Gemini 3.5 Flash-Lite vs DeepSeek V4.1 Flash: The Cache Decides Which Is Cheaper

Same input price, very different bill

Gemini 3.5 Flash-Lite and DeepSeek V4.1 Flash both list $0.30 per million input tokens, so they look like close rivals on a price table. Flash-Lite's output costs about twice as much. Its cached input costs five times as much. That last line turned out to matter more than the others.

We ran Flash-Lite through our open LLM cost per task benchmark from November 2 to 4, 2026. The tasks are checked by plain code: extraction, JSON, SQL, tool calls, bug fixes, regexes and questions over a long document. A failed answer gets up to two retries and every attempt counts toward the cost. The DeepSeek numbers come from our earlier run of the same tasks. Everything is at list price. Current rates are on the Flash-Lite pricing page.

The result, with and without one task

Taken as a whole, DeepSeek wins clearly: Flash-Lite cost 2.6x as much on everyday tasks and 2.7x on hard ones. Take out the long-document task and the order flips.

One task made up 74% of Flash-Lite's everyday bill and 75% of its hard bill. It sends a handbook of about 20K tokens with every question. DeepSeek served 98.8% of that input from cache and paid $0.00023 per correct answer. Flash-Lite got one hit in 50 calls, 1.6% of the input, and paid $0.00604, about 26 times as much. The answers were only a few tokens long, so nearly all of that money was input.

On short prompts Flash-Lite was cheaper despite the higher output price, because it writes far less. It reported no reasoning tokens at all. On the 9 shared everyday tasks it wrote 25,733 output tokens against DeepSeek's 101,129, and 81% of DeepSeek's were reasoning you never see. Bug fixes show it best: $0.00027 per correct answer on Flash-Lite, $0.00078 on DeepSeek.

Task by task the two split almost evenly, 7 to Flash-Lite and 8 to DeepSeek. DeepSeek tended to win where prompts are long and repeat, such as tool schemas and JSON schemas. Flash-Lite won bug fixes, record cleanup and regexes, where DeepSeek reasons at length.

What the bill looks like at other hit rates

Since the free key may have held the hit rate down, here is the same run repriced. The repriced bars are arithmetic, not measurements: same tokens, same answers, with a different share of the long-document input billed at the cached rate.

At DeepSeek's hit rate Flash-Lite would be the cheaper model, $310 against $347 on everyday tasks and $560 against $596 on hard ones. To tie, it needs about 93% of the long-document input from cache on the everyday set and about 96% on the hard set. So the winner comes down to one number, and that is the number our run could not pin down.

A paid key may do better. Google's caching guide describes implicit caching as on by default, gives no promise of a hit, and suggests sending requests with a similar prefix in a short amount of time. Our free quota spaced requests about 7.5 seconds apart. Google also lists a storage price of $1.00 per million tokens per hour for this model. We did not test explicit caching, and the numbers above leave that fee out.

What to do with this

• Short prompts and no big repeated context: Flash-Lite looks like the better buy, about 27% cheaper in this run.
• A large document or system prompt on every call: check your own hit rate before you decide. Your provider's usage report shows cached tokens per call, and a few hundred real requests will tell you more than our run can.
• Comparing any two models: price the cached input line, not only the input line. Here it swung the bill by nearly three times.

To turn your own numbers into a comparison, use the model switch cost calculator with your measured success rate and cached share.

Limits of this test

It is one run per model, on synthetic tasks, in English. DeepSeek's prices here are its Peak rates, and off-peak hours bill half of that. Flash-Lite missed 6 of 270 hard items, all on one regex task that both DeepSeek models passed, and its hard-tier cost per correct answer already includes those misses. The tasks do not cover long agent runs or open-ended writing.

The raw results list every task, and the runner is open source. If you have a paid Gemini key, the two long-document tasks cost well under a dollar to run, and that one number would settle this. Open an issue with your result.

FAQ