# How to track LLM cost on OpenRouter > A router can serve a different model than you requested. Culpa prices the model that actually answered, which is the only figure that matches the bill. URL: https://getculpa.com/track-llm-cost-openrouter Last reviewed: 2026-08-02 ## Answer A routing layer can answer with a different model than the one you requested, so pricing the request rather than the response produces a number that matches nothing. Culpa, a local-first LLM cost, margin, and forecast ledger, reads the model the response says served the call and prices that one, which is the only version of the arithmetic your invoice agrees with. ## Why this happens On a direct provider the model you ask for is the model you get, so nobody separates the two ideas. A router breaks that quietly. Availability, fallbacks and routing rules can all put a different model behind your call, and the response says which one answered. A ledger keyed on the request would price a month of traffic against a model that never ran, and it would look internally consistent while doing it, because every row would agree with every other row and none of them with the bill. ## What this usually looks like - Your routed spend disagrees with the invoice by an amount that moves month to month. - Nobody checks whether the model that answered matches the model requested. - A fallback is configured and nobody has measured how often it fires. - Your cost per call shifts without any change to prompts or volume. ## Common mistakes - Pricing routed traffic against the model you requested. Why it hurts: The response can name a different model, and the invoice follows the response rather than the request. Do instead: Read the served model from the response and price that. It's the only key the bill agrees with. - Leaving a fallback unmeasured. Why it hurts: A fallback that fires often is a pricing change nobody approved and nobody can see. Do instead: Count how many calls were served by something other than your first choice, monthly. - Assuming routed and direct calls of the same model reconcile identically. Why it hurts: Upstream providers on a router price the same model differently, so the spread lands in your ledger. Do instead: Compare your routed cost per call against the direct rate and treat the difference as a decision. ## Self-check - For last month, count calls where the served model differs from the requested one. - Check that your cost records key on the model that answered rather than the one you asked for. - Price the same month both ways. The difference is what keying on the request would have cost you. - Look at whether your fallback fires on a schedule, which usually means capacity rather than failure. - Compare your routed unit cost against the direct rate for the same model. ## One month priced on the request, and on the response (illustrative) 1 million routed calls a month, each sending 4,000 input tokens and returning 600. The request names Claude Opus 4.7 at $5.00 and $25.00 per million. A share of the traffic is served by Claude Sonnet 4.6 at $3.00 and $15.00 instead. Volumes and the fallback share are modelled. Priced as Opus 4.7: (4 x $0.005) + (0.6 x $0.025) = $0.020 + $0.015 = $0.035 per call Priced as Sonnet 4.6: (4 x $0.003) + (0.6 x $0.015) = $0.012 + $0.009 = $0.021 per call All 1,000,000 calls keyed on the request: $35,000 With 30% actually served by Sonnet: (700,000 x $0.035) + (300,000 x $0.021) = $24,500 + $6,300 = $30,800 Keying on the request overstates by $4,200, and every row still agrees with every other row A ledger can be perfectly self-consistent and $4,200 away from the bill. On a router the served model is the only key that closes the loop. ## Cost figures Every figure carries its confidence and its source. No figure on this site is provider-reported. - $30,800 to $35,000 — modelled monthly routed spend priced on the served model against priced on the requested one [estimated] Source: Both endpoints from the teardown arithmetic at real Claude Opus 4.7 and Sonnet 4.6 rates from the price book, effective 2026-07-02. A range because call volume and the fallback share are modelled. - $4,200 — modelled overstatement from keying a month of routed cost on the requested model [calculated] Source: $35,000 minus $30,800, from the teardown arithmetic at real price-book rates. ## FAQ Q: Can OpenRouter serve a different model than I asked for? A: A routing layer decides which upstream answers, and the response names the model that did. That's why a ledger has to read the served model rather than trusting the request, and why the two can differ across a month without anyone changing code. Q: Why does my routed cost per call keep moving? A: Two reasons compound. The model that answers can change, and the same model can be served by upstream providers charging different rates. Both move your unit cost without any change on your side unless you pin explicitly. Q: How do I make routed spend reconcile? A: Key every cost row on the model that answered, then compare the monthly total against the invoice. If it still drifts, the remaining gap is usually the price spread between upstream providers serving the same model. ## Sources - OpenRouter models: https://openrouter.ai/models - Anthropic pricing: https://platform.claude.com/docs/en/docs/about-claude/pricing Run the free Cost Leak Scan: https://app.getculpa.com/scan?source=pseo&slug=track-llm-cost-openrouter&cluster=integration Machine-readable index of every guide: https://getculpa.com/api/pages Human-readable index of every guide: https://getculpa.com/guides Site overview: https://app.getculpa.com/llms.txt Privacy: Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.