MiniMax M3.1 Flash vs DeepSeek V4.1 Flash: same rate, 8x the cost
Short version: for everyday prompts, writing and Japanese text, use DeepSeek V4.1 Flash. MiniMax M3.1 Flash thinks at length before it answers, and on our 14 prompts it produced about 8 times the output tokens. Both bill output at the same rate, so it cost about 8 times as much per answer. Its reasoning may earn its keep on the long agentic jobs MiniMax built it for; one-shot questions are not that.
The two models
Both are low-cost Flash models, and both bill output at the same rate on FastMetal.
| MiniMax M3.1 Flash | DeepSeek V4.1 Flash | |
|---|---|---|
| Model id | minimax-m3.1-flash | deepseek-v4.1-flash |
| Input per 1M tokens | $0.1583 | $0.211 |
| Output per 1M tokens | $0.633 | $0.633 |
| Context | 1,048,576 | 1,048,576 |
| Image input | yes | yes |
| PDF input | no | yes |
On the price sheet, M3.1 Flash looks slightly cheaper because of its input rate.
The test
We took 14 of our shared comparison prompts and sent each one to both models on October 20, 2026: 8 in English, 6 in Japanese. They covered debugging, SQL optimization, data interpretation, translation, word problems and Japanese proofreading. Both were called with streaming and max_tokens set to 32,768.
| 14 prompts | DeepSeek V4.1 Flash | MiniMax M3.1 Flash |
|---|---|---|
| Output tokens, total | 13,787 | 112,226 |
| Output tokens, median | 829 | 2,672 |
| Cost, total | $0.0091 | $0.0715 |
| Time to first visible text | 0.3 to 1.5 s | 3.5 to 402 s |
The gap is reasoning. DeepSeek V4.1 Flash does not reason by default and started writing within 1.5 seconds on every prompt. M3.1 Flash reasoned every time before writing: 402 seconds of nothing on the screen for the SQL prompt, and over two minutes for code review and data interpretation. Reasoning bills as output, so the 14 answers cost 7.8 times as much. Per answer that is about $0.00065 against $0.0051: small either way, but it scales directly with call volume.
What the answers looked like
- Word problems, in English and Japanese: both correct. M3.1 Flash added practical notes, such as how many whole hours to run a machine.
- Debugging: both found the cause and gave the right fix.
- Japanese proofreading: DeepSeek V4.1 Flash caught the duplicated character in ございいます and explained every correction in a table. M3.1 Flash missed that typo, flagged a correct phrase as a grammar error, and its explanation carried words from other languages (Korean 현대적인, a Chinese fragment).
- Translation: the second half of M3.1 Flash's answer was its own English working notes ("No, need final copy. Let's ensure…").
If you publish Japanese text straight from the model, check M3.1 Flash's output first.
Which to use
- DeepSeek V4.1 Flash for chat replies, drafting and summarizing, Japanese proofreading, and anything latency-sensitive. Cost and wait per answer stay small.
- MiniMax M3.1 Flash for multi-step jobs where a long think is acceptable. MiniMax reports 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2; our 14 one-shot prompts do not measure that.
On the API, call M3.1 Flash with "stream": true and a generous max_tokens. FastMetal allows up to 32,768 output tokens on it when you leave max_tokens out.
Try it
Switching is a change of model name on the same key:
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4.1-flash",
"messages": [{"role": "user", "content": "Fix the typos in this paragraph: ..."}]
}'
Both models' answers to all 43 shared prompts are side by side on the comparison page, and on each model page: MiniMax M3.1 Flash, DeepSeek V4.1 Flash. Rates are on pricing.