LONG AI
Home › Blog › When Should You Use a Reasoning Model? 6

When Should You Use a Reasoning Model? 6 Models, 8 Questions, Measured

Category:选购指南 · 8 min read · 2026-11-10

Conclusion first: six models ran the same batch of 5 everyday questions - 30 calls, all correct - with the priciest charging ¥0.011027 and the cheapest ¥0.000346, a 31.9x spread. I added 3 questions that "look like they need thought" (make 24, the 12-coin balance puzzle, a 9-step chain) - still all correct. Across the questions I tested, the extra money bought more invisible thinking tokens and longer waits, not more accurate answers.

Method: same key, same prompts, no temperature set; timing from our server logs, amounts from actual charges in usage records, not estimated from unit prices. Every model name is searchable on the model plaza; prices follow the plaza in real time.

1. Five everyday questions: all correct, 32x the cost

Model5-question charge (¥)Total timeCorrect
gemini-3.8-flash0.01102767.4 s5 / 5
gemini-3.8-flash-high0.00423435.6 s5 / 5
glm-5.30.00358868.9 s5 / 5
deepseek-v4-pro0.00249858.8 s5 / 5
glm-5.3-flash0.00036611.1 s5 / 5
deepseek-v4.1-flash0.00034620.3 s5 / 5

The questions: extract a payable amount, rewrite in 30 words, a three-step calculation, a truth-lie logic puzzle, and reading code output.

Six models run the same 5 questions, all correct, charges from 0.000346 to 0.011027 yuan, a 31.9x spread
Fig 1 - Same questions, all correct, a 31.9x cost spread (log scale, which visually shrinks the gap). Data from our 2026-11-10 bills.

2. Per question: a 30-word rewrite, 56x apart

The widest gap was the rewrite:

Both answers are usable. The money differed 55.7x and the time 8.8x, and all you got was a difference in wording - the classic shape of a misused model isn't worse results, it's pure spending with zero return.

3. Where the money goes: thinking tokens are 99%

The same question output 7,052 tokens, of which 7,035 were reasoning_tokens - the actual answer was a dozen or so words. Thinking is billed at the output rate, and output is usually 5x the input price.

usage comparison across four models: gemini outputs 7052 tokens with 7035 thinking, deepseek has no reasoning_tokens field, minimal tier outputs only 13
Fig 2 - Same prompt, four usage shapes. deepseek has no reasoning_tokens field at all - its thinking hides inside completion_tokens - so scripts estimating cost by word count will be off by an order of magnitude.

And a more hidden leak: the input side. The same 88-character question logged 7,437 input tokens on glm-5.3-flash, while glm-5.3 on the same question took only 76. A model's built-in system prompt and tool definitions all bill at the input rate - that question cost ¥0.000336 on glm-5.3-flash vs ¥0.000167 on glm-5.3. Ten times cheaper per token, and the bill came out higher.

4. One parameter cuts the cost to about 1/240 of default

Every reasoning model exposes a thinking-budget knob - the catch is the parameter name differs per vendor, and a wrong name doesn't error, it's silently ignored while you keep paying. The same rewrite, run twice per setting, numbers reproduced stably:

gemini-3.8-flash-high parameterOutput tokensTimeCharge (¥)
reasoning_effort = minimal13 ~ 144.7 s0.0000326 ~ 0.0000339
reasoning_effort = low17 ~ 183.4 ~ 7.8 s0.0000381 ~ 0.0000395
reasoning_effort = high228917.0 s0.0031754
no parameter (default)570144.7 s0.0078868
reasoning_effort = medium785853.4 s0.0108653
Thinking tier scan: minimal charges 0.0000339, medium 0.0108653, a 320x spread; glm needs reasoning_effort=none to truly turn it off
Fig 3 - Left: gemini tier scan. Right: glm's two "turn off thinking" spellings. Same rewrite, most expensive vs cheapest differ 320x.

Two counter-intuitive points:

So don't trust the docs - send it and check the returned usage: output tokens dropping to single digits is the only proof it's really off.

5. The "hard" questions didn't separate them either

Three questions: make 24 from 3, 3, 8, 8 (answer 8/(3-8/3)); the 12-coin balance puzzle with unknown heavy/light (minimum 3 weighings); a 9-step chain (237). Not one fumbled. deepseek-v4.1-flash solved the chain in 81 tokens and 2.4s, and the make-24 in 111 tokens and 5.5s.

It was deepseek-v4-pro that spent 503 tokens and 10.7s on the same make-24, charging 14x for the identical answer. So "looks hard" is not a reason to turn on reasoning; what actually separates models is long chains, multiple constraints, and tasks that need repeated self-checking - and you can only find that by testing with your own questions.

6. A three-minute self-test

  1. Pull 10 real inputs from production - with dirty data, mixed lengths; not "1+1" style.
  2. Run cheap and expensive once each, same prompt, same input batch.
  3. Look at input tokens, output tokens, and actual charge in usage records, and score the answers at the same time.
  4. Keep only questions where "cheap gets it wrong, expensive gets it right"; if that's fewer than 2, your workload doesn't benefit from reasoning - a cheaper model plus one validation rule is the better deal.

Why 10: a single question can be luck; you need three repeats to check stability. Prices follow the model plaza in real time; the same model has different discounts per seller.

7. Common questions

1) Why is completion_tokens dozens of times the answer length?
Thinking tokens bill as output, but most providers don't put the thinking into the answer text. gemini lists reasoning_tokens separately; deepseek only has reasoning_content with the tokens merged into completion; glm is inconsistent. Seeing 7,052 with a 20-word answer is normal.

2) I set max_tokens, why is it still slow or the answer truncated?
Thinking counts against the max_tokens budget. Give it too small a budget and the model runs out mid-thought and can't produce the answer. Leave 3-5x headroom over your expected answer length.

3) I passed the "turn off thinking" parameter, why didn't the bill drop?
A wrong parameter name gets ignored. There's only one way to tell: check whether the returned usage's output tokens dropped. glm-5.3 accepts reasoning_effort=none but not thinking.type=disabled - measured on Nov 10.

4) Why is my input tokens so much longer than my prompt?
The model's built-in system prompt, tool definitions, and format constraints all count as input. This batch saw the same sentence at 76 and 7,437 input tokens - scripts estimating cost by word count must calibrate against the bill.

5) How to fall back when a cheap model occasionally gets it wrong?
Run the cheap tier first; on an obvious failure signal (empty output, JSON parse failure, "cannot determine") escalate to a reasoning tier and rerun once. 95% of requests take the cheap path, only 5% pay the expensive rate. How to switch models under one key - see the site docs.

8. Three routing rules

All amounts here are real charges from 2026-11-10. Prices move with sellers and promotions - check the plaza before ordering; a model's default thinking behavior may also change after updates, so rerun the self-test quarterly.

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa