Model logging & data policies

This page explains how AI Code Router handles prompts, completions, and logs when you call our supported models. We sit in front of multiple providers and never train on your prompts; our job is to route traffic and surface usage, not to build foundation models.

Each model you can call through AI Code Router ultimately runs on a third‑party provider’s infrastructure. Those providers have their own terms for logging and retention. We mirror those policies at a high level so you can choose the right model for your use case.

Our model lineup

Today we expose nine “latest generation” models. Older or legacy variants are not available through the public router — we always point you at the current flagship or widely adopted release from each lab:

• GPT‑5.2 — OpenAI’s general‑purpose flagship, optimized for complex reasoning, long context (256K tokens), and code. Best default choice when you are unsure which model to pick.
• GPT‑5 mini — A smaller OpenAI model tuned for latency and cost. Ideal for chatbots, high‑traffic assistants, and background tasks where response time matters more than peak capability.
• Claude 4 Sonnet — Anthropic’s balanced model for production apps. Strong reasoning and writing quality while remaining affordable enough for large workloads.
• Claude 4 Opus — Anthropic’s highest‑end model. Use it when you need the deepest reasoning, analysis, or long‑form writing and are willing to trade some cost for quality.
• Gemini 3.1 Pro — Google’s general‑purpose model with very long context (up to 2M tokens). A good fit for retrieval‑augmented apps and document‑heavy workflows.
• Gemini 3.1 Flash — A fast, budget‑friendly Gemini variant. Great for streaming UIs, autocomplete‑style assistants, and any workload where you care about throughput.
• Llama 4 405B — Meta’s latest open‑weights Llama release, hosted for you by providers. Useful when you want strong performance with more flexibility around regions and deployment.
• Mistral Large 2 — Mistral’s production‑grade model, competitive on code and multilingual tasks. Often a good alternative when you want non‑US hosting options.
• DeepSeek V4 — A cost‑efficient model that shines on reasoning‑heavy workloads and tools. Popular for agents, analysis, and experimentation.

How we log requests

AI Code Router stores minimal metadata required to operate the service and show you usage in the dashboard:

• Per‑key usage — We track total tokens and request counts per API key and per model so you can see which keys are active and how much they are used.
• Timestamps and status — We log when a request happened, which model it used, and whether it succeeded, timed out, or failed.
• No training on prompts — We do not train models on your prompts or completions. Training policies are controlled by the underlying provider (OpenAI, Anthropic, Google, etc.).

Provider logging & regions

Because each model is hosted by a different lab, logging and retention policies vary. In general:

• OpenAI (GPT‑5.2, GPT‑5 mini) — Follows OpenAI’s enterprise and API data‑handling terms. You can choose whether data is retained for abuse monitoring only or longer for debugging, depending on your OpenAI agreement.
• Anthropic (Claude 4 Sonnet, Claude 4 Opus) — Uses Anthropic’s Claude API policies for retention and training. Requests are processed in the regions documented by Anthropic.
• Google (Gemini 3.1 Pro, Gemini 3.1 Flash) — Uses Google’s Gemini API policies. Some workloads can be pinned to specific regions; check the Gemini documentation for exact options.
• Meta, Mistral, DeepSeek (Llama 4 405B, Mistral Large 2, DeepSeek V4) — Hosted by providers that surface their own terms. We route your traffic according to those terms and display which provider you’re using on the models page.

For current details, always refer to the provider’s official documentation and terms. AI Code Router does not override or weaken those guarantees; we simply forward your requests and surface usage in one place.

To explore every model and its routing details, see the models page. For request/response formats and example code, visit the API Reference and SDK pages.