New AI Agent Tools This Week, Checked (13 November 2026)

6 verified AI agent releases from 9–12 November 2026: self-hosted Claude Managed Agents, n8n 3.2.0 pre-release, kagent 1.0 GA, Fireflies Fred, Retell Tell-1 at $0.035/min and Meta Muse Ember at $0.10/$0.50 per million tokens.

New AI Agent Tools This Week, Checked (13 November 2026)

Short answer: The agent releases we could verify this week are: self-hosted Claude Managed Agents (private beta; price on request; hosted version is tokens + $0.08 per session-hour); n8n 3.2.0 pre-release (stops agent runs when the Gateway credits wallet is empty; stable 3.x still due early December); kagent 1.0 GA on Agent Substrate (Apache-2.0, needs Kubernetes 1.37+); Fireflies Fred as an AI teammate (paid plans from $10 per seat a month, billed annually, plus AI credits); Retell Tell-1 at $0.035/min for the model line; and Meta Muse Ember at $0.10 / $0.50 per million input/output tokens.

Checked 13 November 2026 (Asia/Tehran) against vendor announcements, docs, pricing pages and GitHub releases dated 9–12 November. We have not benchmarked these tools.

Disclosure: some links in this article are affiliate links. If you buy through them, Sophrai may earn a commission at no extra cost to you. It never changes our rankings. Read our affiliate disclosure.

This week in one line: Anthropic put its whole agent runtime inside customers’ own perimeters, n8n kept refining the 3.x line (and walked back two breaking changes), Kubernetes got a GA agent platform that sleeps idle agents, Fireflies turned a notetaker into a credit-metered teammate, and voice and small-model prices kept falling. Below are 6 releases from 9–12 November 2026, each checked against the vendor’s own page today, plus three shorter notes.

The week at a glance

ToolTypeLaunchedPrice / licenceBest for
Self-hosted Claude Managed AgentsAgent runtime (private beta)12 NovContact Anthropic; hosted = tokens + $0.08/session-hourRegulated teams that need agent data in-house
n8n 3.2.0Automation / agents (pre-release)10 NovOpen source; stable 3.x early DecTesting the 3.x path on a copy
kagent 1.0 + Agent SubstrateKubernetes agent platform (GA)10 NovApache-2.0; enterprise price not publicPlatform teams running many agent sessions
Fireflies FredAI teammate (meetings, email, Slack)11 NovFree; Pro $10 / Business $19 per seat (annual) + creditsTeams already recording meetings
Retell Tell-1Voice-agent LLM (beta)12 Nov$0.035/min model lineContact-centre voice agents
Meta Muse EmberSmall model for agent workloads9 Nov$0.10 / $0.50 per 1M tokens (≤200K input)Routing, extraction, tool calls

1. Self-hosted Claude Managed Agents (private beta)

Launched: 12 November 2026. What it is: Anthropic now lets you run the whole Managed Agents service on your own infrastructure, not only the sandbox. You operate the server, its database (PostgreSQL 16 or later) and the workers where agents run shell and file tools. Anthropic ships a Docker Compose stack and a Helm chart. Agents, sessions and session history stay in your database. Model requests go to the provider you choose: Anthropic by default, or Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry, or your own gateway. Features only the self-hosted server has: a choice of inference provider, single sign-on through your OpenID Connect provider, your own secret manager for vault credentials, and short-lived signed agent-identity tokens for MCP calls.

Price / licence (checked 13 Nov 2026): No public price. Anthropic’s pricing page says to contact your account team for self-hosted pricing. For reference, the hosted service bills tokens at normal model rates plus $0.08 per session-hour of running time (idle time doesn’t count) and $10 per 1,000 web searches. Anthropic’s own worked example: a one-hour Claude Opus 5.5 session with 50,000 input and 15,000 output tokens costs $0.58 hosted.

What it means for n8n users: The self-hosted server sends no webhooks, per the differences page. An n8n workflow that waits for a webhook will never fire. Poll the session’s event stream instead (Schedule Trigger plus HTTP Request), or keep your own worker process running. Credentials that need an OAuth refresh (mcp_oauth) are rejected; store a static_bearer token and rotate it yourself.

The catch: Access is through your Anthropic account team only. Nothing is copied from the hosted service, so agents, skills, vault credentials and memory stores have to be recreated. Also missing on self-hosted: cloud environments, the Files API, the prebuilt docx/pdf/pptx/xlsx skills, and scheduled deployments. You own patching and security for the server and workers.

Source: Anthropic announcement (12 Nov 2026), release notes, differences from hosted, self-hosted docs, pricing

2. n8n 3.2.0 (pre-release)

Launched: 10 November 2026 (GitHub tag [email protected]). What it is: The 3.2.0 release notes add agent and admin features on top of the 3.x beta line. Agent runs now stop when the Gateway credits wallet is empty. The default n8n Assistant model moves to Claude Opus 5.5. There is a new Microsoft Work IQ OAuth2 credential, App bindings UI, and an element picker in Apps preview. The chat hub module is now off by default. The Migration Report hides archived workflows by default, and the Paddle node is renamed Paddle Classic. Two planned breaking changes were reverted: the Execute Sub-workflow node keeps its “Run once for each item” mode, and the “Any workflow” caller policy stays.

Price / licence (checked 13 Nov 2026): Open source on GitHub under n8n’s own licence (GitHub lists it as “Other”); Cloud and self-hosted paid plans are unchanged by this tag. If you want to test 3.x on Cloud, start from n8n. The n8n 3.0 announcement still says the stable 3.x release follows in early December, and 2.x then gets security and critical fixes for 12 months. The current stable build is 2.47.5 (12 November).

What it means for n8n users: The two reverts are good news if your workflows use per-item sub-workflow calls or open caller policies: fewer fixes before 3.x. Re-run the Migration Report on your latest 2.x instance, because the 3.x breaking-change list is still moving. If you use n8n’s built-in AI credits, decide what an empty wallet should do to production agents before you upgrade. Our n8n 3.0 breaking changes guide covers the rest.

The catch: This is a pre-release. Do not move production onto it. Release notes for a pre-release can change before stable, and the reverts show the breaking-change list isn’t final yet.

Source: GitHub [email protected], n8n 3.0 blog

3. kagent 1.0 on Agent Substrate (GA)

Launched: 10 November 2026. What it is: kagent, the open-source Kubernetes-native agent platform, reached 1.0, and every agent now runs on Agent Substrate, which also hit 1.0. Substrate suspends idle agent “actors” (snapshotting their workspace), resumes them on a free worker when the next request arrives, and runs each actor in its own gVisor or microVM sandbox. kagent agents are reachable over A2A and MCP, with a UI, a CLI and OpenTelemetry traces, metrics and token usage. Both are CNCF Sandbox projects. Solo.io announced GA of its enterprise releases the same day (access control, tracing, evaluations, air-gapped installs).

Price / licence (checked 13 Nov 2026): Open source, Apache-2.0 on GitHub (4,266 stars when we checked). You pay for your cluster and model provider (OpenAI is the default). Solo.io’s enterprise distribution has no public price. Vendor claims, labelled: Agent Substrate has 6,000+ GitHub stars and 100+ contributors from 40+ organisations; launch customers include OutSystems and Swissquote.

What it means for n8n users: If your agents already live on Kubernetes, kagent can expose them over MCP, so an n8n workflow could call them with the MCP Client Tool node. We have not tested that connection. For most small teams, n8n’s own Agent node is simpler; kagent fits platform teams running many long-lived agent sessions.

The catch: You need Kubernetes 1.37 or later (Substrate uses certificate APIs that went GA there), Helm and a model key. kagent install sets up Substrate, PostgreSQL and kagent together. Check that your managed cluster offers 1.37 before planning a rollout.

Source: kagent 1.0 release post, Solo.io press release (10 Nov 2026), GitHub repo

What this means if you automate with n8n

Three patterns from this week are worth copying into any n8n setup: keep agent data where you can audit it (Anthropic’s self-hosted server, kagent’s OpenTelemetry traces), put a hard stop on AI spend (n8n 3.2.0 stopping agents when credits run out, Fireflies’ credit packs), and use small models for the busy steps (Muse Ember, Tell-1). Before you add any of these to production, run your workflow export through our free n8n Workflow Auditor, estimate monthly cost in the n8n Cost & Hosting Planner, and compare AI subscriptions in the AI SaaS Pricing Comparison. If you plan to move to 3.x, start with our n8n 3.0 breaking changes guide.

4. Fireflies Fred becomes an AI teammate

Launched: 11 November 2026. What it is: Fireflies turned its notetaker into an assistant that works across meetings, email, Slack and connected apps. Live today: meeting notes and Live Assist, daily briefs and meeting prep, free unlimited voice typing (Fireflies Talk) with a Command Mode, an Email Assistant that drafts replies for review, a Slack Assistant, Artifacts (docs, decks, dashboards from conversations), Knowledge Search, and Voice Agents that run screening, discovery, support and interview calls. Connected apps grew to 240+ tools, including more than 100 MCP connectors (vendor figures).

Price / licence (checked 13 Nov 2026): The pricing page lists Free $0, Pro $18 per seat a month ($10 billed annually), Business $29 ($19 annually) and Enterprise $39 (annual only). The blog says email and Slack assistants, Voice Agents, FredBots and connected tools are on paid plans “subject to usage credits”. Per the billing guide, accounts get one-time credits on sign-up (Free 20, Pro 20, Business 30, Enterprise 50), shared across the team; extra packs run from 50 credits for $5 to 10,000 for $600 a month (20% less on annual), and unused purchased credits don’t roll over.

What it means for n8n users: Fred’s CRM updates and follow-ups overlap with what many teams build in n8n. If you already route meeting outcomes through n8n, compare Fred’s credit cost per task with your current model costs before switching. We did not check Fireflies’ API or an n8n node today.

The catch: Several pieces are not live yet: FredBots roll out gradually from launch day; Scheduler and Clips over the next couple of weeks; Knowledge Base and Fireflies Call over the next few weeks; the Fly hardware early next year. Fireflies doesn’t publish how many credits a FredBot task or a Voice Agent call uses.

Source: Fireflies blog (11 Nov 2026), pricing, billing and AI credits guide

5. Retell Tell-1 and Retell Dialer

Launched: 12 November 2026. What it is: Retell launched Tell-1, a language model built for live contact-centre calls, trained (Retell says) on hundreds of millions of calls. It uses an interleaved reasoning design so the model can think while it speaks, aimed at partial answers, topic detours and tool lookups mid-call. The same day Retell introduced Retell Dialer for outbound campaigns across voice and SMS agents: callbacks at the contact’s requested local time, hand-offs to specialist agents, SMS nurture, local presence, calling windows, number rotation, spam monitoring, retries and CRM integration.

Price / licence (checked 13 Nov 2026): Retell’s pricing page lists Tell-1 (Beta) at $0.035 per minute for voice and $0.011 per AI message for chat. On the same page GPT 5.6 Terra is $0.064/min ($0.128 on the fast tier) and GPT 6.1 Terra $0.032/min. Retell Voice Infra is $0.055/min and its Cartesia, OpenAI and Inworld voices $0.015/min. Our arithmetic: Tell-1 + infra + one of those voices = about $0.105 a minute ($105 per 1,000 minutes) before telephony, versus about $0.134 with GPT 5.6 Terra. Vendor claims, labelled: under 100 ms responses on Retell’s live benchmark, built to stay under 300 ms at peak, and the top overall score on Retell’s own benchmark.

What it means for n8n users: Retell calls are usually started and logged from outside: n8n can launch calls through Retell’s API with the HTTP Request node and write call outcomes back to a CRM. If you are choosing a voice platform, see our Vapi vs Retell vs Bland comparison for per-minute totals.

The catch: Tell-1 is in beta, and the benchmark is Retell’s own (vCX benchmark post). We found no public Dialer price. Telephony and any premium voices are billed on top.

Source: Retell press release (12 Nov 2026, via TMCnet), Retell pricing, Retell benchmark

6. Meta Muse Ember

Launched: 9 November 2026. What it is: Meta’s smallest, most efficient model in the Muse family, aimed at the high-volume work inside agents: tool calls, routing, extraction, document parsing and coding subtasks. It has a 1M-token context window, up to 512K output tokens, supports the Responses API, and is available on the Meta Model API and in Muse Code as muse-ember-1.0.

Price / licence (checked 13 Nov 2026): Per Meta’s pricing page: up to 200K input tokens, $0.10 per million input, $0.004 cached input and $0.50 output. Requests over 200K input tokens are billed in full at $0.30 / $0.012 / $1.50. Standard mode only (the discounted Contributor mode isn’t offered for Ember). Default limits: 3,000 requests and 4,000,000 tokens per minute. Our arithmetic: an agent step with 5,000 input and 500 output tokens costs about $0.00075, so 100,000 such steps is about $75.

What it means for n8n users: Cheap routing and extraction models matter in n8n agents because every tool loop re-sends context. Use a small model for classification and field extraction, and keep a larger one for the final answer. Whether your n8n model node can point at the Meta Model API depends on its OpenAI-compatible settings; we have not tested it.

The catch: The 200K threshold is a cliff, not a step: one long request is billed entirely at the higher rate. Meta’s benchmark charts are its own and Artificial Analysis’s; we didn’t verify quality ourselves.

Source: Meta announcement (9 Nov 2026), Meta Model API pricing and rate limits

Also this week

  • OpenAI Businesses in ChatGPT (beta, 10 Nov). Businesses can claim a profile in Business Hub, verify with Persona, add product feeds and see 28-day insights on mentions, website clicks and discovery topics. Eligible US ecommerce and online-service businesses with a website can publish a Business Agent that answers product and policy questions. Per OpenAI’s help article, the agent can’t complete purchases, make bookings or check orders, and businesses don’t see customer conversations. No price is listed, and you don’t need an Ads account to claim. Sources: OpenAI announcement, OpenAI help article.
  • Nectar Creator Agent (12 Nov). An agent that finds Instagram and TikTok creators, checks their content and past partnerships, sends personalised outreach and tracks campaigns; Nectar says it searches 300+ million creator profiles. Available globally. Source: TechCrunch.
  • Microsoft Agent Framework python-1.25.0 (7 Nov). Adds callable MCP discovery approval policies and honours Foundry Toolbox approval metadata (tool-name and read-only filters). MIT licence. Source: GitHub release.
  • Already covered on Sophrai: Databricks always-on Genie pricing and the 30 November MCP deadline (11 Nov) and n8n Labs’ open-source Nod agent (10 Nov).
  • Previous roundups: 6 November 2026 and 30 October 2026.

Which one should you try first?

  • Regulated or security-heavy teams on Claude: ask your Anthropic account team about the self-hosted private beta, and plan for polling instead of webhooks.
  • n8n builders: stay on 2.47.x in production; spin up a disposable 3.2.0 instance to check the Migration Report and the new credits-wallet behaviour.
  • Platform teams on Kubernetes 1.37+: try kagent 1.0 on a kind cluster first; it installs Substrate and PostgreSQL in one command.
  • Sales and support teams: test Fireflies Fred on the free plan, then price credits for the FredBots you actually need.
  • Voice-agent builders: run a side-by-side test of Tell-1 against your current model on real call scripts. At list prices it saves about $0.03 a minute on the model line versus GPT 5.6 Terra (our arithmetic).
  • Anyone paying for high-volume extraction: benchmark Muse Ember on your own documents, and keep requests under 200K input tokens.

What we couldn’t verify

  • Self-hosted Claude Managed Agents pricing (Anthropic says to contact the account team) and any hardware sizing guidance beyond PostgreSQL 16+.
  • How many Fireflies AI credits a FredBot task, an Email Assistant action or a Voice Agent call consumes.
  • Retell Tell-1’s latency and quality claims (Retell’s own benchmark) and the price of Retell Dialer.
  • Muse Ember’s quality on agent tasks (charts are Meta’s and Artificial Analysis’s) and whether n8n’s model nodes work with the Meta Model API.
  • Solo.io’s enterprise kagent price, and whether the major managed Kubernetes services already offer version 1.37.
  • n8n 3.2.0’s Gateway credits wallet pricing (the release notes describe the stop behaviour, not the price).

Frequently asked questions

What were the most important new AI agent tools this week?

The six releases we verified from primary sources on 13 November 2026 are self-hosted Claude Managed Agents (private beta, 12 November), the n8n 3.2.0 pre-release (10 November), kagent 1.0 on Agent Substrate (GA, 10 November), Fireflies’ new Fred AI teammate (11 November), Retell’s Tell-1 voice model at $0.035 a minute (12 November) and Meta’s Muse Ember model at $0.10 / $0.50 per million tokens (9 November). OpenAI also opened Businesses in ChatGPT in beta.

How much do self-hosted Claude Managed Agents cost?

Anthropic does not publish a price. Its pricing page says to contact your Anthropic account team for self-hosted pricing. For comparison, the hosted Managed Agents service bills model tokens at normal rates plus $0.08 per session-hour while a session is running; web searches inside a session cost $10 per 1,000.

Is n8n 3.2.0 safe to use in production?

No. n8n 3.2.0 is tagged as a pre-release on GitHub (10 November 2026). n8n’s own 3.0 announcement says the stable 3.x release follows in early December. The current stable line is 2.47.x (2.47.5 shipped on 12 November). Test 3.2.0 on a copy of your instance and run the Migration Report first.

What does Retell Tell-1 cost per minute?

Retell’s pricing page lists Tell-1 (Beta) at $0.035 per minute for voice calls and $0.011 per AI message for chat agents. That is only the language-model line. Retell Voice Infra is $0.055 a minute and its Cartesia, OpenAI and Inworld voices are $0.015 a minute, so by our arithmetic a call on Tell-1 with one of those voices is about $0.105 a minute before telephony.

How much does Meta Muse Ember cost?

Meta lists muse-ember-1.0 at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens for requests up to 200K input tokens. Requests above 200K input tokens are billed in full at $0.30, $0.012 and $1.50. The default rate limit is 3,000 requests and 4,000,000 tokens per minute.

How did you check these tools?

We read each vendor’s own announcement, documentation, pricing page or GitHub release on 13 November 2026 (Asia/Tehran). Benchmarks and adoption figures are labelled as vendor claims; we did not run them. Cost examples marked as our arithmetic are calculated from the list prices shown, not quotes.

Sources and method

Every date, price and limit above comes from the vendor’s own page, checked on 13 November 2026 (Asia/Tehran): Anthropic self-hosted Managed Agents, Claude release notes, self-hosted differences, self-hosted overview, Claude pricing, [email protected], n8n 3.0 blog, kagent 1.0 post, Solo.io press release, kagent on GitHub, Fireflies blog, Fireflies pricing, Fireflies billing guide, Retell press release, Retell pricing, Retell benchmark, Meta Muse Ember, Meta Model API pricing, OpenAI Businesses in ChatGPT, OpenAI help, TechCrunch on Nectar, Microsoft Agent Framework release. We have not benchmarked these tools. Adoption, latency and quality figures quoted from vendors are their claims.

hamid7777 Avatar

Written & tested by

The Sophrai brief

One calm email when something actually matters.

New tools, real cost breakdowns and the automation news worth your time. No hype, no spam.