Intel
Page 09
Is Jev a Good Multilingual Judge?
Compares the Jev judge model with GPT-6.1 on four human-annotated multilingual benchmarks.
Why it mattersA cheap judge can match a frontier judge on most multilingual evaluation tasks, but it wrongly accepts more incorrect answers in low-resource languages. Use its probabilities for confidence-based routing.

Together Compiler: Continual Compilation for AI Inference
Together AI presents TKCC, a continual agent loop that recompiles inference deployments as models, architectures and traffic change.
Why it mattersShows how an agent hill-climbing loop can keep inference deployments optimized as models and traffic shift, avoiding both low-fidelity kernel benchmarks and slow end-to-end engine launches.

Abstention Is Not Threshold Compliance: Stated Scoring Rules in Open Language Models
Tests stated scoring-rule thresholds on 12 open 3-8B checkpoints.
Why it mattersTelling a model a higher confidence threshold in the prompt barely changes its abstention behavior. Rely on the output format, or an external threshold on calibrated confidence, rather than a stated rule.

Leaderboard Hacking: Preference-Based Model Evaluations are Vulnerable to Manipulation
Researchers show that Arena-style, Bradley-Terry preference leaderboards are vulnerable to strategic voting, strategic prompting and strategic nomination, demonstrate the effects on public leaderboard data, and outline defenses.
Why it mattersPreference-based model rankings can be pushed up by strategic votes, prompt choice and inserting irrelevant models. If you pick models from Arena-style leaderboards, treat rank differences with caution and look for the proposed safeguards.

Cleaning an Agent's History Hurts When the World Still Holds the Error
Freezes tool-using agents mid-task and edits history or world state on AgentCE-Bench.
Why it mattersIf you compact or hand off agent histories, how the correction is phrased matters: a note beside the preserved error helps, while rewriting it as a completed action hurts recovery when the world still holds the error.

codex rust-v0.168.0 — 0.168.0
Codex 0.168.0 adds a model provider browser with per-chat defaults, an experimental side-conversation workspace, audio attachments, animated GIF generation, OSC 777 notifications and fixes to history, resume and steering behavior.
Why it mattersCodex 0.168.0 adds a side-by-side workspace pane, audio recording attachments for audio-capable models, and per-session model provider handling. These change how engineers run and steer long coding sessions.

Brave Search API now accepts agentic payments via the x402 protocol
Brave's Search API now supports pay-per-request agentic payments through the x402 protocol.
Why it mattersAgents can now pay for Brave Search per query with USDC over HTTP 402, with no signup or key management. That removes a credential step when an autonomous agent needs web grounding.

Vercel Sandbox starts up to 4x faster
Vercel Sandbox now serves reused managed and custom image data from regional caches, cutting startup time about 75% in Paris and from roughly 3.5s to 1.05s in Seoul.
Why it mattersIf you run agent code in Vercel Sandbox outside iad1, cold starts drop by up to 4x on cache hits, because image data is read from a regional cache rather than origin storage in Washington, D.C.

Parallel Decisions now available on AI Gateway for free
Vercel added Parallel Decisions Fast to AI Gateway, free through November 23, 2026.
Why it mattersYou can ask typed, web-grounded questions and get probabilities with cited evidence through AI Gateway without a separate Parallel account. Use it for lead routing, CRM enrichment or claim checks while it is free.

Vercel Security Review no longer requires a linked Vercel project
Vercel Security Review can now scan pushes in GitHub repositories with no linked Vercel project.
Why it mattersYou can get automated vulnerability review on every push to repositories that are not deployed on Vercel, enabled in Agent settings, without creating a project first.

Introducing Owe Bench
Mercor releases OWE-Bench, 26 long-horizon engineering tasks across software, physical engineering, ML and scientific computing, graded continuously on unseen cases.
Why it mattersOWE-Bench scores agents on 26 multi-hour engineering tasks with continuous grading. Top score is 54.6% at about 211M tokens per attempt, showing long-horizon engineering is unsolved and costly.

Edge0-145B: Frontier AI Capabilities on Your Own Computer
Edge0 open-sources its 145B and 120B on-device models along with a streaming MoE inference framework and deployment tools.
Why it mattersA 145B open model that reportedly runs on a laptop at 29GB peak memory and 80+ tokens/s makes local coding and agent workloads without per-token cloud fees plausible. The benchmarks are self-reported, so verify them on your hardware.

llm-gemini 0.37
Simon Willison breaks down the Gemini 4 family pricing, including the scheduled December 1 price doubling for Carbon and Argon.
Why it mattersGemini 4 Carbon and Argon cost $2/$10 per million tokens until Nov 30, then double. Hydrogen is $1.50/$7.50, about 15x the input price of Haiku 5.5 and GPT-6 Luna, which affects model choice.

Build once, run on any agent: A practical guide to agent skills
GitHub walks through building agent skills.
Why it mattersShows how to package repeatable tasks as versioned SKILL.md directories that run across agents, and why skills describe the work but do not grant tool access or credentials.

claude-code v2.1.323
Claude Code 2.1.323 adds an experimental Code Mode where Claude writes JavaScript to call tools, engine version constraints for plugins, TypeScript hook type checking in plugin validate.
Why it mattersCode Mode lets Claude call its tools by writing JavaScript, an opt-in change to how the agent acts.

claude-agent-sdk-typescript v0.3.323
Release notes for the Claude Agent SDK for TypeScript v0.3.323.
Why it mattersSDK hosts get machinePolicy.didNotLoad and matchDeniedModels() for policy handling, plus a tool_progress heartbeat during long hooks. Compaction forking and MCP progress fixes remove bugs in session handling.

Primer: Wharton's open multi-agent system for teacher-built simulations
Wharton's Generative AI Lab open-sources Primer, an agent pipeline that interviews teachers and produces playable simulations.
Why it mattersShows a concrete multi-agent authoring pipeline with human approval gates and inspectable artifacts. Agent builders can reuse the orchestration pattern, prompts, and the paper-to-agent build prompt.

Chinese frontier AI models carry the fingerprint of export-controlled GPUs
A forensic report finds six recent models from DeepSeek, Moonshot, MiniMax and Ant Group carry numerical patterns matching Blackwell GPUs, while earlier ones match Hopper.
Why it mattersShows a method for inferring training hardware from released weights and claims several Chinese labs trained recent models on export-controlled Blackwell GPUs.

Foundation Models framework
Apple's Foundation Models framework is a native Swift API for integrating LLMs.
Why it mattersSwift developers can call Apple's on-device and Private Cloud Compute models with guided generation, tool calling and image input, building private LLM features without a third-party API.

Kimi K3.1 Pro is now available in Token Pass
Vercel adds Moonshot AI's Kimi K3.1 Pro to Token Pass.
Why it mattersKimi K3.1 Pro, a 1M-context model with tool use and vision, can now be selected in coding agents connected to Vercel AI Gateway under Token Pass, with zero data retention on requests.

ai ai@7.0.171
AI SDK 7.0.171 caps validation cost for malformed UI messages and adds a fetch option to createDownload, including opt-in guarded fetch with DNS validation, abort-signal passthrough.
Why it mattersIf your app accepts UI messages from clients, upgrade: malformed parts could previously exhaust memory. File downloads can now use a DNS-validating guarded fetch for untrusted URLs.
Anthropic introduces Claude Meetings: Claude joins a Google Meet call as a visible participant and builds docs and whiteboards while the team talks. It is in beta on Team and Enterprise plans. Notes, action items and transcripts land on the meeting page. No technical detail is given.
Why it mattersClaude can now act as a live participant that produces artifacts during a call, extending agent use beyond chat and code. Teams on Team or Enterprise plans can try it on Google Meet now.

Introducing Claude Meetings
Anthropic introduces Claude Meetings in beta for Team and Enterprise plans.
Why it mattersClaude can now join a team call and build documents live while people talk, then keep notes and artifacts on a meeting page. It is in beta on Team and Enterprise plans, Google Meet first.

python-genai v3.3.0
Google's Python GenAI SDK 3.3.0 adds an XHIGH thinking level, the gemini-4-hydrogen model enum in the interactions API, new FinishReason values.
Why it mattersThe Python Gemini SDK now exposes an xhigh thinking level and the gemini-4-hydrogen model ID. Engineers can select the higher reasoning setting and the new model directly from the client.
Anthropic engineers on caching, effort and cost per task
Two Anthropic API engineers explain how models launch alongside features.
Why it mattersChanging effort or system prompt mid-conversation normally breaks prompt caching. Being able to change effort without that loss, and judging upgrades by cost per task rather than per token, directly affects agent run costs.
GitHub walks through an agentic engineering workflow across plan, build, verify, and observe, with Copilot features for each stage. It frames review load as the new bottleneck, citing 1.2B weekly commits and 441% longer PR reviews at high AI adoption.
Why it mattersAs agents speed up code generation, review becomes the bottleneck. The guide shows how stacked PRs, automated review, and agent-driven fixes address it, naming the Copilot features for each stage.

LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B
IBM introduces LLMON, an LLM-native markup that tags context components with reserved tokens and IDs.
Why it mattersExplicit instruction/data separation in the context window is a model-level defense against indirect prompt injection. Granite 4.1 3B and 8B preview models and a toolkit let you test it.

Gemini 4 Hydrogen and Orchestrator
Google Antigravity adds Gemini 4 Hydrogen and a new Orchestrator model option that combines Hydrogen with stronger models such as Gemini 4 Carbon for long-horizon tasks, balancing quality, efficiency and latency.
Why it mattersAntigravity users can now pick Orchestrator, which pairs the cheaper Hydrogen model with Gemini 4 Carbon for long-horizon work, trading off quality, efficiency and latency.
Closing The Gap
Runway explains how it replaced third-party classifiers that scored fixed categories like blood or weapons with LLM-based moderation built on its own policy.
Why it mattersFixed-category classifiers could not distinguish fiction from real harm, forcing over-blocking. LLM-based moderation lets a product write its own policy and revise it in hours, a pattern applicable to any generative product.
Cognition adds Gemini 4 Hydrogen to Devin Desktop and Devin CLI, reporting 38.2% on its FrontierCode 2.0 benchmark of mergeable real-world engineering tasks, just behind GPT-6.1 Sol. Benchmark details were not collected.
Why it mattersDevin users can now select Gemini 4 Hydrogen. Cognition's own FrontierCode 2.0 puts it at 38.2%, just behind GPT-6.1 Sol, a vendor-run signal on mergeable engineering work.
Cursor adds Gemini 4 Hydrogen, reporting 54.3% on CursorBench at default effort (up from 39.6% for Gemini 3.8 Flash) at $1.50/$7.50 per million tokens. Effort-level comparisons are linked but were not collected.
Why it mattersCursor users can now pick Gemini 4 Hydrogen at $1.50/$7.50 per million tokens. Cursor reports 54.3% on CursorBench at default effort, up from 39.6% for Gemini 3.8 Flash.
World Labs opens general access to its Atlas world model. Viewfinder builds worlds from one to 64 images and exports images, video or Gaussian splats, while Chisel uses agentic workflows for 3D geometry. All endpoints behind the apps are exposed via API for developers and agents.
Why it mattersAtlas world-generation endpoints are open and can be called directly by agents or your own tools. Developers get keyframes, video and Gaussian splats from image or text input without a waitlist.
Artificial Analysis benchmarks Gemini 4 Hydrogen: 54 on its Intelligence Index at xhigh, 8% hallucination rate, strong visual reasoning, but 54% on Terminal-Bench 4.0 versus 64% for Sonnet 5.5. Cost per task ranges from $0.29 at low to $3.16 at xhigh.
Why it mattersHydrogen nearly matches Sonnet 5.5 and GPT-6.1 Sol on intelligence at about a third of Sonnet's cost, but trails on terminal-based agentic coding. The high setting gives Argon-level scores for far less than xhigh.

Nvidia and the World Without Jensen
SemiAnalysis examines Jensen Huang's departure as Nvidia CEO, citing its Accelerator Model data that over 1.5GW of restricted Nvidia chips reached China.
Why it mattersExport-control tightening on GPUs and remote access for Chinese labs could change chip availability and compute pricing.
Google adds Gemini 4 Hydrogen to the Gemini 4 series, aimed at fast everyday coding and agentic work. It claims 2.5x the speed of Gemini 3.8 Flash and tunable thinking levels. It is available in the Gemini API, AI Studio, Android Studio and Antigravity.
Why it mattersA faster Gemini tier with adjustable thinking is available today in the API and in Antigravity for coding and agent loops. Engineers choosing a model for latency-sensitive agent work have a new option to evaluate.

Introducing Gemini 4 Hydrogen
Google DeepMind releases Gemini 4 Hydrogen, a faster tier for coding and agentic work.
Why it mattersA new Gemini 4 tier targets fast coding and agentic workflows at flat $1.50/$7.50 per million tokens across 1M context. It reportedly beats Argon on the Artificial Analysis index, which changes the cost-to-capability tradeoff.

Introducing Gemini 4 Hydrogen
Google DeepMind releases Gemini 4 Hydrogen, a fast tier for coding and agentic work.
Why it mattersA fast Gemini 4 tier now has flat 1M-context pricing of $1.50 per million input and $7.50 per million output tokens. It scores above Argon on one index, making it a cheaper option for coding agents.
Anthropic puts self-hosted Claude Managed Agents into private beta. Teams can run the agent harness, sandboxes and session history in their own cloud or on-premises, and choose where requests to Claude are routed.
Why it mattersTeams with data residency or on-prem requirements can run the agent harness, sandboxes and session history in their own environment while still using Claude. This removes a common blocker for adopting managed agents.

Building the Claude API for the Frontier
Two Claude Platform engineers discuss how API features are chosen and retired, model migrations, cost management, changing effort mid-conversation while preserving prompt caching.
Why it mattersCovers how to change effort mid-conversation without invalidating prompt caching, plus cost and migration practices. Also explains that Mythos and Fable share weights with different safeguards, which affects access planning.

Fish Audio launches Fish Agent on Live 3 full-duplex voice
Fish Audio launches Fish Agent, a no-code builder for voice agents that take and place calls, powered by its Live 3 full-duplex model.
Why it mattersFull-duplex models handle turn-taking inside one model instead of a speech-to-text, LLM and text-to-speech chain. Voice agent builders get a platform with pricing, tooling and testing to compare against cascaded stacks.

browser-use 1.0.3 — browser-use 1.0.3
browser-use 1.0.3 changes how the CLI finds a browser.
Why it mattersThe browser-use CLI now launches its own persistent Chromium when no debuggable browser is available, and a bad remote-debugging URL fails in 3 seconds instead of 30.
Salesforce's SOV-Transcribe-1 tops open-weights speech-to-text on the AA-WER Index at 2.3% with 4.6B active parameters (26.6B total), ranking #7 of 44 overall. It also hits 1.8% on voice-agent conversations and runs at 60x real time on Salesforce's endpoint.
Why it mattersAn open-weights transcription model beats larger open peers (2.3% vs 2.8% WER) with about 5x fewer active parameters than Voxtral Small. This matters for self-hosted voice-agent pipelines.

What Huawei’s Horinger data center reveals about China’s compute gap
Epoch AI uses satellite imagery and public documents to size Huawei's Horinger AI campus at up to 286 MW of IT power.
Why it mattersSatellite-based estimates show Huawei's largest AI campus holds under a tenth of the capacity of the biggest US data center, and likely stays far behind until at least 2028. This grounds expectations about China's frontier compute.

Workflow SDK 5: built for production
Vercel ships Workflow SDK 5, the durable async JavaScript runtime for automations and long-running agents, citing 4 billion steps and 700 million runs.
Why it mattersLong-running agents need durability across crashes and deployments, and the event log holds sensitive step data. Version 5 targets security, reliability and performance for exactly these workloads.
The Jev Revolution: What Engineers Are Already Building with System One Models
Reporting on the AI Engineer Code Summit keynote about system one models, which return scores or choices instead of generated text.
Why it mattersIt argues for replacing string-concatenation prompting with fast classifier-style model calls inside application code. That offers a cheaper, more reliable alternative to reasoning LLMs for decision steps.
WTF are evals? How to test agents and the software they use
Vercel lays out how to eval agents and the APIs, CLIs and MCP servers they call.
Why it mattersAgents can complete a request that looks successful in logs but is wrong (hot latte instead of iced). Repeated, graded evals show how often a fix works, which one manual test cannot.
SemiAnalysis reports that NVIDIA and Inferact submitted a vLLM Rubin NVL72 result for MiniMax M3.1 Flash showing 6.5x better perf per TCO than GB300 NVL72, using CUDA 13.4's locality-aware MoE feature, which adds about 1.2x realized HBM bandwidth.
Why it mattersInference cost on next-generation hardware may drop sharply: a reported 6.5x perf per TCO gain for Rubin NVL72 over GB300 on a MoE model with vLLM. It affects future token pricing and availability, but the submission details were not read.

Rethinking Claude Code Best Practices — Lydia Hallie, Anthropic
Anthropic's Lydia Hallie argues that much of the scaffolding added to Claude Code setups over the past year now limits the Claude 5.5 models.
Why it mattersSetups built for earlier models can now hold newer ones back. This talk from an Anthropic Claude Code developer-relations engineer says which habits to retain and which to drop for Claude 5.5.

pydantic-ai v2.68.0 — v2.68.0 (2026-11-11)
pydantic-ai 2.68.0 adds ModelFallback on max_tokens truncation, RequestInput for approvals during a run, and durable prompt-injection classification on Temporal.
Why it mattersAgent builders get automatic model fallback on truncated output and in-run approval requests. Several compatibility changes, such as renamed sub-agent events and changed frame-text handling, may need code updates when upgrading.

Vercel Sandbox now supports managed npm credentials
Vercel Sandbox adds managed npm credentials via a credentials option.
Why it mattersAgents running in Vercel Sandbox can install private npm packages without ever holding the tokens, because the firewall adds authentication to registry requests. This removes a common credential-leak path for coding agents.
An index of the vibe-coding frontier. Corrections welcome.