Notion Rebuilds Its AI Agent: 24-Hour Tasks, 90% Cost Cut

Notion shipped its sixth-generation agent harness on October 19, introducing a queue-backed durable execution model that lets agent sessions run for up to 24 hours and resume from their last checkpoint after a process restart, replacing a previous hard five-minute cap. Notion said its upgraded "Auto" router, which dispatches sub-tasks across seven new models including GPT-6.1-Sol, GPT-6 Luna, Opus 5.5 and Sonnet 5.5 in partnership with Anthropic and OpenAI, cuts costs by 90% versus frontier-only usage across more than 20 common task types, a figure Notion conditions on mixed workloads where most sub-tasks are simple enough for a smaller model. The release also adds three safety layers (Plan Mode, "Review for me" mode and prompt-injection detection), MCP connections to 150+ tools, and a split credit system giving Business and Enterprise seats 1,500 Standard and 500 Select credits per month, with new workspaces on the system from October 19 and existing workspaces migrating November 2.

Notion’s agents could run for five minutes. Now they run for twenty-four hours — and the difference is not a setting tweak. The sixth-generation agent harness Notion shipped on October 19 introduces a queue-backed durable execution model that survives process restarts. If you build on Notion’s API or use Custom Agents in production, this is a fundamentally different runtime. A New Execution Model, Not a Longer Timer The previous harness had a hard five-minute cap because agent sessions ran as single-process jobs — any interruption killed them. The new infrastructure runs every session on a dedicated service with a queue-backed worker loop. If the process restarts, the session resumes from its last checkpoint rather than starting over. That sounds like infrastructure plumbing, but the practical difference is significant. Long document analysis across hundreds of pages, nightly pipelines that pull data from external sources and write back to Notion databases, and multi-step research tasks that open URLs and run code mid-session — none of these were viable in five minutes. Twenty-four hours makes them production-ready. The harness also got smarter about context. Long conversations get compacted using each model provider’s native compaction. Large tool results are offloaded from the context window instead of flooding it. Agents now write short scripts and execute them as code rather than spending language model tokens on mechanical work. The result is fewer wasted steps and lower token burn per task. The Auto Router and the 90% Cost Claim Notion added seven new models — including GPT-6.1-Sol, GPT-6 Luna, Opus 5.5, and Sonnet 5.5 — in partnership with Anthropic, OpenAI, and open-weight providers. The upgraded “Auto” router sends each sub-task to the model that best balances quality, speed, and cost. The headline number is 90% cost reduction versus frontier-only usage, benchmarked across more than 20 common task types. That figure is real but conditional: it applies to mixed workloads where most sub-tasks are simple enough for a smaller model. Tasks that genuinely require frontier reasoning the whole way through won’t see that saving. Use it as a baseline, not a guarantee. The credit system now splits into two pools. Business and Enterprise seats get 1,500 Standard credits per month for Auto-routed tasks and 500 Select credits per month for tasks where you manually pick the model or run Custom Agents. The split exists because task types burn credits at very different rates — a few complex agent runs can exhaust a single shared pool quickly. New workspaces moved to this system on October 19; existing workspaces migrate on November 2. Three Safety Layers That Unblock Enterprise The safety story is buried in the announcement but matters for anyone selling to enterprise. Notion now ships three layers of protection for agent actions. Plan Mode shows the agent intended steps in a read-only sandbox before it is allowed to touch anything. Review for me mode uses a separate model to check whether an action is safe before execution and flags risky actions for user approval. Prompt-injection detection looks for hidden commands buried in uploaded files or linked pages. Each layer alone is not novel. Together they form a defense-in-depth posture that enterprise compliance teams can point to. If your Notion agent deployments have been stalling in security reviews, this is worth re-opening those conversations. MCP Expansion, BYOK, and What Is Still Coming Notion agents can now connect to 150+ tools via MCP, up from a few dozen earlier this summer. The official Notion MCP server https://github.com/makenotion/notion-mcp-server handles the connection layer for Custom Agents and the standard Notion Agent alike. Bring Your Own AI is now live in two forms. Any user on a Business or Enterprise workspace can connect their ChatGPT Plus or Pro plan so agent sessions draw from the ChatGPT subscription rather than workspace Notion AI credits. Bring Your Own Key BYOK for enterprise API keys is in private alpha. For teams that negotiate direct model pricing at volume, this signals Notion will eventually let you route through your own contracts entirely. One item worth marking on the calendar: the AI Command Center is not yet available. When it ships, admins will get per-member and per-agent credit usage, spending limits, model controls, and agent governance. Right now, finance teams have limited visibility into which agent tasks are burning through which pools. The Bigger Picture: This Is Not a Notion Story The same week Notion shipped its sixth-gen harness, OpenAI added computer use to the Agents API https://openai.com/index/devday-2026-recap/ , Microsoft launched Guardian Agents for containment and dynamic permissions, and Salesforce had already shipped long-horizon agents to Agentforce in September. Every major platform is answering the same question: what does an agent need to keep working after a chat session ends? The answer is converging on the same three components: durable execution checkpointed state that survives restarts , model routing right model for right sub-task , and safety layers defense-in-depth for autonomous actions . Notion shipped all three on October 19. For developers, the implication is not just that Notion is better now. It is that the infrastructure expectations for any agent platform are rising. If you are choosing where to build autonomous workflows, the baseline now includes durable execution, not as a premium feature, but as table stakes. Notion official announcement https://www.notion.com/blog/notion-ai-more-for-less and the full October 19 release notes https://www.notion.com/releases/2026-10-19 are the right starting points for scoping what moves from your backlog to your sprint.