Notion, Cloudflare and DeepSeek ship serving, not models

Three of today's AI releases ship no new model. Notion rewrote the harness its agents run inside, Cloudflare put a price on a narrow extraction model, and DeepSeek released the inference framework it already serves production traffic with. All three change how a model is run or which model takes the call, and all three are where the cost now moves.

The session length is a product boundary

Notion rebuilt the harness behind its agents so each session runs on a dedicated service with a queue-backed worker loop, which lifts a single run from a cap of roughly five minutes to more than 24 hours, and shipped an Auto model router the company's own tests put at about 90% lower average cost than frontier models across more than 20 tasks (Notion).

Neither figure describes a model. The five minutes was a timeout, and the 90% is a routing decision about which request deserves an expensive answer. What changes for a user is the kind of work that can be handed over: a task that takes an afternoon was previously not a task an agent could be given at all.

That matters to Notion more than to most. The company reported $600M of recurring revenue as of December 2025, roughly half of it from AI products (Forbes), and sold $270M of employee shares at an $11B valuation in January 2026 (Notion). Customers had built more than a million Custom Agents in the three months to May (TechCrunch). A 90% average cut at the router lands on the cost of serving half a company's revenue, not on a benchmark.

Zero on output tokens is a statement of scope

Cloudflare put Clairic into preview on Workers AI, an extraction model that returns typed values with confidence scores from text and up to four images, priced at $0.10 per million input tokens with no charge at all on output tokens, with open weights still to come (MarkTechPost).

Charging nothing for output only works if output is small by construction. Extraction returns a handful of typed fields, so the pricing says what the model is for and what it refuses to be: not a model you hold a conversation with, a function you call on a document. A general model priced this way would be a loss.

The framework is the release

DeepSeek open-sourced DeepInfer under Apache 2.0, the framework that serves DeepSeek-V4.1-Flash in production, bundling optimised kernels, distributed inference and speculative decoding, and running on Nvidia SM100 GPUs and Huawei Ascend 950 NPUs alike (The New Stack).

The weights are not the news here; the serving layer is. Speculative decoding and distributed inference are throughput work, and the two-vendor support means the throughput is not tied to one supplier's silicon. A framework that runs on either is a hedge anyone downstream can take.

What is checkable and what is not

Two of the three claims are the vendor's own. Notion's 90% comes from Notion's tests on Notion's task mix, and the task count, more than 20, is small enough that the mix decides the number. Cloudflare's model is in preview and its weights are not open yet, so the only verifiable part of that release today is the price.

DeepSeek's is the claim that can be read: it is code, under a licence that permits reading it, serving a model the company runs itself. That asymmetry is worth keeping in view as more of the stated savings in this market come from the serving layer, because a harness is easy to measure in private and harder to measure from outside.