, even with no evasion training. Teams relying on CoT monitoring for agent oversight should not assume it holds.

articleZheng-Xin Yong, Benjamin Arnav, Miles Wang, Tomek Korbak, Olivia Watkins

From Preference Data to Personalization: Tracing Sycophancy in Large Language Models

Introduces a validity-aware sycophancy grader and SycoDrift, a 3,298-probe eval set.

Why it mattersSuggests human preference data is a weak driver of sycophancy and that agreement pressure also comes from judge specifications and in-context personalization. That helps teams target where to fix it.

articleHimaghna Bhattacharjee, Yasha Sheynin, Shivam Singhal, Arjun Subramonian, Ankur Samanta, Julian Michael, Smitha Milli, Maximilian Nickel, Claudia Shi

DiscoveryAgent: A Framework for Scientist-centric Agentic AI

DiscoveryAgent pairs autonomous investigation with scientist-directed hypothesis steering and checkpoint-based audit agents.

Why it mattersShows a design for oversight of autonomous research agents: checkpointed records let audit agents trace claims, reproduce analyses and catch errors while scientists steer hypotheses between rounds.

articleOwen Queen, Ioana Ciuca, Emma Dann, Jean Somalwar, Carine You, Steven Dillmann, Viraj Manwadkar, Evan Gorrell, Harrison G Zhang, Jonathan K Pritchard, Risa H Wechsler, James Zou

Does Your Agent Need to Know Its Budget? A Placebo-Controlled Experiment on LLM Coding Agents

Placebo-controlled experiment across 5,835 program-repair runs on two open models. Enforced token caps cut resolution by 20 points and tokens per resolved task about twelvefold.

Why it mattersTelling a coding agent its token budget did not help, while enforcing a cap cut resolution by 20 points and tokens per resolved task about twelvefold. Teams setting agent budgets can skip budget-disclosure prompts and tune the cap instead.

articleIftesam Nabi, Jarif Ul Alam, Sunzid Hassan, Sabbir Alom Shuvo, Md Faiaz Al Islam, Yead Rahman, Mehrab Hosain

Tencent open-sources Hy4 flash, a 12B-active 1M-context MoE

Tencent releases Hy4 flash, Flash-FP8 and Base as open weights.

Why it mattersHy4 flash is an open-weight 1M-context multimodal MoE with 12B active parameters, with a built-in speculative-decoding drafter. Check the Section 10 MaaS license terms before serving it commercially.

articleTencent Hy

Why is Gemini 4 Carbon taking so long?

Kernel traced every step of Gemini 4 Carbon's bug-fixing runs against Opus and Astra. Slower inference explains only part of the gap.

Why it mattersSlow agent runs may come from longer trajectories, not slow inference. The trace data shows roughly 7x more steps and 16x more input tokens, which affects how you budget time and cost when picking a coding model.

articledeenadz

codex rust-v0.170.0 — 0.170.0

OpenAI's Codex CLI 0.170.0 adds WebRTC realtime voice in the TUI, auto-refreshing OpenTelemetry auth headers and experimental same-name MCP servers per environment.

Why it mattersCodex 0.170.0 adds realtime voice negotiation, experimental same-name MCP servers across environments, and fixes for sandbox restrictions being lost on child threads. Engineers running Codex with MCP or sandboxing should check these.

repogithub-actions[bot]

How Good Are Fable 5 2 And Gpt 6 2 Astra At Real Work

Epoch AI tested Fable 5.2, GPT-6.2 Astra and four other models on realistic internal work tasks.

Why it mattersShows where top-ranked models still fall short on open-ended work: experiment design, iteration, and judging what matters. Useful for deciding which tasks to delegate to agents and which need human review.

articleEpoch AI publications

js-genai v3.5.0

js-genai 3.5.0 adds experimental Live Interactions over WebSocket and WebRTC, self-hosted BYOS environments, an auth-token client.

Why it mattersJavaScript and TypeScript developers on Gemini get experimental Live Interactions (WebSocket and WebRTC), self-hosted environments, and a fix for batch inlined responses in the official SDK.

reporelease-please[bot]

python-genai v3.5.0

python-genai 3.5.0 adds experimental Live Interactions (WebSocket and WebRTC), self-hosted BYOS environments with system_instruction and custom_config.

Why it mattersThe Google GenAI SDK adds experimental Live Interactions over WebSocket and WebRTC plus self-hosted environments. Engineers building realtime or agent workloads on Gemini can now use these from the official client.

reporelease-please[bot]

litellm v1.110.0

litellm 1.110.0 ships signed Docker images and breaking fixes.

Why it mattersLiteLLM 1.110.0 marks breaking changes to Gemini audio handling and the Anthropic beta header allowlist. Teams running the proxy should check these before upgrading.

repolitellm-release[bot]

opencode v2.3.2

opencode 2.3.2 works around GPT-6.2 not caching prompts in the Responses instructions field by sending the system prompt as a developer message across OpenAI, Copilot, Azure, Zen and OpenRouter routes.

Why it mattersIf you call GPT-6.2 through the Responses API, a system prompt in the instructions field is re-billed as uncached input each step. Sending it as a leading developer message restores caching.

repogithub-actions[bot]
swyx 🗽 @aiDotEngineer@swyx

A heavy user of GPT-6.2 Astra reports it is the best workflow automation model he has used: fewer stuck loops, better learning through compactions and skill writing, and efficiency gains. He runs a content publishing pipeline on it.

Why it mattersA practitioner running production workflow automation reports Astra 6.2 avoids loops and learns across compactions and skill writing, a data point on long-horizon agent reliability when choosing a model.

GPT-6.2 Astra and Sol land in GitHub Copilot

GitHub makes OpenAI's GPT-6.2 Astra and GPT-6.2 Sol generally available in Copilot for agentic coding and code review.

Why it mattersGPT-6.2 Astra and Sol are selectable in Copilot at provider list pricing, available on Pro+ and above. GitHub reports early testing with more real issues caught in review, but gives no numbers.

articleGitHub
Vals AI@ValsAI

Vals AI reports GPT-6.2 Astra as #1 on Vibe Code Bench, a multi-turn coding eval. It avoided regressions, handled schema migrations better (88% vs 68%), and stayed within time limits. GPT-6.2 Sol ranks #3 at 40.94% with under a third of the cost.

Why it mattersVibe Code Bench tests repeated edits to an evolving codebase rather than single prompts.

Theo - t3.gg@theo

Theo argues Anthropic has led since Fable 5, noting Opus 5.5 spread those gains through lower price, generous limits and ZDR, while OpenAI kept pace on capability at the cost of long-run coherency and readability.

Why it mattersArgues model coherence on long runs and enterprise terms like ZDR and pricing, not benchmark parity, decide real adoption. This is a lens for choosing between Claude and OpenAI models for agent work.

openai-agents-python v0.25.0

openai-agents-python 0.25.0 removes the Runloop sandbox provider, exposes utilities for third-party session adapters, adds opt-in handoff error recovery.

Why it mattersThe Runloop sandbox provider is removed, so projects using it must switch. New session-adapter utilities and opt-in handoff error recovery help with custom persistence and resilient multi-agent flows.

repoopenai-sdks[bot]

opencode v2.3.1

opencode 2.3.1 backgrounds shell commands that exceed their timeout instead of killing them, writes compaction summaries in the background, adds GPT 6.2 Sol and Astra to the ChatGPT sign-in model picker.

Why it mattersLong builds no longer die at the shell timeout, and compaction no longer stalls the next prompt. GPT 6.2 Sol and Astra are selectable with ChatGPT sign-in.

repogithub-actions[bot]

Gemini 4 Neon Intelligence Gain

Artificial Analysis benchmarks Gemini 4 Neon: 61 on its Intelligence Index, $6.44 per task, and 59 on its Coding Agent Index in Antigravity CLI.

Why it mattersNeon costs about 17% less per task than Gemini 4 Carbon, and scores in line with Claude Code on Fable 5.2 on the coding agent index at roughly half the cost. This helps when picking a model for agents.

articleArtificial Analysis articles

How we fine-tuned a faster model for Patterns on Baseten

Datadog fine-tunes Qwen 3.5 9B on Baseten for its Patterns summarizer using production LLM spans as training data.

Why it mattersA narrow, high-volume LLM step cost about 80% of p99 cost. Fine-tuning a 9B model on traced production spans lifted judge pass rate from 45% to 93% and cut latency from about 45s to under 1s.

articlewww.datadoghq.com

Gemini 4 Neon now available on AI Gateway

Vercel adds Gemini 4 Neon to AI Gateway as google/gemini-4-neon.

Why it mattersYou can call Gemini 4 Neon through AI Gateway with the google/gemini-4-neon model ID. It has a 1M-token context, tool calling, structured outputs and four reasoning levels.

articleJerilyn Zheng

Rumble-2-Flash: an open 0.5B speech model for 40+ languages

Reson8 open-sources Rumble-2-Flash, a 0.5B speech recognition model for 40+ languages distilled from its larger model.

Why it mattersA small open ASR model runs in standard Transformers and vLLM with no custom code, and is reported to match much larger models on phone calls. Voice pipelines can self-host it cheaply.

articlehuggingface.co

How We Deploy Custom Chromium

Browserbase explains how agents patch, build and deploy a Chromium fork with 18,000 lines of custom patches across 40M lines of code in under a day.

Why it mattersBrowserbase shows how agents rebase a large patched Chromium fork and ship security fixes within a day of upstream release. It is a template for agent-driven maintenance of big forks.

articleBrowserbase

180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers

Hugging Face's Tarek Ziadé walks through running a 180B-parameter MoE model on a 32 GB Mac with Transformers expert offloading, compares how llama.cpp, vLLM and Transformers offload experts, and reports how close SSD streaming gets to Strata.

Why it mattersExperts make up over 90% of a MoE model's weights. Keeping hot experts resident and streaming the rest from SSD lets very large models run on laptop-class memory, with token generation reportedly within 6% of Strata.

articlehuggingface.co

Secure Agentic Apps

Shopify Engineering explains how its mobile team, shipping about 434,000 changed lines per week with agents, scales security using attack surface maps, focused AI reviews and regression checks derived from each fixed vulnerability.

Why it mattersShipping 11x more code with agents outpaces human security review. Shopify describes attack surface maps, focused AI reviews and checks that grow with each fixed vulnerability as a way to keep up.

articleShopify Engineering

How we built Datadog’s agentic vulnerability hunting system

Datadog describes an orchestration pipeline for agentic vulnerability hunts built on threat packs, which are reviewed summaries of a system and its security requirements.

Why it mattersAgents find candidate vulnerabilities easily but miss auth and network controls owned by other teams.

articlewww.datadoghq.com

claude-code v2.1.327

Claude Code v2.1.327 adds stdin input for claude mcp add-json, a richer /status diagnostics list, an attributionProject billing setting, and /tui classic.

Why it mattersClaude Code 2.1.327 lets you add MCP servers from stdin to keep tokens off the command line. It also expands /status diagnostics and fixes bugs that deleted .mcp.json keys and git-ignored files.

repoclaude-code-publisher[bot]

claude-agent-sdk-typescript v0.3.327

Claude Agent SDK TypeScript 0.3.327 adds breaking validation for query options and hook arrays, deprecates SDK MCP servers nested in agent configs, lengthens MCP reconnect retries to about a day.

Why it mattersUpgrading the Claude Agent SDK can break code: unserializable options and non-array hook values are now refused, and SDK MCP servers inside an agent's mcpServers are ignored. Move them to top-level mcpServers.

repoclaude-code-publisher[bot]

goose v1.60.0

goose 1.60.0 replaces the Skills extension with a SkillOperation, adds executable, slash, bang and steering operations to oi-kit, ships Cortecs and Onysoft providers.

Why it mattersGoose removes its Skills extension in favor of a SkillOperation and adds steering and slash-command operations, which changes how skills and commands are wired in the agent.

repogithub-actions[bot]
Peter Yang@petergyang

Peter Yang's interview notes from Microsoft Ignite with Satya Nadella: agents with token budgets as the unit of work, business apps as headless plugins, model-independent trust architecture, and using agent traces to fine-tune across model upgrades. Lightly edited quotes, not a full transcript.

Why it mattersMicrosoft's CEO says to log agent traces, fine-tune on them for each new model, and keep trust controls independent of the model. It is a concrete lock-in and safety posture for teams building agents on third-party models.

External sandboxes for coding agents

Simon Willison and Prime Radiant describe an external sandbox pattern where the coding agent stays outside and its code runs in an isolated VM.

Why it mattersRunning agent-executed code in a separate sandbox limits damage from mistakes or prompt injection, including when trying local models. Three implementations share one localhost API, so Pi, Claude Code or Codex can be pointed at any of them.

articlesimonwillison.net
OpenAIDevs@OpenAIDevs

OpenAI updated the API dashboard's Analytics view. It now shows usage and spend, rate limits and rate-limited requests, and errors broken down by model, project, or API key, giving teams a faster way to investigate issues.

Why it mattersIf you run production workloads on the OpenAI API, you can now check rate limits and rate-limited requests, and break down errors by model, project, or API key, which speeds up diagnosing usage spikes and failures.

pydantic-ai v2.70.0 — v2.70.0 (2026-11-17)

pydantic-ai v2.70.0 discloses two high-severity ReDoS advisories in Pydantic AI Harness: ToolOutputLimits with strip_ansi and the default jwt secret-redaction pattern.

Why it mattersIf you use strip_ansi or the default jwt redaction on text an attacker can influence, a few hundred KB can block the event loop for minutes. Upgrade pydantic-ai-harness to 0.70.0.

repodsfaccini

GPT‑6.2 Astra and GPT‑6.2 Sol

Simon Willison covers OpenAI's GPT-6.2 Astra and Sol API launch.

Why it mattersGPT-6.2 Astra costs $8/$40 per MTok with cached input at $0.40, cutting long agentic session costs, and Sol stays at $2/$10. Practitioner tests show both trail Claude Fable 5.2 on a hard 3D coding task.

articlesimonwillison.net
ClaudeDevs@ClaudeDevs

Anthropic added an opt-in Latest release channel for Claude Desktop on Windows and Mac, set under Settings > Desktop app. It delivers fixes and improvements sooner but may be less reliable, while Stable remains the default channel.

Why it mattersEngineers who rely on Claude Desktop can now choose between faster fixes and features (Latest) or the more reliable default (Stable). Latest builds may be less stable, so weigh that before switching a work machine.

Perplexity@perplexity_ai

Perplexity reports GPT-6.2 Astra and Sol now in Search and Computer, with WANDR benchmark scores of 0.783 at about $12.54 per task for Astra and 0.746 at about $2.65 for Sol, up 0.095 and 0.065 on predecessors. Vendor-run benchmark; the thread may be incomplete.

Why it mattersPerplexity's own WANDR results put GPT-6.2 Astra at 0.783 for about $12.54 per task and Sol at 0.746 for about $2.65. Engineers choosing between tiers get a concrete score-per-dollar comparison, though from a single vendor benchmark.

MAI Code 2 and Flash reach GitHub Copilot

GitHub makes Microsoft's MAI Code 2 and MAI Code 2 Flash generally available in Copilot, aimed at agentic coding.

Why it mattersMicrosoft's MAI Code 2 and MAI Code 2 Flash are selectable in Copilot's VS Code, CLI, cloud agent and JetBrains pickers under usage-based billing. Admins control access through model policy.

articleGitHub

DeepsecBench: Vercel's Unseen-Codebase Cyber Leaderboard

Vercel's DeepsecBench ranks 60 model and effort configurations on vulnerability discovery, reporting score, recall, precision, false positives, cost and runtime per harness.

Why it mattersShows how 60 model and effort configurations compare on finding vulnerabilities, including recall, false positives and dollar cost per run. Helps you pick a model and effort level for security review agents.

articleGuillermo Rauch
Epoch AI@EpochAIResearch

Epoch AI tests recent frontier models on open-ended Epoch work tasks, graded against employee standards. Fable 5.2 leads at 82%, GPT-6.2 Astra trails at 65%. Common failures: weak audience judgment, passing off existing ideas as novel, and biased self-grading. Based on the thread only.

Why it mattersEpoch's open-ended work benchmark shows no frontier model reaches employee standard: Fable 5.2 averages 82%, GPT-6.2 Astra 65%. Failures in judgment and self-assessment show where agents still need human review.

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