Article

Choose Clairic or Clairic-flash before December 1

For teams extracting invoice or receipt fields, compare the December input rates, repeated billing, shared limits and quality trade-offs.

Editorial illustration for Choose Clairic or Clairic-flash before December 1: a document represents the research briefing. Not documentary evidence.

Cloudflare’s November 18, 2026 update gives teams extracting invoice, receipt or support-ticket fields a choice before Clairic’s December 1 price rise. Consider Clairic-flash if its results meet your needs at US$0.10 per million input tokens; keep Clairic when its results justify US$0.24 from December 1. Tokens are units used to count model input, including text and images—not documents.

What changes on December 1?

The existing @cf/cloudflare/clairic endpoint—the model address your application calls—is being upgraded in place to the 27-billion-parameter model. The separate @cf/cloudflare/clairic-flash endpoint provides the updated 9-billion-parameter option. Cloudflare says existing Clairic applications receive the upgrade automatically; retaining the old address does not preserve the old model.

Model

Until December 1, 2026

From December 1, 2026

Clairic

US$0.10 per million input tokens

US$0.24 per million input tokens

Clairic-flash

US$0.10 per million input tokens

US$0.10 per million input tokens

These are Cloudflare’s published input rates. Neither model charges for output tokens. Clairic’s rate becomes 2.4 times its promotional rate—a 140% increase—but that does not mean the whole application bill rises by 140%.

December 1 is a price-change date, not an announced shutdown deadline. The announcement does not specify a cutover hour or timezone.

When is Clairic worth the higher rate?

Use extraction quality on your own records as the deciding factor. Flash is the lower-rate candidate when it meets your requirements for correct fields, missing values and complete records. Keep Clairic when a comparison on representative documents shows that it avoids enough errors or correction work to justify the extra charge.

Cloudflare’s internal evaluations favor Clairic on some field-level checks and show ties on others. They are not independent evidence of either model’s accuracy on your invoices. The company’s speed figures also came from a self-hosted server, not a Workers AI response-time guarantee.

Compare cost per accepted document, including retries and correction work, rather than token price alone. A correctly formatted amount can still be the wrong amount; define which errors require review before choosing a model.

How do you estimate the input charge?

Counting the original document once can understate usage. The model-page billing guidance describes separate processing passes for text fields, with source text and images counted again on each pass. Requesting more fields can therefore add input charges even though output is free.

There is an unresolved detail: the model-page prose calls usage.passes a count of processing passes, while the output schema describes sequential decoding steps. Do not reconstruct an exact bill from that counter.

The schema defines usage.input_tokens as aggregate prompt tokens, including retries. For a gross estimate, sum that reported usage across requests—including additional requests your application retries—and apply the model’s rate for the relevant date. Do not multiply the aggregate by passes again or add its included retry tokens a second time.

Gross inference cost = total reported input tokens ÷ 1,000,000 × the applicable input rate.

For an illustrative workload, assume 100,000 documents each consume 10,000 aggregate input tokens on either model, all after the price change. That assumed count already includes fields, media, repeated processing and retries: 100,000 × 10,000 = one billion input tokens.

At the published rates, the gross amounts are US$240 for Clairic and US$100 for Flash, a US$140 difference. Equal token use is an assumption, not an established result. These amounts exclude free allocations, account charges, taxes, other services and human review; they are not final invoices or promised savings.

That US$140 is the useful comparison: would Clairic avoid more than US$140 of correction work or error cost on those records? The vendor’s evaluation results cannot answer that for your application.

Under ordinary Workers billing, the daily free allocation is 10,000 Neurons, Cloudflare’s compute billing unit—not 10,000 tokens. It is shared across account usage and resets at 00:00 UTC. Paid excess costs US$0.011 per 1,000 Neurons. Do not deduct a fresh allowance for each model or turn unused daily allocations into a monthly pool.

Usage beyond that allowance requires Workers Paid, whose minimum account charge is US$5 a month. An existing paid account does not need another minimum charge for each model. Reconcile the estimate with your dashboard, and check AI Gateway prepaid billing or enterprise arrangements separately.

If extracted records then feed an agent, downstream input is a separate cost. The guide to keeping large JSON out of AgentCore prompts covers reducing repeated tool-result input while retaining access to the full evidence.

Which document limits do both models share?

Paying for the larger model does not buy a larger document allowance. The Clairic input reference and Flash input reference specify the same limits:

  • Four images or rendered PDF pages in total. This is not permission to send four PDFs of arbitrary length.

  • Media-size caps. Each embedded image or PDF is limited to 4 MiB, with 8 MiB of decoded media combined and a 16 MiB request-body cap. Remote media URLs and encrypted PDFs are not accepted.

  • Shared context capacity. Both advertise 65,536 tokens, while their field specifications set a 65,273-input-token prompt limit. Source material, images and field definitions must fit; that is not a source-text-only allowance.

A five-page document therefore needs page selection or multiple requests, with results reconciled afterward. Preserve page identifiers and cross-page relationships: dropping a page to fit the limit may also drop the evidence needed for a total or line item.

For a different document workflow, the guide to selecting pages and spreadsheet cells with the Gemini API explains how to narrow model input. It is not an interchangeable Clairic interface.

The Clairic documentation also gives conflicting accounts of oversized text: its billing explanation says text can be shortened to fit, while its parameter definitions say input is never shortened and oversized requests are rejected. Until that is clarified, enforce explicit input-size limits and handle failures; do not rely on automatic shortening or assume that every submitted character was processed.

What should you check before switching?

  • Change both model selectors where applicable. The Flash examples use @cf/cloudflare/clairic-flash as the endpoint and model: "clairic-flash" in the request. Review both settings in your integration.

  • Check missing and incomplete answers. The current response schema exposes valid and complete metadata for generated answers. Its truncated flag describes invalid or incomplete generated output, not a reliable detector of shortened source input. A valid structure does not establish factual correctness.

  • Compare the same real task. Use representative document formats and important fields. Record aggregate input, retries, missing values, incorrect values and review effort on each model. Include every request needed to reach an accepted record.

Does downloading the weights avoid hosted charges?

The Clairic and Clairic-flash model cards identify Apache-2.0 releases. Running those weights yourself is a separate deployment option, with its own infrastructure and operating costs; it is not evidence that self-hosting will be cheaper.

That licence also does not replace the hosted service’s terms, which make no accuracy or continued-availability guarantee. For a hosted application, the immediate decision remains straightforward: choose the model whose accepted results justify its total cost, and budget for Clairic’s December 1 price change if you keep it.