Gemini 3.1 Flash Lite is the cheapest Gemini model on every task I compared, but DeepSeek V4.1 Flash is cheaper still on proofreading, email replies and product descriptions. On the CSV task, DeepSeek's low cost comes from a failed run, so it is not a real saving. Gemini 3.8 Flash costs the most per job and has the highest LMArena Text score of the four.
The four models on price and setup
All four models share the same 1,048,576-token context window, so context length does not separate them. The differences are price and reasoning mode.
| Model | Input / output per 1M tokens | Reasoning | LMArena Text |
|---|---|---|---|
| DeepSeek V4.1 Flash | $0.072 / $0.14 | Default | 1475 |
| Gemini 3.1 Flash Lite | $0.25 / $1.50 | Default | Not listed |
| Gemini 3.5 Flash Lite | $0.30 / $2.50 | Always | 1455 |
| Gemini 3.8 Flash | $0.82 / $4.12 | Always | 1496 |
Reasoning tokens are billed as output tokens, so a model that thinks at length costs more even when its list price is low. That matters for the Gemini models that always reason, and for DeepSeek on the CSV task.
Proofreading and email replies
Proofreading a 500-word page has a median of 0.25¢ per task across all 336 models I track.
| Model | Per task | Per 1,000 tasks | Basis |
|---|---|---|---|
| DeepSeek V4.1 Flash | 0.04¢ | $0.43 | Measured |
| Gemini 3.1 Flash Lite | 0.18¢ | $1.82 | Estimated |
| Gemini 3.5 Flash Lite | 0.30¢ | $2.95 | Estimated |
| Gemini 3.8 Flash | 0.51¢ | $5.10 | Measured |
DeepSeek used 1,954 reasoning tokens on this task. Each Gemini model used 443. The Gemini models reasoned less, but their output price is much higher, so they still cost more per job.
Customer email replies are cheaper again, with a median of 0.10¢ per task.
- DeepSeek V4.1 Flash: 0.03¢ per task, $0.30 per 1,000 (measured).
- Gemini 3.1 Flash Lite: 0.09¢ per task, $0.90 per 1,000 (estimated).
- Gemini 3.5 Flash Lite: 0.15¢ per task, $1.46 per 1,000 (estimated).
- Gemini 3.8 Flash: 0.25¢ per task, $2.50 per 1,000 (measured).
The ranking is the same on both tasks: DeepSeek, then Gemini 3.1 Flash Lite, then Gemini 3.5 Flash Lite, then Gemini 3.8 Flash. Only the DeepSeek and Gemini 3.8 Flash figures come from direct runs. The two Flash Lite figures are estimates, so treat them as approximate. See the proofreading and email reply task pages for the full lists.
Product descriptions
Writing 10 product descriptions costs more because the output is longer. The median is 0.60¢ per task.
- DeepSeek V4.1 Flash: 0.04¢ per task, $0.41 per 1,000 (measured). 870 output tokens and 1,842 reasoning tokens.
- Gemini 3.1 Flash Lite: 0.45¢ per task, $4.49 per 1,000 (estimated).
- Gemini 3.5 Flash Lite: 0.74¢ per task, $7.45 per 1,000 (estimated).
- Gemini 3.8 Flash: 1.2¢ per task, $12.39 per 1,000 (measured). 1,127 output tokens and 1,817 reasoning tokens.
The gap is wider here than on proofreading: $0.41 versus $12.39 per 1,000 tasks between DeepSeek and Gemini 3.8 Flash. The higher benchmark score of Gemini 3.8 Flash may justify that on high-stakes copy. For bulk catalog text, the cost gap is hard to ignore. More models are listed on the product descriptions page.
CSV analysis: read this one carefully
This task is a 200-row spreadsheet analysis. The median is 2.1¢ per task.
- DeepSeek V4.1 Flash: 0.26¢ per task, $2.65 per 1,000 (measured, hit the output cap).
- Gemini 3.1 Flash Lite: 1.4¢ per task, $13.75 per 1,000 (estimated).
- Gemini 3.5 Flash Lite: 2.2¢ per task, $22.19 per 1,000 (estimated).
- Gemini 3.8 Flash: 3.9¢ per task, $38.69 per 1,000 (measured).
The DeepSeek figure looks like a win, but it is not a fair comparison. In my run, DeepSeek hit the 16,000-token output cap with 0 visible output tokens. It spent its whole budget on reasoning and returned no analysis. The low cost reflects a failed answer, not a cheap successful one.
All three Gemini models produced a visible answer, with 1,018 output tokens each. Before using DeepSeek V4.1 Flash for analysis like this, test whether it returns an answer on your own prompts. The spreadsheet task page has the full ranking.
Which family to pick for each volume tier
I am not setting volume bands, because the per-1,000 figures above already show the gap at any scale. Use them with the calculator for your own inputs.
- High volume text work (proofreading, email replies, product descriptions): DeepSeek V4.1 Flash had the lowest cost per job on all three tasks. If its quality fits your copy, it is the default choice.
- High volume where you want a Google model: Gemini 3.1 Flash Lite is the cheapest Gemini option and sits second on every text task.
- Low volume, quality-sensitive work: Gemini 3.8 Flash has the highest LMArena Text score of the four (1496) and the highest cost per job. At low volume the absolute difference is small, so use it where a wrong answer costs more than the price gap.
- Spreadsheet analysis at any volume: Start with Gemini 3.1 Flash Lite. DeepSeek's CSV run produced no usable answer, so its low cost does not count as a saving.
For how I produce these numbers, see the methodology.
FAQ
Is DeepSeek V4.1 Flash always the cheapest option?
In my measurements it was the cheapest of the four on proofreading, email replies and product descriptions. On the CSV task it hit the output cap without a visible answer, so its low cost is not a real saving there.
Why does Gemini 3.8 Flash cost so much more per job?
Its list price is the highest of the four, at $0.82 input and $4.12 output per 1M tokens. It also always reasons, and reasoning tokens are billed as output tokens.
Which Gemini model should I start with?
Start with Gemini 3.1 Flash Lite. It was the cheapest Gemini model on every task I compared, and its reasoning mode is default rather than always on.