Summary

  • On 9 October, Cloudflare lowered Clef-flash’s listed input rate from $0.090 to $0.038 per million tokens, while reducing its hosted context window from 64K to 24K. The price falls 57.8%; the advertised context falls 62.5%.
  • Cloudflare says only 0.24% of requests exceed 24K input tokens. It does not publish the observation period, request count, customer distribution or method behind that figure, so it cannot show how many customers or workflows fit the cheaper model.
  • Workers AI’s per-token prices are comparison units within a Neuron-based billing system. A buyer still needs its own request mix, fallback rate and decision-quality results to estimate cost per accepted decision.

Analysis

A lower price usually reads as a wider door. Cloudflare’s 9 October change to Workers AI’s Clef family comes with a narrower one. Clef-flash now costs $0.038 per million input tokens instead of $0.090, but its hosted context window has moved from 64K tokens to 24K. The reduction in price is 57.8%; the reduction in the advertised window is 62.5%.

That pairing makes this more than a discount. It sets a boundary around the work Cloudflare wants to sell at the lowest displayed rate. The model family is designed for typed decisions: a customer sends a state and a set of allowed answers, and the model returns scores or probabilities rather than a paragraph. For a repeated classifier or router, the amount of state sent on each call is therefore central to the bill.

Hosted model Displayed price per million input tokens Context window
Clef-flash $0.038 24,576 tokens
Clef $0.240 65,536 tokens
Clef-omni $0.150 64,000 tokens

The SKUs do not form a quality ranking. Clef-omni accepts text, images, audio and video; the standard Clef page lists a separate 64K context model; Clef-flash’s current page specifies 24,576 tokens. A team whose state exceeds the flash limit could shorten or segment it, route the request to another hosted model, or operate released weights itself. Cloudflare says those Clef-flash weights remain unchanged and support up to 256K context when self-hosted. That option shifts serving, reliability and utilization work to the customer; the public materials do not quantify those costs.

Cloudflare’s defence of the smaller hosted window is a demand statistic: it says just 0.24% of requests exceed 24K input tokens. That is a request-level share, not a customer-level one. The company has not published the observation period, request denominator, distribution by account or workflow, or how it counted the long tail. A small number of heavy users could account for a meaningful share of customer workloads even if their requests are rare across the platform. The figure also cannot establish whether a request fits before media inputs and typed questions consume context.

The competing list prices show how sharp the entry rate looks, with limits. Cloudflare lists Clef-flash at $0.038 per million input tokens; TypeSafe lists Jev at $0.042, with output free. Clef-flash is 9.5% lower on that input-token measure. The comparison says nothing about equal task quality, latency, availability or cost per correct decision. It is a price reference, not a head-to-head result.

There is another meter behind the displayed one. Cloudflare says its pricing page presents model rates in token-equivalent units, while Workers AI accounts are billed in Neurons. For Clef-flash, the page maps each million input tokens to 3,455 Neurons; paid usage above the daily free allocation is priced at $0.011 per 1,000 Neurons. That makes the published token price useful for comparing model rates, but the buyer’s account-level cost still depends on measured Neuron use and the shape of each request.

The relevant commercial test is consequently not whether $0.038 is small in isolation. It is whether a buyer’s production requests remain inside 24K after the state, questions and media are counted; how often the application retries or falls back; and whether the resulting decisions are good enough for the action they trigger. Cloudflare has disclosed a platform-wide request statistic, not customer adoption, workload economics, margins or lower bills. Until those are visible, the price cut is a product offer with a clear constraint, not evidence of realized savings.

Sources