Reasoning and cache affinity

Reasoning controls are model-specific. The gateway preserves the native controls for its current backend; it does not guarantee that every combination works on every model.

Protocol	Controls
Chat	`reasoning_effort`, `reasoning`, `include_reasoning`, and legacy `thinking`.
Responses	`reasoning` and `include_reasoning`.
Messages	`thinking` and `output_config.effort`.

Chat and Responses accept reasoning effort values documented in the API reference, including none, minimal, low, medium, high, xhigh, max, and ultracode. The active model decides which are meaningful. Prefer one control style in a request.

const result = await client.chat.completions.create({
  model: process.env.LIMITLESS_MODEL,
  messages: [{ role: "user", content: "Explain your answer briefly: what is 17 × 23?" }],
  reasoning_effort: "low",
  max_tokens: 1024,
});
console.log(result.choices[0]?.message.content);

For Messages, thinking.type can be enabled, disabled, or adaptive, with a model-supported budget_tokens when appropriate. Reasoning can consume output tokens before visible text appears. A short visible answer or empty content field does not imply zero output cost.

Cache affinity

Send a stable X-Limitless-Session-ID for requests in one conversation. X-Abliteration-Session-ID is also accepted for client compatibility and takes precedence if both are supplied. Chat and Responses can use prompt_cache_key as a fallback.

const result = await client.chat.completions.create({
  model: process.env.LIMITLESS_MODEL,
  messages: [{ role: "user", content: "Summarize the context supplied above." }],
  max_tokens: 512,
}, {
  headers: { "X-Limitless-Session-ID": "conversation-123" },
});

The hint is limited to 512 characters and is isolated to your organization and project before it reaches the backend. Reusing a hint across different customers does not join their cache namespace. A hint neither stores conversation history nor guarantees a cache hit. Continue to send the needed messages and inspect provider-reported cache usage. Cached input uses the configured cached-token rate. Messages cache-read tokens are discounted; cache-write tokens use the normal input rate. See billing.

Start here

Build

Administer

Operate

Reasoning and cache affinity

Cache affinity

​Cache affinity

Cache affinity