The workload
A solo developer runs a multi-model AI pipeline — heavy agent-style inference with large context windows (avg ~168K input tokens per request), spanning 11 models from GLM 5.2 to Claude Sonnet 5.
Scale: 21,739 requests · 3.61B input tokens + 13.05M output tokens · 11 models · 53 days (Jun 28 – Aug 19, 2026)
Prior bill
At OpenRouter list prices, the identical token mix would have cost:
| Model | Requests | Input tokens | Qubax paid | OpenRouter list |
|---|---|---|---|---|
| GLM 5.2 | 20,273 | 3.39B | $212.69 | $3,314.89 |
| GPT-5.6 Sol | 444 | 67.6M | $50.27 | $172.23 |
| Kimi K3 | 614 | 84.0M | $41.82 | $255.78 |
| Claude Sonnet 5 | 115 | 35.6M | $18.15 | $71.91 |
| GPT-5.6 Luna | 267 | 30.8M | $1.16 | $6.23 |
| Other (6 models) | 26 | — | $0.04 | $0.42 |
| Total | 21,739 | 3.61B | $324.13 | $3,821.47 |
Savings: $3,497.34 (91.5%)
Migration time
Under 1 hour. The workload moved by swapping one base URL and one API key — no SDK changes, no code changes. Every model kept the same OpenAI-compatible request shape.
Reliability
- 0 errors recorded across the full 53-day window (error_events: zero entries)
- 100% success rate in 671 traced requests (avg 9.7s TTFT on long-context agent calls)
- Platform-wide 30-day completion success: 100% across 2,923 sampled requests
What the workload looks like month over month
| Month | Qubax paid | OpenRouter equivalent | Saved |
|---|---|---|---|
| Jul 2026 | $107.22 | $628.52 | 82.9% |
| Aug 2026 (19 days) | $216.88 | $3,192.85 | 93.2% |
Savings increased as the workload scaled — August ran 93% below list as volume discounts deepened.
Verification
Every figure above is computed from actual request logs (tokens in/out, billed micros) cross-referenced against public OpenRouter list prices at the time of each request. No estimates, no annual-contract pricing, no rounded-up marketing math.