AI news: November 3–10, 2026
Two frontier models built for long agentic work shipped a day apart: Anthropic's Claude Fable 5.2 on Monday and Google's Gemini 4 Carbon on Tuesday. The tools moved in the same direction, with Claude Code, Cursor and Antigravity all adding ways to run agents in the background or in the cloud. Anthropic also published a detailed account of how agents should remember things between sessions.
AI models
Claude Fable 5.2: cheaper and faster, Mythos 5.2 only for verified organizations
Anthropic released Claude Fable 5.2 (claude-fable-5-2) on November 9 with a 1M-token context window, priced at $8 per million input tokens and $40 per million output tokens (Fable 5.1 cost $10 and $50). Anthropic says it generates output up to 30% faster than 5.1, and an optional fast mode gives up to 2.5x more output tokens per second at $16/$80. Claude Mythos 5.2 is the same model without some of the cyber and biology safeguards, available only to organizations in Anthropic's verification programs. Anthropic itself recommends starting most tasks with Claude Opus 5.5 and moving up to Fable, or down to Sonnet 5.5 or Haiku 5.5, as needed.
Source: Anthropic, Claude Platform release notes
Gemini 4 Carbon reaches the API, with a price step after November
On November 10 Google opened Gemini 4 Carbon to all paying Gemini API developers (via AI Studio) and Gemini Enterprise customers; Google AI Ultra subscribers get it in Antigravity, the Gemini app and AI Mode in Search. The introductory price matches Gemini 4 Argon, $2 per million input tokens and $10 per million output tokens, until November 30, then rises to $4 and $20. Google aims Carbon at complex multi-step coding, debugging and research, notes that it spends more tokens to explore and verify, and says Argon "remains the ideal choice for the majority of software engineering and knowledge work".
Source: Google
Meta Muse Ember: a small model for the repetitive part of agent work
Meta released Muse Ember 1.0, its smallest multimodal model, built for high-volume agentic workflows; Google Cloud lists November 9 as the release date. It has a 1M-token context and costs $0.10 per million input tokens and $0.50 per million output tokens for prompts under 200k tokens ($0.30/$1.50 above that), through Meta's Model API, OpenRouter and Google Cloud. In Meta's own table it beats Claude Haiku 5.5 on AutomationBench-AA (37.8 vs 13.9) but trails it on FrontierCode 2.0-AA (27.2 vs 34.9), which fits how Meta describes it: a model for the work agents repeat all day.
Source: Meta, Google Cloud
Kimi K3.1 Pro weights are on Hugging Face
Moonshot AI published the weights of Kimi K3.1 Pro, which was previously available through Kimi's apps and API (TechNode reported the release on November 5). It is a mixture-of-experts model with 2.8 trillion total and about 104 billion active parameters and a 1,048,576-token context; the model card lists vLLM, SGLang and TokenSpeed for inference. Moonshot's own numbers put it close to Claude Opus 5.5 on Terminal-Bench 4.0 (65.7 vs 66.4). The weights ship under a custom "Kimi K3.1 Pro License", so read it before any commercial use.
Source: Hugging Face, TechNode
Coding with AI
Claude Code: MCP tasks in the background and a mode that denies instead of asking
Five Claude Code releases shipped this week, from 2.1.317 to 2.1.322. Version 2.1.319 (November 6) moves MCP tool calls that a server runs as tasks into the background and reports the result when they finish, and adds permissions.denyInsteadOfPrompting, which denies a tool call that would wait on a permission prompt so the turn can continue. Version 2.1.317 lets Agent SDK hook callbacks set onFailure: "block", so a failing hook blocks the action instead of letting it run; since 2.1.321, Fable 5.2 is the default Fable model.
Source: Claude Code changelog
Cursor builds and hosts apps, and agents now take voice calls
On November 5 Cursor added Build Live Apps: a cloud agent scaffolds a web app, sets up a Postgres database when needed, puts authentication behind Cursor SSO, provisions hosting and shares the app via a link. Before anything goes live, Cursor runs deterministic and agentic security checks (for example for secrets written into app files), and team admins can add their own review rules and unpublish any app; the feature is on all paid plans with cloud agents enabled. A day later Cursor added realtime voice calls with agents, running on Grok Voice, with zero data retention when Privacy Mode is on.
Source: Cursor: Build Live Apps, Cursor changelog
Google moves agentic coding into Antigravity
Antigravity 2.29.0 (November 3) started rolling out remote agents, conversations that run in Google's cloud instead of on your machine, to Google AI Ultra and Pro subscribers; version 2.30.0 (November 9) reports the rollout to Ultra subscribers as finished. Version 2.31.0, also from November 9, adds /swarm, which coordinates multiple subagents over a message board to complete long-horizon tasks. Separately, on November 6 Google said the Gemini Code Assist entitlement bundled with Gemini Enterprise Standard and Plus ends on March 31, 2027, is limited to Gemini 3.x models until then, and pointed developers to Antigravity for agentic coding.
Source: Antigravity changelog, Gemini Enterprise release notes
Conductor joins Meta; its cloud shuts down on November 20
Conductor, an app for running a team of coding agents, announced on November 4 that it is joining Meta. Conductor Cloud workspaces go read-only on November 13, and the service shuts down on November 20, when deletion of cloud user data begins; the Mac app keeps getting new models and critical fixes until early 2027, then winds down. Cloud workspaces can be moved with "Transfer to local", and a full archive is under Settings → Account → Export data.
Source: Conductor
AI agents
Anthropic on agent memory: four layers and "dreaming"
In an engineering post on November 6, Anthropic splits agent memory into four layers: searchable transcripts, memory notes, structured representations such as wikis, and context injected at the start of every session. It also describes "dreaming", an offline pass in which Claude rereads logs, corrects notes and decides what to preload next time; on a 25-question documentation task, wiki-only scores rose from 13% to 48% as dreaming scaled from 2 to 128 sessions, and wiki plus search reached 61% against 53% for search alone. The practical advice: keep the always-loaded layer small (Claude Code reads only the first 200 lines or 25 KB of MEMORY.md), have one agent write each page, and quote critical facts verbatim instead of summarizing them.
Source: Anthropic
AWS adds web search and cited answers as agent tools
Amazon Bedrock AgentCore got two web tools this week. Web Answer, generally available since November 5, returns a synthesized answer with citations in one call; it attaches to an AgentCore Gateway and is called over MCP with IAM authentication, for now only in US East (N. Virginia). The Web Search API (November 10) returns ranked results from a single InvokeWebSearch call with domain and publication-time filters, costs $7 per 1,000 queries and runs in six regions, including Europe (Ireland).
Source: AWS: Web Answer, AWS: Web Search API
Mistral Studio gets managed agents and app hosting
On November 5 Mistral expanded its Studio platform with managed agents (runtime, skills, sandboxes, subagents and hooks) and apps with hosting, authentication and access control, deployed with one command. It also added OpenTelemetry traces for every run, evaluations on your own datasets, and a Mistral MCP server, so you can build with Mistral Vibe or another coding agent and deploy to Studio. Deployment can run in Mistral's cloud, your own VPC or on-premises, which matters for teams with EU data requirements; the announcement lists no prices.
Source: Mistral
Other
OpenAI GPT-6 Cyber: offensive capabilities behind identity checks
On November 5 OpenAI released GPT-6 Cyber, a variant of GPT-6 Astra for authorized security work such as vulnerability discovery, exploit development and web penetration testing. Access goes through the Daybreak Red program with identity verification, SSO, MFA and API IP allowlisting, and OpenAI rates the model "Critical" for cybersecurity under its Preparedness Framework. According to the system card it scores 48.45% on ExploitGym (GPT-6 Astra: 42.43%) and, without safety instructions, exploited an out-of-scope honeypot in 0.07% of graded runs, against 49.5% for GPT-5.6 Sol.
Source: OpenAI system card, OpenAI API changelog
METR: AI is roughly doubling the pace of progress at Anthropic
On November 5 METR published an independent assessment based on internal Anthropic data it reviewed from August 12 to October 29. It estimates that Anthropic's capabilities on well-specified tasks are improving about twice as fast as between March 2024 and October 2025 (1.9x on Anthropic's internal benchmark trend, 95% interval 1.5–2.8x) and attributes roughly 80% of the current acceleration to AI use itself, with substantial uncertainty. METR also notes that Anthropic redacted most of the evidence METR found most informative.
Source: METR
Anthropic starts its IPO roadshow
On November 10 Anthropic launched the roadshow for its initial public offering: 825 million Class A shares at an expected $70–78 per share, to be listed on the Nasdaq Global Select Market under the symbol "ANT". Of those, 801,351,354 shares are offered by the company and 23,648,646 by existing stockholders. Anthropic filed its S-1 on October 26.
Source: Anthropic
What it means if you build with an agent
- Before you switch an agent loop to Fable 5.2, check the API changes: a request that edits the system prompt, tools or an earlier message and then sends back a 5.2 thinking block returns a 400 error, and computer use needs the
computer_toolset_20260801toolset (release notes). - Gemini 4 Carbon tests run at the introductory price until November 30. If you plan to keep it, budget for $4/$20 per million tokens from December.
- For agent memory, keep what loads into every session short, keep transcripts searchable as the source of truth, and let one agent own each note.
/swarm, remote agents and background MCP tasks do more work while you are not watching. Settle the plan and the acceptance criteria before you start them, because you will not see each step.- If you use Conductor Cloud, export your data before November 20.
Planning a project with an AI agent? NodePlan helps you describe it before the first line of code is written.
Join