Platform capabilities
Reasoning Models
Choose reasoning models for complex analysis, mathematics, planning, and code tasks, and manage long runtimes and output budgets.
Reasoning models suit multi-step analysis, complex planning, mathematics, and difficult coding tasks. They share POST /v1/chat/completions with ordinary chat models; the difference is response time, output budget, and parameter behavior, not the interface.
Selection criteria
Reasoning and ordinary chat models share one interface; they differ in thinking cost and output budget. When choosing, confirm:
| Dimension | What to confirm |
|---|---|
| Context length | Long-chain tasks need larger input and output budgets. |
| Thinking cost | Whether the reasoning process is billed separately or counted in token usage. |
| Parameter behavior | Whether sampling parameters such as temperature are restricted or ignored. |
| Output ceiling | Whether the maximum output per request leaves room for the full derivation and the conclusion. |
Model identifiers come from the model catalog; this page does not list a fixed set. A minimal request to confirm availability before scaling input:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["APIQIK_API_KEY"],
base_url="<BASE_URL>/v1",
)
response = client.chat.completions.create(
model="<reasoning-model-id>",
messages=[
{
"role": "user",
"content": (
"Compare three caching strategies. Describe constraints, trade-offs, and a recommendation, "
"then list the metrics needed to validate that recommendation."
),
}
],
max_tokens=1800,
)
print(response.choices[0].message.content)When reasoning models are not needed
| Good fit for reasoning models | No need for reasoning models |
|---|---|
| Comparing options under complex constraints | Keyword classification or simple rewriting |
| Mathematics, logic, and multi-step analysis | Fixed-format template filling |
| Large code problems and root-cause analysis | Low-latency chat and autocomplete |
| Planning tasks that must validate assumptions | Results deterministic code can compute directly |
Calling notes
- The thinking process consumes tokens, so
max_tokensmust be generous; with512the common symptom is being cut off before the conclusion. - Some reasoning models restrict or ignore sampling parameters such as
temperature; do not rely on low temperature to stabilize conclusions. - Interactive products need an acceptable overall deadline per request; consider asynchronous or background queues for long tasks.
- Ask for conclusions, assumptions, and validation methods; do not equate longer answers with higher quality.
- Verify key conclusions with tests, calculators, retrieval, or business rules before persisting them.
Cost control
- Confirm a model is available with a small request before scaling input.
- Keep a smaller-model baseline for the same task class and upgrade to a reasoning model only when quality falls short.
- Record the model identifier, token usage, and latency of each request, and break costs down by business line.