Claude Code Plugin Evals, Notion 3.7 Agent SDK, Suno v6 Licensing Revolution: September 2026 AI Digest

September was the month AI tools chose maturity over spectacle.

Instead of announcing flashy new features, three major AI tools faced their own weaknesses directly. Claude Code finally answered the question developers had long been asking: "Does my plugin actually work?" Notion elevated its agent capabilities from a solo tool to team infrastructure that everyone shares. Suno addressed the copyright problem that has weighed on the entire AI music industry — not by dodging it, but by licensing the music and redesigning the editing experience at the same time.

Reliability updates rarely make headlines. But these three changes could mark the turning point from "trying out AI tools" to "fully integrating them into real work."

Claude Code Plugin Evals: Now Plugins Have to Pass a Test
Managed MCP Servers and the September Feature Bundle
Notion 3.7: Agent SDK Public Beta — Call Your Agent From Any App
AI Search 50% Faster, Model Controls for Admins: Notion's Quiet Revolution
Suno v6: Solving Copyright with Licensing and Reinventing Music Editing
Claude Code Projects Beta: Run Long Jobs Like Experiments
Notion 3.7's Agent Skills Library: SKILL.md for Any Agent
Gemini 3.5 Transcribe and 3.8 Flash: New Baselines for Voice and Speed
Claude Fable 5.1: 75% Cheaper Cache Reads and Five Levels of Thinking Depth
Suno v6 Terms of Use: Download Caps and the Label-Deal Timeline
The Signal All Three Tools Are Sending

1. Claude Code Plugin Evals: Now Plugins Have to Pass a Test

In the first week of September 2026, Claude Code shipped its plugin evaluation (eval) system. Before you deploy a plugin, you can now run test cases against it, score the results, and compare performance against a no-plugin baseline — all from a single CLI command.

Until now, Claude Code plugin developers had no reliable way to measure whether their plugin actually helped in real tasks. They relied on subjective feel, or ran manual one-off tests. This update fills that gap directly.

What claude plugin eval Does

# Auto-draft test cases and graders for your plugin
claude plugin eval init

# Run evaluation (with baseline comparison)
claude plugin eval run

Running claude plugin eval init causes Claude to analyze the plugin's purpose and functionality, then auto-draft test cases and scoring criteria. You review and refine them, then run claude plugin eval run to get results.

The output shows three things: per-test-case scores, comparison against the no-plugin baseline, and identification of cases where the plugin underperforms. Anyone who has heard "I vibe-coded this plugin but I'm not sure it actually works" will understand why this feature was overdue.

Other Features in the Same Release

Feature	Description
Desktop pane pop-out	Detach any Claude Code Desktop pane into its own window and re-dock it later
maxEffortLevel setting	Cap the effort level applied to every provider request
Memory warnings	Visual warning when memory is critical, with steps to free memory or restart safely
v2.1.270 bug fix	Fixed git commands unexpectedly requesting permission during long sessions

2. Managed MCP Servers and the September Feature Bundle

The second pillar of Claude Code's September update is managed MCP (Model Context Protocol) servers. Where setting up an MCP server used to be a barrier in itself, Claude Code now handles the server management infrastructure.

Managed MCP servers consolidate installation, updates, and permission management inside the Claude Code environment. Headless unattended permission controls, GitLab MR recognition, and JSON plugin validation were added alongside this.

In team and enterprise settings, the hardest question around MCP servers has always been: "Who has access to which server with what permissions?" Claude Code now provides the foundation to answer that question from a central place.

"We use MCP but I have no idea what my teammates are doing with it" — that excuse no longer holds. Admins can see it and control it.

3. Notion 3.7: Agent SDK Public Beta — Call Your Agent From Any App

On September 15, 2026, Notion released version 3.7 and moved the Agent SDK to public beta. External apps can now call Notion agents directly, continue conversations, and stream results back in real time.

Notion's agent features had previously worked only inside the Notion app — an "internal AI." The Agent SDK public beta dissolves that boundary. Developers can now start conversations with Notion Agent from their own apps, pick up previous conversations, and stream outputs to their app's UI.

What Changed in Notion 3.7

Build skills by chatting: Chat with Notion Agent to create skills from scratch — new-hire onboarding checklists, workflow automations — without any complex configuration UI. Start by saying what you want.

Developer section: The new sidebar section shows every Worker connected to your workspace, which apps each Worker is linked to, and the execution logs for each run.

Three core capabilities of the Agent SDK public beta:

Continue conversations: Pick up a previous Notion Agent conversation from an external app
Trigger from your app: Start a Custom Agent directly from outside Notion
Stream results: Get agent output in real time on your app's screen

4. AI Search 50% Faster, Model Controls for Admins: Notion's Quiet Revolution

Hidden behind Notion's headline Agent SDK, two more changes arrived in September. AI Search is now 50% faster, and workspace owners can now choose which AI models their agents use.

The AI Search speed improvement is a difference of experience, not just a number. The psychological gap between "I'm waiting for an answer" and "the answer is already here" is larger than most people expect. Making AI faster to respond means making it easier to reach for in everyday work.

Model Controls, added in the September 9 release, let workspace owners configure which AI models are available in Notion Agent and Custom Agents separately, and set a default model for Custom Agents.

Notion Inbox AI Prioritization

Prioritize with AI button: AI reads your piled-up notifications, flags what needs attention, and helps you clear the rest
Personal Agent Effort setting: The Effort control that arrived on Custom Agents now extends to the Personal Agent
Notion Mail deprecation: The standalone Notion Mail inbox closes on September 22, 2026

5. Suno v6: Solving Copyright with Licensing and Reinventing Music Editing

On September 9, 2026, Suno launched v6, replacing every previous model. Two things define it: three new models trained from scratch on licensed music from Warner Music Group, BMG, and Believe; and section editing that lets you rewrite any specific part of a finished song in plain language.

AI music services have been battered by copyright lawsuits for two years. Suno approached the problem in v6 not with technical workarounds but with contractual resolution. All three new models — v6, v6-wild, and v6-mini — were trained from scratch on a licensed music catalogue.

Suno v6 Model Comparison

Model	Users	Characteristics
v6	Paid subscribers	High quality, full feature set
v6-wild	Paid subscribers	More experimental, unpredictable creative output
v6-mini	Free users	Fast generation, core features

Section Editing and Multimodal Prompting

Section editing: Edit a specific part of a finished song (verse, chorus, bridge) in plain language. "Make this chorus more energetic" regenerates only that section
Mashups: Combine multiple sources in a single request to create a new track
Multimodal prompting: Use text, audio, images, or video as reference inputs
Distribution partnership: Tracks made with Suno's licensed model are eligible for distribution through Believe and TuneCore, reaching Spotify, Apple Music, Amazon Music, and YouTube

Every v6 output carries durable, tamper-resistant, inaudible audio watermarking and fingerprinting — a system designed to track the origin of AI-generated music.

6. Claude Code Projects Beta: Run Long Jobs Like Experiments

Around the same time as plugin evals, Anthropic opened Projects in Claude Code as a beta. A project is now the unit for long-running work that spans multiple repositories and threads.

Close a Claude Code session and its context used to vanish; open the same codebase the next day and the agent had to read everything again. Projects changes that.

Parallel thread coordination: Multiple work threads run at once inside one project, and results roll up at the project level. If one thread gets stuck, a human can redirect it while the rest keep running.
Shared memory: What one thread learned about the codebase's structure and patterns carries over to the next — no repeated explanations.
Project library: Group several repositories into one project and package recurring tasks for reuse.
Remote session sync: Session files and state are saved to your Claude account, so you can pick up work on web and mobile (Cowork).

Typical uses: split a large refactor across threads and merge the results into one PR, or run a feature thread and a test-writing thread side by side. If plugin evals measure the quality of your tools, Projects shifts your role to running a project that agents work inside.

7. Notion 3.7's Agent Skills Library: SKILL.md for Any Agent

If the Agent SDK is a door to the outside, agent skills are the storeroom where a team's way of working accumulates.

Every team's AI know-how tends to be scattered across personal prompt files and chat threads. Notion 3.7 answers with a team-wide skills library: write a recurring instruction once — "a monthly business review in our team's format" — and anyone on the team gets the same result.

Feature	What it does
Database-managed	Write, edit, and version skills in a Notion database
Cross-agent export	Download as SKILL.md for Claude Code, Codex, Cursor, Gemini, and Grok
Update tracking	A badge flags changed skills so no one runs a stale version
Subagents	Custom Agents can call other Custom Agents for multi-step workflows
Custom MCP connections	Connect external tools such as GitHub and Amplitude
Suggest-changes mode	Agents propose edits for approval instead of applying them directly

Notion also released a separate Notion Agents iOS app for voice, photo, and delegated tasks on mobile. A practical start: turn three recurring tasks — proposal reviews, meeting summaries, customer-reply drafts — into skills, and sync the SKILL.md files to your team's Git repository for version control.

8. Gemini 3.5 Transcribe and 3.8 Flash: New Baselines for Voice and Speed

Google launched Gemini 3.5 Transcribe on August 26 as the successor to Chirp 3, and made it generally available in the Gemini API and Google AI Studio in September.

Item	Spec
Average word error rate (WER)	2.6% non-streaming, 4.0% streaming
Languages	85+ (per-utterance auto-detection)
Diarization and word timestamps	Supported
Custom vocabulary	Up to 1,000 terms
Smart utterance filtering	Filler removal + recognition of mid-sentence corrections

Smart utterance filtering records only the final intent when a speaker corrects themselves — "I can have it done by tomorrow… no, the day after." The low-latency 3.5 Transcribe Live streams over WebSocket for real-time captions and call transcription. It fits lecture and seminar transcription and multi-speaker meeting notes out of the box, and you can try it free in AI Studio first.

In the same window, Gemini 3.8 Flash reached general availability on September 2. It focuses on token efficiency and on code and agentic planning, with an introductory price of $0.75 per 1M input tokens through December 31, 2026.

9. Claude Fable 5.1: 75% Cheaper Cache Reads and Five Levels of Thinking Depth

Behind Claude Code's September feature bundle is the new default model released on September 1: Claude Fable 5.1 (plus its restricted-access twin, Mythos 5.1).

Input and output prices match Fable 5, but prompt cache reads fell from $1.00 to $0.25 per 1M tokens. That cuts costs by about 25% for typical workloads and up to 45% for cache-heavy agent work. An agent that reads a 100K-token system prompt 1,000 times a day drops from $100 to $25 per day in cache costs — about $2,250 saved per month. It offers a 1M-token context window, 128K max output, and always-on adaptive thinking.

Effort	Best for
low	Simple Q&A, summaries, translation
medium	Everyday coding and document drafts (default in Claude.ai and Cowork)
high	Complex debugging, multi-step analysis (default in Claude Code)
xhigh	Architecture design, long-horizon agent planning
max	Scientific research and maximum-accuracy work

To get the most out of caching, pin system prompts and frequently referenced documents at the front of the context, and use cache_control in the API to raise hit rates. Benchmarks, watermarking, and more are covered in our Fable 5.1 · Suno v6 · Gemini 3.7 Flash roundup.

10. Suno v6 Terms of Use: Download Caps and the Label-Deal Timeline

The conditions you need to check before using v6 have changed too.

Suno reached a turning point on August 13 with a global licensing alliance with BMG, and v6 is the result. CEO Mikey Shulman framed the strategy as building the future of music responsibly. Alongside it, new download caps took effect on September 3:

Pro subscribers: 20 songs per month
Premier subscribers: 60 songs per month
Suno Studio Premier: Unlimited

Capture a frame from a YouTube video and feed it to v6 as an image prompt for background music that matches the mood. With download caps in place, it is efficient to iterate on drafts with v6-mini and download only the final version.

11. The Signal All Three Tools Are Sending

Claude Code's plugin evals, Notion's Agent SDK, and Suno's licensed models look like completely different updates on the surface. But all three point in the same direction: AI tools are entering a phase of building verifiable trust.

Plugins can now be measured for real effectiveness. Admins can see what model their agents are running and what they're doing. The music an AI generates can be traced to a source. This is a signal that AI tools are no longer in the experimental-curiosity territory.

Whether you're an educator, a creator, or a team lead, the question to ask when choosing an AI tool has changed. It's no longer "is this feature impressive?" The key question is now: "Can my team actually trust this tool enough to rely on it?"

Closing Thoughts

If these three updates can be summed up in one sentence: AI is growing up.

Tools that prioritize measurability, controllability, and legal stability over novelty are the ones you can build on for the long term. Take a moment to look at your own workflow and identify where the most "unverified" AI is. Then check whether September's updates can fill that space.

Related Posts

Have you tried any of these September updates? Which feature felt the most different? Let us know in the comments!

Sources:

Table of Contents