OpenRouter vs direct OpenAI pricing: worksheet

See also / canonical cluster. Pillar for token & cost estimation: How to estimate LLM token costs before you ship. Interactive: token & cost estimator. This page stays on a different intent — see the intro.

Intent here: gateway vs direct all-in pricing and ops tradeoffs — not “how to estimate any LLM feature.” Base token math still follows the estimation pillar; use the token estimator for editable rates.

What you are comparing

Dimension	Direct vendor API	Gateway / router
List token prices	Vendor dashboard	Gateway price list (may differ)
Extra margin / fees	Usually none beyond tokens	Often baked into rates or billed separately
Model variety	That vendor’s catalog	Many vendors behind one key
Invoicing	One vendor relationship	One gateway relationship
Data processing terms	Vendor DPA	Gateway + upstream — read both
Debugging	Vendor error shapes	Extra hop, unified errors
Rate limits	Per vendor account	Gateway + upstream interplay

Your worksheet should have two scenarios with the same traffic shape.

Worksheet skeleton

Fix a traffic profile first:

• Model id(s) you actually call
• p50 / p90 input tokens, avg output tokens
• Calls / month, retry %, agent rounds
• Required regions / retention

Then columns:

1. Effective input $/1M and output $/1M (paste dated sources)
2. Estimated monthly token $ (use token estimator)
3. Minimum monthly fees / seat / top-up friction (if any)
4. Eng hours for integration + secret rotation
5. Legal review status (DPA, subprocessors)
6. Lock-in / exit notes (prompt formats, SDK assumptions)

All-in monthly ≈ token $ + fixed fees + (eng hours × loaded cost) / amortization months

For early experiments, eng hours dominate. For high volume, token $ dominates.

Price myths to avoid

• “Gateway is always more expensive” — not always; promotions, batching, and which model you land on matter.
• “Direct is always cheaper” — true only for the same model at the same rate with the same failure rate.
• “One key to many models is free complexity” — product and security still need an allowlist.

Always date-stamp the rate rows. Re-pull before quarterly planning.

Operational differences that change cost

• Retries: Extra hop can change timeout behavior; fold into retry budget.
• Fallbacks: Auto-routing to another model can save uptime and surprise you on quality/price — require allowlists.
• Caching / batch: Direct vendors may offer batch or cached-input tariffs gateways expose differently.
• Observability: If you cannot attribute spend per feature, unit economics fail regardless of vendor.

Decision patterns

• Single-vendor, high volume, strict DPA → direct often wins on clarity once legal is done.
• Multi-model experiments, small volume → gateway can reduce integration time; still allowlist models.
• Production with failover needs → either multi-direct clients you own, or a gateway with explicit failover policy and caps.

Write the choice as a reversible decision with a review date.

Security and compliance checklist (short)

• Subprocessors documented
• Retention defaults acceptable
• Key scoped and rotatable
• Logging redacts prompts if required
• Model allowlist enforced in code

Pair with your org’s vendor review — this article is not legal advice.

Practical next steps

1. Duplicate your traffic profile into two estimator runs with each rate table.
2. Add a line for eng-time amortization.
3. Document allowlisted model ids in the service config.
4. Revisit when either side changes rates or you cross a volume threshold (e.g. 10× calls).

The “winner” is the row with acceptable risk and the lower all-in cost for your traffic — not the loudest screenshot of a price page.

Sample all-in comparison (placeholders)

Assume 5M input + 1M output tokens/month on the same model class:

• Direct: list rates → $X token; eng already integrated; DPA done
• Gateway: list rates → $Y token; +2 eng days to productionize allowlist and logging; DPA in progress

If Y - X is smaller than the loaded cost of those eng days over three months and legal is acceptable, gateway can win short-term. If you are already integrated direct and volume is high, shaving $Y toward $X matters more. Put both stories in the worksheet with dates.

Exit plan

Whichever you pick, keep prompts and tool schemas vendor-neutral enough to move in a quarter. Record: SDK wrappers, model id mapping table, and who owns the cutover checklist. Pricing wins evaporate if exit takes a year.