Notion AI agents now run up to 24 hours as Auto targets lower costs
Notion adds longer-running agents, browser use and dynamic model routing. Here's what its 90% credit claim, MCP upgrades and rollout limits mean.
On October 19, Notion announced new agent infrastructure that can keep a task running for up to 24 hours, compared with roughly five minutes before. Its upgraded Auto model router reportedly matched frontier-model results across more than 20 internal work scenarios at around 90 percent lower average credit cost. Notion is also changing credits, integrations and administrative controls. Some features are live; others remain in private alpha or are scheduled for later.
Notion's agents had shorter task windows and higher exposure to context and credit overhead. Choosing models and connecting external agents required more configuration.
A queued execution service, browser/computer use and context compaction extend work sessions. Auto routes tasks among models while Notion changes the AI credit system and integration options.
From a five-minute chat to a long-running job
Notion describes the sixth generation of its agent harness. A dedicated service and queue-backed worker loop are designed to keep work progressing through interruptions, rather than ending with a short chat exchange. Notion says tasks can now last up to 24 hours, up from about five minutes. Agents also gain a browser and computer for tool execution, including short scripts, while context compaction and offloading of large tool results reduce repeated processing.
A new 'Review for me' mode asks a separate model to assess whether an action looks safe before execution. This is a useful review layer but cannot replace scoped permissions, deterministic enforcement or human approval for high-impact operations. Likewise, Notion's claim of more than 99.9 percent infrastructure uptime does not prove that complicated jobs always finish correctly. Runtime durability and task accuracy are different measurements.
Reading the 90-percent cost reduction carefully
Auto now routes straightforward requests to efficient models and harder analysis or multi-step work to more capable ones. If progress stalls, the router can switch models. Notion reports frontier-like outcomes across more than 20 common work scenarios at roughly 90 percent lower average credit consumption than the high-end comparison models. These are vendor evaluations in Notion's credit economy, not an independent audit or a universal reduction in public model API prices.
The revised system gives Business and Enterprise seats 1,500 Standard credits per month for automatic model selection and 500 Select credits for chosen models or custom agents, according to Notion. Existing customers are due to transition on November 2, 2026. Frequent custom runs may still exhaust their allowance. Teams should compare real end-to-end tasks, including retries and human review, rather than projecting the reported savings onto every workload.
MCP and bring-your-own-AI: Availability matters
Notion now supports linking existing ChatGPT Plus, Pro or Go subscriptions to the personal Notion Agent. That does not initially extend to Custom Agents. Bring Your Own Key is in private alpha for Enterprise, not broadly available. The revised Notion MCP lets outside agents use Notion as a context layer with memory synchronization; MCP-based Skills are described as coming later.
The AI Command Center, which would provide central usage views, spending limits and governance over models and agents, is also announced as coming soon. Enterprises should distinguish live connections from planned control features. An honest pilot should test what information an agent can retrieve through MCP, whether potentially sensitive actions require approval, how many credits a long task consumes and how a stalled process is stopped and audited.
AgentenCode analysis
The strategic change is the combination of durable execution, model routing and credit visibility, not simply a longer timer. Notion is positioning its workspace as a shared context layer for external and native agents. That could become sticky infrastructure, but planned governance features and vendor-run 90-percent benchmarks do not establish production reliability or independent economic superiority.
Sources and verification
As of October 20, 2026. Feature, price-credit and performance comparisons come from Notion's October 19 release. No independent AgentenCode runtime evaluation; BYOK is private alpha and the AI Command Center is upcoming.