Not every model should think the same way. DeepSeek R1 was the model that made reasoning-first inference famous: spend extra tokens working through the problem step by step, answer second. That idea won — so thoroughly that it stopped being a separate model. Today the same capability lives inside DeepSeek V4-Pro's thinking mode, and R1 itself is retired: the legacy deepseek-reasoner alias stops working on July 24, 2026. The judgment call this page covers — when deliberate reasoning is worth the cost — matters as much as ever; only the model name changed.
What a reasoning mode actually buys
A general pass answers in one shot. A reasoning pass generates an internal chain of thought first, then the final answer. That deliberation trades latency and output tokens for accuracy on problems where one wrong intermediate step ruins the result — which is why reasoning modes top the math and logic benchmarks.
The tradeoff is real: thinking is slower, and you pay for the thinking tokens. V4-Pro charges the same $0.435/$0.87 per million either way; what changes is how many tokens a hard question consumes.
When to turn thinking on
- Math and quantitative reasoning: multi-step calculations, proofs, word problems.
- Logic and planning: constraint problems, decision trees, working out the steps before acting.
- Careful analysis: cases where a wrong intermediate step poisons the answer and you want the model to check itself.
When to leave it off
- Simple, high-volume tasks: classification, extraction, short replies. Use DeepSeek V4-Flash — reasoning overhead there is wasted cost and latency.
- Latency-sensitive UX: chat that must feel instant answers faster without deliberation.
- Most coding: V4-Pro's standard mode (LiveCodeBench 93.5) is usually the better pick; save thinking for the genuinely hard algorithmic knots.
Calling it today
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["TURILOOP_API_KEY"],
base_url="https://api.turiloop.com/v1",
)
r = client.chat.completions.create(
model="deepseek-v4-pro", # thinking mode: the successor to deepseek-r1
messages=[{"role": "user", "content": "A train leaves A at 60 km/h..."}],
)
answer = r.choices[0].message.contentIf your code still references deepseek-r1 or deepseek-reasoner, migrate before July 24, 2026 — the V4 release tracker has the deadline details, and the migration guide has the mechanics.
The rule of thumb
Turn thinking on when the question is *hard to get right* and the cost of a wrong answer is high. For everything else, V4-Flash for cheap volume and V4-Pro's standard mode for quality are faster and cheaper. Both sit behind one OpenAI-compatible key on Turiloop, so routing each request to the right mode is a parameter, not a plumbing change. Pay-as-you-go, international card.