Value
November 2026. Your traffic with us, next to the same traffic without us. Sample data.
Kept away
16,770
requests that would have hit flagged providers
Tool calls saved
893
that would have broken
Better answers
2,157
that would have scored below the reference
Long prompts saved
366
that would have been cut off
With us and without us
48,210 requests · strict mode| Without us | With us | Difference | |
|---|---|---|---|
| Requests to flagged providers | 16,770 | 0 | 16,770 fewer |
| Requests to providers on watch | 12,540 | 0 | 12,540 fewer |
| Broken tool calls | 1,277 | 383 | 893 fewer |
| Answers below the reference | 2,157 | 0 | 2,157 fewer |
| Long prompts cut off | 366 | 0 | 366 fewer |
| Model spend | $72.70 | $84.97 | $12.26 more |
| Added latency | 0 ms | 22 ms | 22 ms slower |
Was it worth it?
$
- Failures avoided
- 3,415 × $0.25 = $854
- Model spend
- $12 more than without us
- Your plan
- $29
- Net this month
- $813 ahead
DeepSeek V4.1 Flash
26,400 requests · where they went| Provider | Status | Without us | With us |
|---|---|---|---|
| Provider A | allowed | 22% · 5,804 | 61% · 16,131 |
| Provider B | on watch | 29% · 7,763 | 0% · 0 |
| Provider C | flagged | 35% · 9,138 | 0% · 0 |
| Provider D | allowed | 14% · 3,694 | 39% · 10,269 |
Kimi K3
13,900 requests · where they went| Provider | Status | Without us | With us |
|---|---|---|---|
| Provider A | allowed | 20% · 2,811 | 55% · 7,663 |
| Provider E | allowed | 16% · 2,288 | 45% · 6,237 |
| Provider F | flagged | 29% · 4,025 | 0% · 0 |
| Provider G | on watch | 34% · 4,777 | 0% · 0 |
GLM 5.3
7,910 requests · where they went| Provider | Status | Without us | With us |
|---|---|---|---|
| Provider H | allowed | 26% · 2,029 | 47% · 3,731 |
| Provider I | flagged | 46% · 3,607 | 0% · 0 |
| Provider J | allowed | 29% · 2,273 | 53% · 4,179 |
How we estimate this
Without us is OpenRouter's default routing: it spreads requests across every provider for a model and favours cheaper ones (weight is one over price squared). We replay your real request mix through that split.
Failure rates come from our own checks on each provider: tool call validity, scores on known answer questions, and recall on long prompts. Your prompts are never stored or graded.
These are estimates, not counts. Some failures would have happened anyway, and we show those too in the With us column.