Notion Co-founder Akshay Kothari: Claude Sonnet 5.5 Released to All Users
Notion Co-founder Akshay Kothari: Sonnet 5.5 Fully Released
Notion Co-founder Akshay Kothari stated that Claude Sonnet 5.5 has been released to all users, but Opus 5.5 is not yet fully online. The reason is that the influx of multiple models has caused traffic congestion, with more models to be deployed in the future.
On the same day, Anthropic released Sonnet 5.5, positioned as the second model in the Claude 5.5 family, priced the same as Sonnet 5: $2 per million input tokens and $10 per million output tokens. The official claim is that speed has increased by over 30%, with task costs potentially reduced by up to 30%. The context is 1 million tokens, with a maximum output of 128,000 tokens. Terminal-Bench 4.0 scored 70.6%, higher than Sonnet 5's 10.3% and Opus 5.5's 66.4%.
Opus 5.5 went live on Anthropic's official channels on September 22, priced at $4 per million input tokens and $20 per million output tokens, about 20% lower than Opus 5. The company claims typical workload running costs are about 40% lower. This model is aimed at long-range agent programming and knowledge work, with default adaptive thinking, knowledge cutoff in June 2026, and also supports 1 million token context. Kothari mentioned that Notion is still running Opus 5, with unit costs about twice that of 5.5.
Notion has turned Claude into a managed agent within its workspace: Business and Enterprise users can create Claude agents, hosted by Anthropic's infrastructure, without needing their own accounts. The agents can read and write pages, databases, comments, and connect to GitHub to create pull requests. The external agent interface has previously allowed Claude, Codex, Cursor, and others to enter the same workspace. There have been about 18,000 Claude agents on the client side, with around 140,000 executions, about 90% of which run on automatic triggers rather than conversations.
Sonnet 5.5 is officially defined as a fast model for daily coding, bug fixing, document slides, and spreadsheets, while Opus 5.5 is reserved for open tasks requiring continuous judgment. On CursorBench 4.0, Sonnet 5.5 scored 55.5% and Opus 5.5 scored 57.8%; on GDPval-AA v2.1, both are nearly equal (1844 vs. 1846). The application side connects to both tiers, which will stack inference queues on the same quota and gateway during peak times.
What is purchased is the default model switching rights in the enterprise workspace: users will not leave Notion to open a separate Claude window, with traffic ordered from Notion to Anthropic. What is sold is Anthropic's new model capacity. The event-driven aspect comes from the simultaneous release of the 5.5 family, with funding shifting from purchasing Opus 5 at the old price to the cheaper 5.5, but the launch window is congested by concurrent releases. The beneficiaries are the end users who first access Sonnet 5.5 and Notion, which can reduce token bills; the pressured parties are those still billed at Opus 5 and the inference and gateway layers that need to handle multiple model traffic in the same week.
Source: Public Information
ABAB AI Insight
In recent years, Notion's actions have not focused on training large models but rather on transforming its document library into an agent operating system: first creating Custom Agents, then integrating Claude Managed Agents, and finally opening up External Agents to allow Cursor, Codex, and Decagon to work on the same task board. Kothari's update continues this logic—whoever embeds the new model into the existing pages and triggers of the team first will capture the knowledge workflow traffic, rather than competing with Anthropic in the lab.
The capital path involves the application layer using subscription and seat revenue to purchase tokens from the lab. Opus 5 is about twice as expensive as 5.5, giving the application side a clear incentive to switch models immediately, but the switch is limited by the other party's capacity and its own routing queue. Money flows from Notion's AI quota to Anthropic's inference bills, then back to the product team that can deploy the new model into the workspace on the same day. The strategy is to occupy the "distribution window on model release day," rather than building its own computing power.
Similar positions can be seen with Microsoft embedding OpenAI models into Office, Salesforce embedding models into Agentforce, and Cursor embedding models into editors: the platform first becomes the default desktop, while the lab is responsible for supply. Notion is in the middle of transitioning from a note-taking tool to a workflow control layer, with the model itself still being an external purchase, pricing power with Anthropic, and distribution power with Notion.
This is a layer in the restructuring of the industry chain: the pace of model releases is faster than the pace of application integration, with the bottleneck shifting from "is there a new model" to "who can integrate the new model into existing workflows on release day without overwhelming the queue." The mechanism is that labs use price reductions to encourage switches, while applications use double the old cost to wait, making the routing and quotas the new scarce resources.
ABAB News · Cognitive Laws
On the first day of a new model, queuing is more expensive than parameters.
Price reductions do not sell unless the channel can connect on the same day.
Whoever controls the workspace places orders with the lab.