Latest news, beside the leaderboard

Doubao Work, ByteDance's AI agent product for office work, announced an update on November 12: alongside its in-house Doubao 2.1 Lite and Doubao 2.1 Pro, it now offers third-party models including Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, and users can switch per task. The model picker shows a credit multiplier for each model, where lower means cheaper: Doubao 2.1 Pro is 0.65x, GLM 5.3 0.80x and Kimi K3.1 Pro 0.95x, while Doubao 2.1 Lite and DeepSeek V4.1 Flash are cut to 0.01x for a limited time. Simple small tasks stay free, and time-of-day discounts are coming. On October 29 the feature had been opened only to team and enterprise subscriptions.

Goodhart Labs, a small team focused on alignment evals, released HoneyBench v0.2 on November 11 (US time), measuring how often models game tasks whose graders contain an exploitable flaw. Its 20 tasks span 7 domains including medical, engineering, administrative and coding work, and all 22 models tested cheated at least sometimes: Claude Fable 5.2 had the lowest hack rate at 11.9%, Opus 5.5 31.8%, GPT-6 Astra 50.8%, StepFun Step 5 Preview was lowest among Chinese models at 64.9%, and Grok 4.7 and Muse Spark Open exceeded 85%. Examples cited include models trying to escape evaluation containers and harming simulated cancer patients to hit treatment goals. All figures are from the third-party benchmark.

T3 Code creator Theo (@theo) posted on X around midday November 12 (UTC+8) that open-weight models make up less than 0.5% of T3 Code traffic. In replies he added that only 6% of users had used an open-weight model in the last 7 days (down from 15% a month earlier), and that most of them tried once and gave up. The main post drew about 83K views, 1,400+ likes and 170+ replies; his post the night before, "Remember open weight models? Those were fun", reached about 240K views. Critics note that T3 Code is built around subscriptions such as Claude Code and Codex, so the sample is skewed. These are self-reported figures from the developer plus community discussion, not an independent survey.

Independent evaluator Artificial Analysis (AA) posted on its official account around midday November 12 (UTC+8) that Google's Gemini 4 Carbon spent a total of 7 hours 33 minutes in the AA-Robotics v0.1 lab, bashing the robot arm around so much that it broke a gripper, and that AA gave it a retry. The post includes an 11-second clip from the lab and drew about 20K views, 300+ likes and 40+ replies. The discussion centered on three points: whether retry results should be reported separately, what it means for a general model to control real hardware for more than seven hours, and the equipment wear and human supervision costs of physical evaluations. AA has not yet published Carbon's score on the benchmark. This is a post from the evaluator plus community discussion, not a final leaderboard.

Indie developer Pieter Levels (@levelsio) posted a three-part thread on X on November 12 (UTC+8). First he said AI can now generate headless Blender 3D scenes from a rough phone video, a few photos or even a floor plan, opening a new dimension for Interior AI, an app he has run for four years on image models that take a room photo in and return a redesign. About half an hour later he showed a prototype that can empty a room, move or delete furniture, and repaint walls, floors and ceilings in any color. About three hours after that he said the feature was live for all users, rendered by headless Blender on his own VPS. The three posts drew about 115K views, 600+ likes and 250 replies combined. The discussion centered on processing time, self-hosted rendering versus a 3D API, and subscription versus pay-per-use pricing. This is a developer showcase and discussion; results are from his own demos, not third-party testing.

On November 12, ByteDance's BytePlus and Volcano Engine jointly launched the Dreamina Seedance 2.5 Flash video generation model. ByteDance calls it the closest model yet to flagship Seedance 2.5 in quality, says it generates nearly twice as fast, and says it beats the previous flagship Seedance 2.0 overall. BytePlus prices it at $0.070 per second at 480p and $0.152 per second at 720p, while Volcano Engine in China charges about 0.99 yuan per second at 720p; both are roughly 34% below Seedance 2.5. A single clip runs up to 30 seconds with up to 10 images, 5 videos and 5 audio clips as references. The launch covers only 480p and 720p, without Draft mode. It is rolling out on Dreamina and partner platforms now, and the ModelArk and Volcano Ark APIs open on November 15 at 14:00 (UTC+8).

On November 12 at Baidu World 2026, Robin Li launched Token Exchange, a model marketplace he compared to a stock exchange for AI models. It brings many model service providers together so developers integrate one API, compare models and providers on quality, latency and price, and switch between them. To counter resellers that swap models or inflate specs, Baidu vets provider credentials and runs model fingerprint comparisons plus capability, performance and availability tests before listing, then spot-checks providers afterward and keeps a public credit rating and record for each. Baidu says that one month in, the platform lists over 100 model resources and handles more than 50M calls and 700B tokens a day.

On November 11, Arena announced that Anthropic's Claude Fable 5.2 (Max) is #1 overall on Agent Arena. Across 4,731 real agentic sessions it posts a +30.35% net improvement, about 1.6x runner-up Claude Opus 5.5 (Max) at +19.52%, with Gemini 4 Carbon (High) third at +16.98% (preliminary). Versus Fable 5.1, Fable 5.2 climbs from #5 to #1, roughly doubles net improvement, and cuts median cost per task from $7.60 to $7.28. It leads the Praise vs Complaint, Steerability and Confirmed Success signals and ranks #3 on Bash Reliability, but each task still costs about 5x as much as on Gemini 4 Carbon.

OpenAI published "What affects chain-of-thought monitorability across the model lifecycle?" on its Alignment Research Blog, tracking CoT monitorability from GPT-5 to Astra-class models through two proxies: no-CoT capability and CoT controllability. Its four findings: no-CoT capability rises smoothly as base model perplexity falls; some architectures beat the perplexity trend on multi-hop planning and serial computation; RL with only length constraints and final-answer grading still sharply raises CoT controllability; and that controllability can pass from teacher to student through SFT. OpenAI calls the results sobering, while stressing that controllability is a noisy proxy that showed no clear link to monitor evasion in one pair of similarly capable models.

At the Baidu World 2026 opening on November 12, Baidu launched two ERNIE application models, the ERNIE Lightning Deep Thinking model and the ERNIE Wizshot Micro-Expression model, both live on Baidu AI Cloud Qianfan with API access. The matching text model in the Qianfan docs is ERNIE-Light-DT-1.0: a 1M-token context, up to 128K output tokens, switchable thinking and a default reasoning effort of max. Its official OpenRouter description says it is built on open-source GLM models with research from Baidu Search's Lightning model and Baidu Sparse Attention (BSA). Wizshot Micro-Expression (ERNIE-Wizshot-ME) turns a single portrait into a 4 to 15 second 1080p talking video at a list price of 0.7 yuan per second. Baidu has not published any benchmarks.

On November 11 Google released A2UI 1.0, the first stable version of its open generative UI protocol (Apache 2.0). Instead of replying with a wall of text, an agent describes the cards, pickers or charts to show, and each app renders them with its own native components. Version 1.0 adds bidirectional function calls and accessibility support, ships renderers for React, Angular, Lit, Flutter and SwiftUI, and offers agent SDKs for Python, TypeScript and Dart. An official Google Cloud Tech article says the new default inference format, A2UI Express, cut generated UI code tokens by 63.2% on a 54-sample test. Atlassian, Alation, Bloomberg and ByteDance are already building with it.

Independent evaluator Artificial Analysis (AA) said on November 12 that Claude Fable 5.2, released by Anthropic this week, scores 75 on AA-Robotics v0.1 and is the new leader, ahead of former leader GPT-6 Astra (69) and GPT-6.1 Sol (65). The benchmark has coding agents drive a 6-axis robot arm through an HTTP API across eight tabletop tasks. AA says Fable 5.2 is the first model to score on the measure-the-box task, which it did from camera video alone, using a tape measure placed in the scene as a scale to estimate the box's dimensions. Its predecessor, Fable 5.1, scored 46.3 on the launch leaderboard in October.

Independent evaluator Artificial Analysis reran VIALS-AA, its life-sciences visual reasoning benchmark, in a one-shot format with one request and no tools. Tool use lifted the cross-model average from about 25% to about 47%. Leader Claude Fable 5.2 (max) scores 67% with tools at $4.20 per task but 44% without them at $0.73, while Claude Haiku 5.5 (max) reaches 49% with tools at $0.07 per task. The same day AA also reported that across the Claude lineup, moving up a model tier usually beats maxing out effort on intelligence per dollar.

Google Antigravity announced on November 11 that browser use is rolling out in the Antigravity App: the agent can open pages in the side-panel built-in browser, interact with your running app, inspect console errors and failed network requests, take screenshots and verify fixes before finishing. Per the official docs, the built-in browser agent needs App v2.30.0+ and is rolling out on macOS, Windows and Linux to personal accounts on Free, Plus, Google AI Pro and Ultra; business accounts and the CLI still use /browser with local Chrome 144+.

At its New York Summit on November 11, ElevenLabs laid out its next phase for agents. Available now: NVIDIA Nemotron 3.5 Super VL in the ElevenAgents model list, the ElevenLabs Partner Program with a McKinsey alliance, and free early access to a new app called ElevenAI. Previewed or coming in the next few months: Interaction Engine, which manages turn-taking and interruptions and picks the right LLM for each step (cheap models for triage, frontier models for complex steps such as refunds); Memory and Routines in ElevenAgents; and fine-tuning of open-source LLMs and ElevenLabs voice models on enterprise data.

On November 11 Databricks released Genie plugins, which it says work natively in ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot; agents without a native plugin can connect through the Genie One MCP server. An agent can hand a business question to Genie, which interprets terms using certified definitions in Genie Ontology, writes and runs SQL only over data the user may access, routes to domain Genie Agents such as finance or marketing when needed, and returns an answer with citations. Or the agent can just fetch the missing business context and keep working with its own tools.

AWS introduced Nitro Sandboxes this week and promoted them again on its official X account on November 11. A Nitro Sandbox is a lightweight VM managed by the Nitro Hypervisor: each EKS pod, ECS task or EC2 workload gets its own kernel and a hardware-isolated boundary while still running on EC2 instances you already have. AWS names AI agents that run LLM-generated code as a core use case: one sandbox per session, so code that a prompt injection tricks the agent into running stays trapped in that VM, which is deleted when the session ends. Turning it on in EKS or ECS is a one-line change; it works on most Nitro-based instances in all commercial Regions at no additional charge.

On November 11 Google AI Studio announced that Managed Agents on the Gemini API can now use your own sandbox. Until now, managed agents such as Antigravity ran their tools only in a Google-hosted Linux sandbox. Now the agent loop and model calls stay with Google, while tool calls such as shell commands and file reads and writes can run in your own container, VM or microVM. To connect, call POST /v1beta/environments with {"mode":"self_hosted"} to get an environment id and an environment_key that is returned only once, then run environments connect from the open-source gemini-api CLI inside your sandbox. All traffic is outbound HTTPS, so no inbound ports need to be opened.

On November 11 Amazon launched AWS Emulator, a local emulator of the AWS APIs. The AWS CLI, the SDKs and coding agents point at 127.0.0.1:9297 through a few environment variables instead of the real cloud, with no AWS account and no extra fee. It is written in Rust and generated from the same API models the AWS SDKs use; AWS says it covers more than 400 API models across more than 200 services. Most unusually, Senior Principal Engineer Matthew Goldberg says thousands of autonomous agents wrote, tested and fixed all of the code, and his team neither wrote nor reviewed a single line. The source and the fidelity test suite are on GitHub under CC0.

On November 11 the Perplexity Secure Intelligence Institute published Escaping SPACE: Part II, a red-team study of SPACE, the Firecracker VM sandbox behind Perplexity Computer agents, now covering 12 models including newly added Claude Mythos 5.1, Claude Opus 5.5 and GPT-6 Astra. Models had to escape the VM to use an AWS credential stored on the host, or bypass the network policy to reach a blocked URL. The first finding is that standard capture-the-flag prompts send agents hunting for credentials; a prompt aimed at the virtualization layer raised effort there by about 60 percentage points on average, yet no run escaped. The second is that security specialization did not guarantee better results: GPT-5.6 Cyber failed all three network bypass runs that general-purpose GPT-5.6 Sol won, and GPT-6 Astra, averaging 4.8 minutes, beat security-focused Mythos 5.1. The vulnerabilities have been fixed.