Reasoning and thinking
How Talos reasons before it answers, and why you should set a generous max_tokens
Talos is a reasoning model. Before it writes the answer, it works through the problem internally. By default talos runs in thinking mode, so most requests reason first and then reply.
Where the reasoning goes
The visible answer stays clean. The model keeps its reasoning separate from the content it returns, so choices[0].message.content holds the answer alone, not the thinking that produced it.
When reasoning runs, the API can return the reasoning trace separately in reasoning_content, and the usage object reports a reasoning_tokens count. Read these if you want the trace. Ignore them and you still get a clean answer in content.
Set a generous max_tokens
Thinking consumes output tokens. Each reasoning mode reserves a minimum generation budget, so the reasoning and the answer both draw from the max_tokens you set. Set it high enough for the reasoning to finish, or the request fails with an error rather than returning a partial answer.
Size max_tokens to the effort:
- fast and low/medium thinking: set it generously, a few hundred tokens or more, so the answer has room after the reasoning.
- high effort (
high/xhigh/max): set it high, 1024 or more. Higher effort raises the reasoning-token budget, so the model can reason longer, runs slower, and holds a concurrent-request slot longer. Plan for the higher latency.
You select mode and depth with reasoning_effort, not by switching models. The same max_tokens guidance applies to whichever effort you pick — max_tokens bounds the actual reasoning, so a small max_tokens at high effort still keeps the reasoning short.
from openai import OpenAI
client = OpenAI(
base_url="https://api.ablatic.ai/v1",
api_key="sk-ablatic-...",
)
resp = client.chat.completions.create(
model="talos",
messages=[{"role": "user", "content": "Plan a three-day trip to Vienna."}],
max_tokens=4096, # leave room for reasoning and the answer
)
print(resp.choices[0].message.content)On the Anthropic-style surface, max_tokens is required, and the same guidance applies: size it for both the reasoning and the answer.
Pick how much the model reasons
Use reasoning_effort to set the depth: none or minimal for fast (no reasoning), low for brief thinking, medium for the default, and high (or xhigh/max) to raise the reasoning-token budget. low measurably shortens the reasoning; effort above medium lifts the ceiling rather than switching mode. See Models for the full mapping.