[BUG] API Cost for OpenRouter Computed Incorrectly

[sawka]

Problem (one or two sentences)

For the past couple weeks I've been seeing much higher API Cost numbers in Roo. At first I thought it was because of the change to use the tool call APIs. But now I'm convinced it is just a miscalculation of API Cost in Roo itself.

I ran a simple task in Roo, and I also pulled all the data from OpenRouter for that task. Net result is Roo claims $0.36 API Cost, but the sum of the 6 OpenRouter requests sums to ~$0.18 (off by almost a factor of 2x).

I then summed up all the tokens and types in OpenRouter and compared with Roo's totals. The input/output/cached tokens matched exactly.

But, the API Cost was computed as if there was zero caching discount (and also zero extra cost for writing to the cache).

Context (who is affected and when)

OpenRouter users.

Reproduction steps

1. Run a simple task in Roo, and observe the API Cost Total
2. Look at the corresponding requests and costs in OpenRouter and sum those totals
3. Compare, and see that they are different

Expected result

The API Cost in Roo matches the API Cost in OpenRouter

Actual result

The API Cost in Roo is much higher than the API Cost in OpenRouter

Variations tried (optional)

No response

App Version

3.47.2

API Provider (optional)

OpenRouter

Model Used (optional)

anthropic/claude-sonnet-4.5

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[hannesrudolph]

Update to 3.47.3 and your api costs will drop significantly. We fixed a bug this morning. Sorry about that.

[hannesrudolph]

Thanks for the report. The 3.47.3 fix mentioned above should have already improved things for you.

Beyond that, we are actively reimplementing the OpenRouter provider to use the official ai-sdk provider package. This will move cost/usage reporting to be derived directly from the provider response rather than computed locally, which eliminates this entire class of discrepancy (cache discount pricing, etc.). That work is underway now.

Going to close this one out as the immediate fix is already shipped and the longer-term solution is in progress.

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause

The bug is in src/api/providers/openrouter.ts at line 531. The current cost calculation is:

totalCost: (lastUsage.cost_details?.upstream_inference_cost || 0) + (lastUsage.cost || 0)

This adds upstream_inference_cost and cost together. For non-BYOK (Bring Your Own Key) OpenRouter users, cost already includes the full amount (upstream provider cost + OpenRouter margin). Adding upstream_inference_cost on top double-counts the cost, producing roughly 2x the actual cost -- which matches the issue report exactly ($0.36 vs $0.18).

This was introduced in commit 1a9aa13b (PR #5145) which replaced the old is_byok flag with cost_details but lost the distinction between BYOK and non-BYOK users.

Proposed Fix

In src/api/providers/openrouter.ts, change the cost calculation to handle both BYOK and non-BYOK scenarios:

const upstreamCost = lastUsage.cost_details?.upstream_inference_cost || 0
const orCost = lastUsage.cost || 0
// For BYOK: upstream_inference_cost > cost (OpenRouter margin is much smaller than provider cost)
// For non-BYOK: cost already includes upstream + margin, so cost >= upstream_inference_cost
const totalCost = upstreamCost > orCost ? upstreamCost + orCost : orCost

This uses a heuristic: if upstream_inference_cost > cost, it's a BYOK user (where cost is just OpenRouter's small fee and upstream_inference_cost is the provider charge billed to the user's own key). Otherwise, cost already includes everything.

Files to change:

1. src/api/providers/openrouter.ts -- fix the cost calculation
2. src/api/providers/__tests__/openrouter.spec.ts -- add tests for both BYOK and non-BYOK cost scenarios

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.