Skip to content
Last updated

Synapses and Neurons

Freddy uses synapses and neurons to measure AI model computation and processing.

What are Synapses?

Synapses are Aitronos' unit of measurement for AI model output generation. They represent the computational units consumed during response creation, including:

  • Output synapses - The visible text, images, or structured data in the response
  • Reasoning synapses - Internal thinking and reasoning computations (when using reasoning models)
  • Tool execution synapses - Processing for function calls and tool use

What are Neurons?

Neurons measure the input processing capacity consumed by your requests, including:

  • Input neurons - Processing your text, images, audio, and file inputs
  • Context neurons - Thread history and conversation context
  • System neurons - Instructions and system prompts

Why Synapses & Neurons?

Unified measurement: Track both input processing (neurons) and output generation (synapses) separately.

Reasoning transparency: Accounts for the model's internal reasoning process, not just visible output.

Fairer pricing: You're charged for the actual computational work performed by the AI.

Granular control: Set limits on both input context and output length independently.

Synapse Limits

Setting maxOutputSynapses

Control the maximum length and cost of responses:

{
  "maxOutputSynapses": 2048,
  "inputs": [...]
}

Common limits:

  • 512 synapses - Short answers, simple queries
  • 2048 synapses - Standard responses (default)
  • 4096 synapses - Detailed explanations, complex reasoning
  • 8192 synapses - Long-form content, multi-step tasks

What happens when limit is reached?

The model stops generating when:

  1. The response is naturally complete, or
  2. maxOutputSynapses is reached

You'll receive a finish_reason indicating why generation stopped:

  • stop - Natural completion
  • length - Hit synapse limit
  • tool_call - Model called a function

Usage Tracking

Every response includes synapse and neuron usage breakdown:

{
  "usage": {
    "inputNeurons": 150,
    "outputSynapses": 423,
    "reasoningSynapses": 89,
    "totalNeurons": 150,
    "totalSynapses": 512
  }
}

Use this data to:

  • Optimize prompts - Reduce input neurons
  • Control costs - Monitor total consumption
  • Debug responses - See reasoning overhead
  • Improve performance - Identify expensive operations

Understanding the Measurement

Synapses (Output)

  • Each piece of generated content consumes synapses
  • Reasoning models use additional synapses for internal thinking
  • Tool calls and function executions count toward synapses
  • Limit with maxOutputSynapses parameter

Neurons (Input)

  • Processing text, images, audio, and files consumes neurons
  • Conversation history and thread context count as neurons
  • System instructions and prompts use neurons
  • Automatically managed based on model context limits

Best Practices

✅ Start with defaults - 2048 synapses works for most cases

✅ Monitor usage - Check usage in responses to optimize

✅ Set limits - Use maxOutputSynapses to prevent runaway costs

✅ Optimize inputs - Reduce neuron consumption with concise prompts

❌ Don't over-limit - Too low limits may cut off important responses

❌ Don't ignore usage data - Track patterns to optimize costs

Related: