POST /v1/chat/completions

The model to use for completion. The model value may include supported model suffixes, including web search (':online', ':online/<provider>'), memory (':memory', ':memory-<days>'), reasoning visibility (':reasoning-exclude'), thinking variants where listed by the model catalog (':thinking'), and provider routing preferences for eligible models (':fast', ':cheap', ':caching', etc.). Model ID or compatibility alias. Append :reasoning-effort/<effort>, where effort is none, low, medium, high, xhigh, or max, to supply a default reasoning_effort. The suffix is stripped before model routing and is not listed as a separate model. Explicit reasoning generation settings take precedence, including enabled thinking when the suffix is none. The none suffix requests disabled reasoning where supported; it does not bypass model/provider restrictions. Other effort levels remain model/provider-specific.

Billing override to force pay-as-you-go without an explicit provider, or to apply saved provider preferences to subscription-included traffic. Accepted values (case-insensitive): paygo, pay-as-you-go, pay_as_you_go, paid, payg.

Alias for billing_mode.

When true, route to an available provider that is marked as prompt/input-caching capable for the requested provider-selection model. If no usable cache-capable provider exists, the request fails instead of falling back to a non-cache-capable provider. This is provider capability routing only; it does not add cache_control markers, configure cache TTLs, or force a cache write. The model suffixes ':caching', ':cache', and ':cached' request the same behavior.

Top-level routing control for caching: true. When caching is true, defaults to true and prefers the previously recorded provider for later matching requests from the same API key or session when still usable. Set false to require a cache-capable provider without sticky routing.

CamelCase alias for top-level stickyprovider. Distinct from prompt_caching.stickyProvider, which controls explicit prompt-cache failover behavior.

Whether to stream the response

Optional processing tier. Use flex for lower-cost variable-capacity processing or fast for higher-priority processing where supported. priority remains accepted as a legacy alias for Fast on OpenAI models. Discover model support with GET /api/v1/models?detailed=true and inspect supported_service_tiers; the basic model list intentionally omits this field.

Classic randomness control. Accepts any decimal between 0-2. If omitted, NanoGPT does not force a value and the routed provider/model default applies

Upper bound on generated tokens. If omitted, NanoGPT does not enforce an explicit default and the routed provider/model default applies

Nucleus sampling. When set below 1.0, trims candidate tokens to the smallest set whose cumulative probability exceeds top_p. Works well as an alternative to tweaking temperature

Penalizes tokens proportionally to how often they appeared previously. Negative values encourage repetition; positive values discourage it

Penalizes tokens based on whether they appeared at all. Good for keeping the model on topic without outright banning words

Penalizes tokens that already appear in the conversation, on most providers including the prompt. 1 turns it off and values above 1 discourage repetition. Values below 1 encourage repetition and can cause misspelled words and endless loops, especially with long prompts. Unlike frequency_penalty and presence_penalty, the neutral value is 1, not 0

Caps sampling to the top-k highest probability tokens per step

Combines top-p and temperature behavior; leave unset unless a model description explicitly calls for it

Ensures each candidate token probability exceeds a floor (0-1). Helpful for stopping models from collapsing into low-entropy loops

Tail free sampling. Values between 0-1 let you shave the long tail of the distribution; 1.0 disables the feature

Cut probabilities as soon as they fall below the specified tail threshold

Cut probabilities as soon as they fall below the specified tail threshold

Typical sampling (aka entropy-based nucleus). Works like top_p but preserves tokens whose surprise matches the expected entropy

Enables Mirostat sampling for models that support it. Set to 1 or 2 to activate

Mirostat target entropy parameter. Used when mirostat_mode is enabled

Mirostat learning rate parameter. Used when mirostat_mode is enabled

For providers that support it, enforces a minimum completion length before stop conditions fire

When true, keeps the stop sequence in the final text. Not supported by many providers

Allows completions to continue even if the model predicts EOS internally. Useful for long creative writing runs

Extension that forbids repeating n-grams of the given size. Not supported by many providers

Requests logprobs on the prompt itself when the upstream API allows it

Optional integer forwarded on model/provider routes that support seeded sampling. This may improve reproducibility but does not guarantee identical output. Results can change if NanoGPT selects a different automatic or fallback route, or if the provider changes its backend.

Maximum number of built-in tool calls, such as native web searches, page opens and searches within a page, the model can make in one response. Applies to requests NanoGPT sends to OpenAI's Responses API, including GPT-5+ and o-series native search and your own OpenAI key (BYOK); calls past the cap are not run or billed. Other routes, and rerouted requests such as Auto Model, ignore it.

Controls reasoning depth and acts as an explicit reasoning-mode signal. Any value other than "none" requests reasoning/thinking behavior. Use "none" to explicitly disable reasoning.

Shorthand for reasoning.delta_field

Shorthand to force legacy reasoning_content fields in the response

Unique identifier for the completion

Object type, always 'chat.completion'

Unix timestamp of when the completion was created

Service tier used (echoed when provided on the request)