# Model reasoning

Source: https://provod.ai/en/docs/reasoning

*Model reasoning*

Reasoning controls depend on the model and API format. Check catalog `supported_parameters` and the selected model documentation first.

Chat Completions uses `reasoning_effort`; Responses uses `reasoning.effort`. Allowed values and constraints can differ. Do not apply one model setting to every model. Omitting the parameter does not necessarily disable reasoning.

Reasoning tokens may count as paid output even when their text is hidden. Allow sufficient output tokens and budget: use a supported `max_tokens` or `max_completion_tokens` in Chat Completions, or `max_output_tokens` in Responses. Do not send both Chat Completions limit names together. Compare the actual response accounting and request record in Usage, rather than estimating cost from visible text length.

Cached input and reasoning output are different categories. Do not add categories again when they are already included in total token counts. Another platform's public cache controls are not parameters of this API.

[Catalog](/en/docs/catalog) · [Responses](/en/docs/responses) · [Costs](/en/docs/usage-costs)

## Send a request with explicit effort

Requires Bash, `curl`, `jq`, an `AI_API_KEY` and an exact `MODEL_ID`. Choose an available model whose contract confirms `reasoning_effort` and the value `low`. Do not discover support by repeating paid requests. Running the example makes one paid request.

```bash
set -euo pipefail
: "${AI_API_KEY:?Set the API key for your workspace}"
: "${MODEL_ID:?Choose a model with documented low reasoning effort support}"
API_BASE="${API_BASE:-https://api.provod.ai}"

REQUEST_JSON="$(jq -n --arg model "$MODEL_ID" '{
  model: $model,
  messages: [{role: "user", content: "What is 17 multiplied by 19?"}],
  reasoning_effort: "low"
}')"
RESPONSE_JSON="$(curl --fail-with-body --silent --show-error "$API_BASE/v1/chat/completions" \
  -H "Authorization: Bearer $AI_API_KEY" \
  -H "Content-Type: application/json" --data-binary "$REQUEST_JSON")"

jq -e '
  if (.choices[0].message.content | type) != "string"
  then error("No text response; inspect the response") else
    {text: .choices[0].message.content,
     finish_reason: .choices[0].finish_reason,
     usage: .usage}
  end
' <<<"$RESPONSE_JSON"
```

Read `finish_reason` and actual `usage`: an absent reasoning breakdown does not mean zero cost. A `length` finish reason means the result was limited and can be incomplete. For Responses use the separate `reasoning: {"effort": "low"}` field only when the selected contract supports it.

## Troubleshooting


**Changing effort produces an error**


Remove unsupported parameters and check the selected model and API format contract. Do not search for valid values with a series of automatic paid requests.
