Latest news, beside the leaderboard

On November 14 Artificial Analysis released Coding Agent Index v2.1, swapping FrontierCode 1.1 for its implementation of Cognition's FrontierCode 2.0 (132 real-world tasks graded on whether a maintainer would merge the change), upgrading Terminal-Bench from 4.0 to 4.1 and keeping SWE-Atlas-QnA-AA. Claude Fable 5.2 (xhigh and high, in Claude Code) and Claude Opus 5.5 (max) share the lead at 58. Because FrontierCode 2.0 is harder, scores fall 5 points on average; only Codex with GPT-6.1 Terra (max) held within 1 point of its prior score.

On November 13 Artificial Analysis said Speechify's Simba 3.5 text to speech model is #1 in its Controlled Voice Arena, statistically level with HeyGen Voice. Every model on this board reads with the same 8 cloned voices (4 US, 4 UK) and users vote blind, so it compares models rather than showcase voices. The leaderboard lists Simba 3.5 at 1192±17 Elo from 1,783 samples, HeyGen Voice at 1190±13 and Qwen-Audio-3.1-TTS-Plus third at 1176; Simba 3.5 costs $6.58 per 1M characters on the API versus $30.00 for HeyGen Voice.

On November 13 Artificial Analysis published the Diarization Index, a benchmark for whether speech to text APIs attribute each word to the person who actually said it. It scores Word Diarization Error Rate (WDER), the share of words given to the wrong speaker or to no speaker, across 97 English conversations of two to eight people, averaging three speaker bands (2, 3-4 and 5-8) with equal weight. NVIDIA's open-weight Nemotron 3.5 Transcribe leads at 3.3%, ahead of Voxtral Small Transcribe 3 at 7.5%, the most accurate priced API; MAI-Transcribe-2 scores 8.8% and is one of the fastest.

On November 13 the FBI Cyber Branch published FLASH-20261113-001 on ARTEX, an open-source, LLM-driven multi-agent penetration testing platform. Since at least August 2026, various threat actors have used it to automate intrusion phases from reconnaissance to data exfiltration against financial, government, technology, defense, semiconductor, energy, healthcare and other sectors, with some intrusions going from targeting to privileged access in under three hours. The original developer pulled the GitHub repo in October, but forks, mirrors and a commercial ARTEX PRO keep it available, and researchers found more than 1,600 internet-exposed ARTEX consoles in early November.

SpaceXSI released Grok Bot 1.0, and the official @bot account then announced that Bots can now join meetings: paste a Google Meet, Zoom or Microsoft Teams link, approve the card, and the Bot joins as a transcribing participant, takes notes, answers when addressed in the meeting chat, and sends a summary and action items after the call. Meetings are on Team, Enterprise and Ultra plans with an Enterprise admin switch; the same release brings Team Bots into Microsoft Teams and adds shared connectors and review of team learnings.

On November 13 OpenAI announced ChatGPT Wallet, built with American Express, Mastercard and Visa, which lets dots agents complete purchases after the user approves them. At payment time the card network issues a token in place of the real card number, so neither the business nor the agent sees it; before each purchase the user reviews the items, total and card and confirms with Face ID or a fingerprint. It is rolling out now in dots to US ChatGPT Pro users aged 18+, alongside a Commerce Partner Playbook for businesses.

A November 13 LMSYS blog post from the RadixArk Miles team says its open-source RL post-training framework Miles trained GLM-4.7-Flash end to end with every agent sandbox on NVIDIA Vera CPUs. With the machines already busy, episodes took 12% longer on average when sandboxes ran on AMD EPYC Turin, and within a fixed latency bar one Vera socket carried 1.3-1.5x the load of Xeon 6 and 1.4-2.1x that of Turin.

Vercel released Workflow SDK 5. It says the SDK has run over 4 billion steps and 700 million runs; v5 rewrites the engine to cut queue hops, taking a 1,020-step run from a 21-minute to a 3-minute median. Step payloads are encrypted with AES-256-GCM using a per-run key, Workflow infrastructure now runs in every Vercel Function region, and the same durability layer backs harnesses such as Pi and OpenCode.

Arena studied 105 models across 44 weekly Code Arena: WebDev boards since December 2025. Xiaomi's MiMo-V2.6-Flash now beats April's leader Claude Opus 4.7 at under a twentieth of the price, the top score has risen about 41 points a month since May (3x the winter pace), and 19 of the 25 price-performance frontier models under $1 per answer are Chinese.

On November 13, Google DeepMind released the two datasets behind its WeatherNext 3 model. SatWeave stitches imagery from geostationary satellites such as GOES, Meteosat and Himawari into one global mosaic across 17 visible, water vapor and infrared channels, updated hourly, with an archive spanning more than a decade. PARDIG is an AI-generated global precipitation dataset trained on NASA GPM spaceborne radar, with 15-minute time steps, roughly 10 km resolution and about five hours of average latency. DeepMind says that, evaluated against radar networks in the US, Europe and Japan and more than 36,000 weather stations, it improves hourly detection of light-to-moderate rain by more than 30% over NASA IMERG-Final, with archives from 1999 to 2026. Live feeds and archives for both are on Google Cloud Storage, and PARDIG historical catalogs are also in Earth Engine.

On November 12, IBM's Granite team published LLMON (LLM Object Notation, pronounced "lemon") on Hugging Face. It wraps semantically tagged blocks such as instructions, data and requirements in reserved special-token delimiters, gives each block a unique ID, and uses an exec block to state which instruction runs on which data. IBM also released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview (Apache 2.0), plus a GitHub toolkit and lossless JSON-to-LLMON converters. IBM says that on the OPI indirect prompt injection benchmark the LLMON models avoided 47% more attacks than the best frontier model it tested and 95% more than the best open-source model, while general capability stays on par with Granite 4.1 Instruct. IBM stresses this is learned behavior, not an enforced guarantee, and the models are an experimental preview, not for production.

On November 13, DeepSeek said the capabilities behind the DeepSeek App are now available through its API. V4.1-Flash (model name deepseek-flash) supports server-side web search (including image search with citations), deep thinking search, file input for PDF, Office and image files, automatic reading of links in user messages, and input/output safety checks across the Messages, Responses and Chat Completions protocols. OCR, speech-to-text (ASR) and text-to-speech (TTS) also launch as standalone APIs. Tool calls are billed on top of token charges: Search and Fetch cost $0.75 per 1K requests, OCR $0.225 per 1K pages, ASR $0.058 per hour and TTS $0.225 per 10K characters (both at a limited-time 50% discount). DeepSeek also open-sourced an MIT-licensed chat front end that can import DeepSeek chat history. Public-facing apps must enable safety checks.

On November 13, Cognition launched Devin Apps. You describe the tool you need, Devin builds it as a web app in a session and serves it from its own cloud machine, with nothing deployed to a third-party host. Each App has a stable URL (name--id.preview.devinapps.com); everyone who opens it must sign in to Devin, Apps are never public, and they can be shared with people, an organization or a whole enterprise. Apps can use any stack, reuse connected MCPs, integrations and code, and keep data such as SQLite or Postgres on the same machine. The machine sleeps 5 minutes after the last request and wakes when someone opens the App; billing is by awake time, while building and changes bill as normal session work. Every update is a new version tied to a commit, and if a new version fails to start the previous one keeps serving. Apps are rolling out gradually.

On November 12 (UTC), Simular, the computer-use agent startup backed by Felicis and NVIDIA, launched box: one cloud Ubuntu desktop per account with Chrome and a persistent shell, exposed as a single remote MCP server (https://mcp.box.simular.ai/mcp, Streamable HTTP, OAuth sign-in). Any MCP client can add it, including claude.ai, ChatGPT, Gemini, Claude Code, Codex, Cursor, OpenCode and VS Code, so different assistants can work on the same machine with shared logins and files. Users watch it live on the web or iPhone and take over to sign in or approve payments. The desktop itself is free; model calls use account credits. The box sleeps when idle, wakes in about a second, and is deleted after 14 straight days asleep. The launch video passed 810K views. Simular has not published a benchmark for its "fastest computer use" claim.

On November 13 at 18:12 Singapore time, AI commentator Chubby (@kimmonismus, about 159K followers, editor-in-chief of Superintel) posted that after a year of switching between Codex/GPT and Claude Code/Opus, Claude Code with Opus 5.5 has been his daily driver for two months: "the most efficient and intelligent model to date," with GPT "good, but not close." He added that Claude's weak spot is computer use, where GPT is "unbeaten," and that he cares more about Claude Duet than ChatGPT dots, though neither is available to him in Europe. In a reply he put his split at about 80% Claude. At writing time the post had about 29K views, 590 likes and 152 replies. It went up during an Anthropic incident that blocked sign-ins to claude.ai and Claude Code from 09:58 to 11:58 UTC. This is one user's opinion and community discussion, not a benchmark.

Continuous Intelligence, a London lab founded in July, released its first time-series foundation model, ci-proto, with weights on Hugging Face and code on GitHub. It forecasts series such as power load, sales or sensor data zero-shot on CPU or GPU. The company says it reads more than 2 million steps of history, 64 times other time-series foundation models, and returns nine quantiles from 0.1 to 0.9. Its results on the 100 fev-bench tasks were merged into the official fev repo by an AutoGluon maintainer: 51.51 SQL skill and 40.80 MASE skill. The licence builds on Apache 2.0 but adds attribution and a one-time notice for external products.

Anthropic put self-hosted Claude Managed Agents into private beta. Until now teams could either let Anthropic run everything or host only the sandboxes. Now the whole service (API, agent loop, session storage and sandboxes) can run in your own cloud account or on-premises, with session history in your own PostgreSQL. Model calls can go to Anthropic, Amazon Bedrock, Claude on Google Cloud, Microsoft Foundry or your own gateway, and the developer API matches the hosted service. Access is limited to eligible organizations through the Anthropic account team.

At NODES 2026, Neo4j launched Relate, a new open-source embedded graph database written in Rust (Apache 2.0). It starts and stops with your application, with no server to deploy and no account to create. Agents can store what they learn as a connected graph and use one Cypher statement to follow relationships and run vector and full-text search together. It ships with a coding-agent skill and an MCP endpoint, and switches to a token-saving output mode when it runs under a coding agent. v1.0.1 is out on GitHub, but the README says it is still an experimental Neo4j Labs project outside the supported product line.

LMSYS published a technical deep dive on a megakernel written from scratch in CUDA C++ and PTX (about 8,000 lines) for Meta's open-weight Muse Spark Open. Each decoder layer runs as a single kernel launch that overlaps compute, memory loads and cross-GPU communication. On 8x GB300, single-request decoding goes from 517 tok/s on the SGLang baseline to 1,109 tok/s, about 2.1x, and each decode step drops from 12.9 ms to 5.9 ms. The code is open source as an SGLang plugin at facebookresearch/sgl-muse-mega (MIT). It only covers batch size 1 and is experimental, so it buys latency rather than cheaper tokens.

On November 12 Lightning Labs released pi402, an extension for the Pi coding agent with an embedded Wavelength Lightning wallet. When the agent calls a paid inference API, it pays automatically over L402 or MPP, and the proof of payment doubles as the API key, so there is no sign-up and no credit card. Sellers put Aperture v0.6.0 in front of an existing inference API and price per token with meterd or lightmeter. The whole stack is open source (MIT). Lightning Labs' example rate of 1 msat per input token and 2 msat per output token works out to about $0.80/$1.61 per million tokens at a bitcoin price of about $80,500.