Public audit

Routing transparency

Our responsible-AI policy commits us to auditing for systematic bias toward or against any provider. This page is that audit, live and anonymized — derived from the last 30 days of routing decisions. No prompts, no user data, no model names; only provider-level shares.

Updated Thu, 04 Jun 2026 23:28:24 GMT

Routed calls

143

Avg savings vs baseline

94%

Concentration (HHI)

3,226

concentrated

Provider share

HHI under 1,500 is considered balanced; 1,500–2,500 moderate; above 2,500 concentrated. Lower is better — it means no single provider dominates routing decisions.

anthropic50.3%
google18.9%
mistralai16.1%
meta-llama6.3%
deepseek4.9%
qwen3.5%

By task type

simple_qa45% of traffic
- anthropic78%
- mistralai11%
- meta-llama9%
- google2%
summarize26% of traffic
- google51%
- mistralai41%
- meta-llama5%
- anthropic3%
creative13% of traffic
- anthropic95%
- qwen5%
drafting7% of traffic
- google70%
- anthropic20%
- meta-llama10%
reasoning4% of traffic
- deepseek100%
long_form2% of traffic
- qwen100%
code1% of traffic
- qwen50%
- deepseek50%
extraction1% of traffic
- mistralai100%

How this is computed

Every assistant message is logged with the model the router picked and the task it classified the prompt as. We aggregate those rows over the last 30 days, group by provider, and compute the Herfindahl–Hirschman Index — the standard market-concentration measure used by the US DOJ.

Prompts, user IDs, and per-model breakdowns are deliberately excluded. If a provider's share drifts above ~60% for a given task without justification, the router's quality scores are re-weighted from real user feedback to rebalance. See our responsible-AI policy for the full commitment.