Choosing a model

GET https://api.brrrain.fun/v1/models (no key needed) lists every model you can use:

Field	Meaning
id	What you send as model
name	Display name
private	Runs in a TEE with attestation (NEAR AI private inference): the provider cannot read the prompt
context_length, max_output_tokens	Limits; max_tokens above max_output_tokens is refused by the model
reasoning	The model can think before it answers (see below)
price.input_per_1m, price.output_per_1m	Power per million tokens
price.cached_input_per_1m	Power per million cached input tokens, when the model has a cache price
price.typical_message	Power for a typical message (1 500 tokens in, 650 out)

default_model is the site's preselection: a cheap private model.

How to pick

1. Routine work (summaries, extraction, classification, simple code, chat): the default model. Messages left today ≈ power_left ÷ its typical_message.
2. Hard reasoning, long careful code, the human asked for a specific model: a frontier model. Check the price first: typical_message ranges from a few Power to several hundred, so one frontier message can cost as much as a hundred cheap ones.
3. Sensitive input: prefer private: true models.
4. Long inputs: check context_length; prices are per token, so a long context dominates the cost of every step.
5. Prices change with NEAR AI's catalog: read /v1/models, never a number remembered from earlier.

Models that can think first (reasoning: true)

They can write a chain of thought before the answer; some always do (GLM, DeepSeek, Qwen thinking models), others only when asked. That thinking is output tokens: it is billed and counts against max_tokens.

• Give them max_tokens of 1 000 or more, even for a one-line answer. A five-word answer can take ~400 tokens of thinking; with max_tokens: 100 the reply ends before any text and you still pay for it.
• Or leave max_tokens out: the API then sends the most you can afford, up to its default.
• An empty content with finish_reason: "length" means the thinking used it all: resend with a larger max_tokens, do not loop on the same one.
• They cost more than typical_message says, since it counts 650 output tokens; reckon two to four times it.
• For short routine replies, a model with reasoning: false is cheaper.

Messages left for a model = power_left ÷ price.typical_message (from GET /v1/me and GET /v1/models); show that to your human when they ask "how much do I have left".