Put three inexpensive rows in one table and the cheapest input price looks like the winner. It is not a winner; it is a different product. At the prices on the listings when this page was updated, the input rates span roughly a factor of eighteen and the output rates a similar range. The rows also differ in modality, license, and hosting story, and those differences eliminate options before price gets a vote.

This comparison is for high-volume work: classification, extraction, short summaries, tagging, routing, and the many calls that make up an agent's inner loop. It is not for the one call where a wrong answer is expensive. For that, use a frontier row and route by consequence.

Price per token is not price per answer

The first correction is the unit. A cheap row that needs a longer prompt to hold a format, or a second attempt on one request in five, is not as cheap as its rate. Multiply the per-token price by the tokens you actually send and by the attempts you actually need.

Three quantities change the ranking:

  • Input length. High-volume work often sends a long instruction or a retrieved document with a short answer. Input price dominates.
  • Output length. Summaries and drafts are output-heavy; classification is not.
  • Retry rate. A row that fails the schema check is a row you pay for twice.

Measure all three on a sample of real requests before you commit. A rate table cannot tell you which row will need the retry.

License decides where it can run

Gemini 3.7 Flash is a hosted, proprietary row. If the data cannot leave your network or cannot be processed by that vendor, the row is out, regardless of price. If it can, the hosted price includes the operations you would otherwise own.

GLM-5.2 and DeepSeek-V4-Flash-0731 are listed as open weights under permissive licenses. That does not make them free: self-hosting means GPUs, a serving stack, upgrades, monitoring, and someone on call. At low utilization the hosted rate is usually cheaper; at steady high utilization the local row can win. The break-even is an operations question, not a token question.

Read the license file rather than the badge. "Open weights" means the parameters are downloadable; the license says what you may do with them. The open source and proprietary guide covers the distinction.

Modality and context are constraints

The Gemini row lists text, image, file, audio, and video input. The other two list text-oriented transformers. If the task includes a scanned invoice, a screenshot, or a call recording, the modality field decides before the price does. Do not build an OCR pipeline to avoid a model that already accepts the image.

All three advertise context windows above a million tokens. Treat that as capacity, not quality. Accuracy tends to fall as input grows, and a detail in the middle of a very long prompt is recovered less reliably than the same detail near the end. For high-volume work, the cheaper and more accurate pattern is to retrieve the relevant passage and send less.

Where each row belongs

  • Gemini 3.7 Flash belongs where the input is multimodal, the vendor is an accepted processor, and the team does not want to operate a serving stack. It is the broadest input surface of the three.
  • GLM-5.2 belongs where an open license matters, the workload is text, and the team can either self-host or accept a hosted endpoint. Its output price is close to Gemini's; its input price is higher.
  • DeepSeek-V4-Flash-0731 belongs where the work is text and volume is high enough that the input rate dominates. It is the lowest-priced row here, which makes it the one to test first for bulk classification and extraction.

Those are placements, not verdicts. A placement says which constraint the row satisfies; it does not say the row will pass your accuracy bar.

What to test

Build a twelve-task set from real high-volume work, as described in test a model on your own tasks. Include the format check and the refusal case, not just the happy path. Run each row twice, score blind, and record tokens, latency, and retry rate.

Then compute cost per accepted answer, not cost per call. That number, multiplied by monthly volume, is the only price comparison worth putting in a document. The token cost guide has the worksheet, and the comparison tool will rebuild this table against any rows you choose.