Reasoning models

One set of thinking controls, mapped to each model's native format.

Controls

The gateway accepts these across providers (OpenAI, Anthropic, Bedrock, DeepSeek) and translates them to each model's native parameters.

Parameter	Type	Description
`reasoning_effort`	`string`	How much the model should think: `low`, `medium` or `high`.
`thinking`	`object`	Anthropic-style thinking configuration, for example a token budget.
`reasoning`	`object \| boolean`	Generic reasoning switch or configuration for providers that use it.
`max_tokens`	`integer`	Reasoning tokens count toward this limit, so leave headroom beyond the visible answer.

curl https://api.eggsplain.com/v1/chat/completions \
  -H "Authorization: Bearer $EGGSPLAIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "sovereign/gpt-6-luna",
  "reasoning_effort": "high",
  "max_tokens": 4000,
  "messages": [
    {
      "role": "user",
      "content": "How many primes are below 100?"
    }
  ]
}'

Tips

Reasoning tokens count toward output, so raise max_tokens.
Stream to see the thinking trace as it is produced.
Models without reasoning support ignore or reject these fields; check List models.