Cost control
Direct integration optimizes for time-to-first-call. ozDNA optimizes for margin at scale — routing, cache, and per-workflow attribution.
Illustrative only — Numbers on this page are hypothetical examples for education and planning. They are not measured production benchmarks. Public, reproducible benchmarks are planned for Q3 2026.
Benchmarks
ozDNA vs
direct provider
How a unified gateway compares to wiring OpenAI (or another provider) directly in application code — on latency, cost control, and operational depth.
Illustrative scenario
Hypothetical B2B SaaS running POST /v1/detect in academic mode — ~50,000 requests/month, mixed EN/TR text, single-region deployment. Used only to structure the comparison below.
- Direct provider: App → OpenAI API (single model, custom retry logic, logs to stdout)
- ozDNA: App → ozDNA Gateway → Model Router → Prompt Registry → provider
Metric comparison
Illustrative* = example numbers for the scenario above, not ozDNA production telemetry.
| Metric | Direct provider | ozDNA | Notes |
|---|---|---|---|
| Latency Illustrative* | ~850 ms p50 | ~880 ms p50 | Example: small gateway hop vs direct call. Your p99 depends on model and region. |
| Token cost Illustrative* | ~$0.042 / 1k req | ~$0.031 / 1k req | Example: blended cost if router sends 40% of traffic to a cheaper capable model. Not a guarantee. |
| Cache hit Illustrative* | — | ~28% example | Prompt-hash / response cache at gateway — requires identical inputs. Direct integration: build yourself. |
| Retries Illustrative* | Manual | Automatic | Example: fallback chain on 429/5xx. Direct: retry loops in app code per provider SDK. |
| Observability | DIY | Integrated | Request IDs, audit events, structured logs — vs stitching provider dashboards and app logs. |
| Model routing | Single model | Policy-based | Route by workflow, cost ceiling, or latency budget without redeploying app code. |
| Prompt versioning | Git / env vars | Registry | Versioned prompts with rollback and environment scopes in Prompt Registry. |
| Cost analytics | Invoice + spreadsheet | Per-workflow | Attribute spend to workflow, account, and model — not only provider monthly bill. |
Where ozDNA adds value
Operational depth
Retries, observability, and prompt versioning are product features — not weekend engineering projects bolted onto a provider SDK.
Latency trade-off
A gateway adds a hop. Illustrative example shows ~30 ms p50 delta — acceptable when routing and cache savings matter more than raw minimum latency.
Methodology & honesty
- We do not publish production customer metrics on this page.
- Illustrative numbers use round, clearly labeled examples — not cherry-picked wins.
- Public benchmarks will document workload, region, model IDs, and reproducible scripts.
- Production proof today: live workloads on TezMakale — qualitative, not quantified here.
Findbelow Ventures operates ozDNA. Provider names (OpenAI, Anthropic, etc.) describe integration targets — not endorsements or measured comparisons against those vendors.
Run your own comparison
Sandbox access lets you measure latency and cost on your corpus — the only benchmark that matters for your product.