---
title: "Model pricing and cost calibration"
description: "Configure per-million-token prices, pricing-model source, and multipliers, then reconcile local cost."
version: "en"
---

> Documentation Index
> Fetch the complete documentation index at: https://docs.ochub.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Model pricing and cost calibration

OcHub cost is a local estimate calculated from request tokens, model prices,
and a multiplier. It is not the provider's invoice. OcHub does not maintain its
own built-in price list. A request can succeed while showing 0 when model IDs
or pricing sources do not match.

## What determines cost

One request can have three model names:

| Name | Meaning |
| --- | --- |
| Request model | Model or alias sent by the client |
| Model | Model returned by or used upstream |
| Pricing model | Model OcHub uses to look up a price |

With gateway rewriting, all three can differ. For example, the client may request
`claude-sonnet-current`, routing may send
`anthropic/claude-sonnet-4-6`, and the pricing model should follow whichever
real ID matches billing.

The price table stores four USD rates per million tokens:

- Input
- Output
- Cache read
- Cache write, also called cache creation internally

OcHub adds these base costs and then applies the multiplier. It also handles
app-specific cache semantics: Claude-style input is usually fresh input,
whereas Codex / Responses input can include cache reads that must be subtracted
before applying the ordinary input rate.

## Price sources and synchronization

The app bundle includes an offline catalog that CI filters, validates, and pins
to a specific revision of LiteLLM's official data. First launch therefore does
not require the network, and the prices are not a separately maintained OcHub
list. OcHub makes at most one conditional background check per day. Use
**Sync now** under **Usage → Model pricing** to check immediately. If an update
fails, the last valid catalog remains active.

Price resolution follows these rules:

1. User-saved manual overrides always win and catalog sync never changes them.
2. The catalog first matches a complete provider or regional model key, then a
   direct model key.
3. An alias is used only when every matching catalog entry has the same four
   rates. OcHub does not guess when providers or regions disagree.
4. When a request used cache tokens but the catalog omits the corresponding
   cache rate, the missing value is not treated as `0`.

Models with tokens but no reliable price are grouped into one missing-price
notice instead of repeated dialogs. After catalog sync or a manual override,
OcHub attempts to backfill zero-cost historical requests with tokens. Existing
nonzero costs are not rewritten in bulk.

## Find the correct model ID

Do not guess from a marketing page:

1. Make a real request.
2. Open it under **Usage → Request logs**.
3. Record Request model, Model, and Pricing model.
4. Compare them with the provider's price list.
5. Add the ID that represents the model actually billed.

If Pricing model is a custom alias, choose among response-model pricing, adding
an equivalent alias price, or changing the route mapping. The most reliable
setup makes Pricing model equal the provider's priced model ID.

## Add a manual price override

1. **Expand Model pricing configuration** on the Usage page and find the manual
   override table.

2. **Enter the model ID.** It must match the pricing model; case, provider
   prefixes, and version suffixes may matter.

3. **Enter a display name.** It improves readability but does not affect
   routing.

4. **Enter all four rates.** Values are USD per million tokens. Use `0` when the
   provider's rules genuinely have no cache-read or cache-write charge.

5. **Save.** Prices must be non-negative numbers. Saving the same ID updates it.

6. **Refresh and inspect a request.** Verify its pricing model and itemized and
   total cost.

After a manual override is saved, OcHub attempts to backfill historical
requests for that model whose token count is nonzero but total cost is still 0.
Existing nonzero historical costs are not all recalculated, so a price change
can create a before-and-after reporting boundary.

> **Pricing changes do not alter provider billing**
>
> Backfill affects only the local OcHub usage database. It cannot change provider
> invoices, balance, or actual charges.

## Request model versus response model

Claude and Codex have a default pricing-model source:

| Choice | Use it when | Risk |
| --- | --- | --- |
| Response model | The upstream returns the actual model after aliases or rewriting | Some upstreams return no model or echo the alias |
| Request model | The client ID is stable and matches the price list | Routing may cause the wrong model price |

Compare both names on a direct request and a routed request, choose the source
that consistently finds the real price, save defaults, then verify the Pricing
model on a new request.

Claude Desktop inherits the Claude default multiplier and pricing source.

## Cost multiplier

The default multiplier `1` leaves base prices unchanged.

| Settlement rule | Multiplier |
| --- | --- |
| Same as list price | `1` |
| 20% markup | `1.2` |
| 20% discount | `0.8` |
| Track tokens but omit cost from totals | `0` |

The multiplier applies after input, output, and cache base costs are added. It
does not change token counts.

If a provider marks up models differently, enter settled per-model rates and
keep the multiplier at `1`.

## Check one calculation by hand

| Item | Tokens | Per-million price | Base cost |
| --- | ---: | ---: | ---: |
| Fresh input | 10,000 | $3.00 | $0.03000 |
| Output | 2,000 | $15.00 | $0.03000 |
| Cache read | 20,000 | $0.30 | $0.00600 |
| Cache write | 5,000 | $3.75 | $0.01875 |

The base total is `$0.08475`. At a `1.2` multiplier:

```text
$0.08475 × 1.2 = $0.10170
```

Use token values from one request detail, not the page-wide Actual tokens
summary. OcHub already normalizes Codex / Responses cache semantics.

## Reconcile with provider billing

When local estimates differ:

1. Use the same date range and timezone.
2. Filter to one provider.
3. Compare request counts, including upstream treatment of failures and
   retries.
4. Verify the pricing model.
5. Check whether system prompts, tools, or reasoning tokens are billed
   separately.
6. Verify input and cache semantics.
7. Check multiplier, discount, or currency conversion.
8. Reconcile one request before comparing daily totals.

Account plans, tiered prices, batch discounts, regions, or minimum charges may
not fit one model price and multiplier. Treat OcHub as a request-level estimate
and the provider invoice as the settlement record.

## Data-source limits

| Source | Usually includes | May omit |
| --- | --- | --- |
| Gateway request | Request ID, status, model, tokens, latency | Fields omitted by upstream |
| CLI session | Model, some tokens, session time | Gateway status and first-token time |
| Legacy local request | Historical tokens and cost | Current pricing-model semantics |
| Synchronized session | Fields in the original session file | Provider or detailed cache split |

OcHub cannot infer reliable cost when model or token data is absent.

## Common problems

| Symptom | Check | Fix |
| --- | --- | --- |
| Tokens exist but cost is 0 | Missing-price notice, pricing model, and cache rates | Sync or override that exact ID |
| New requests are right, old ones are not | Old records already had nonzero cost | Treat the price change as a boundary |
| Cache cost looks too high | Input and cache read overlap | Confirm app and upstream semantics |
| Routed request prices the client alias | Source is Request model | Use Response model or price the alias |
| Cost changes suddenly | Multiplier, suffix, or price update | Compare pricing model and multiplier |
| Total differs slightly from invoice | Rounding, timezone, failures, plan rules | Reconcile individual requests first |
| Price will not save | Empty display name, negative, or invalid number | Enter a name and non-negative decimals |

After setup, verify one new request with a known model. See
[View sessions and usage](/guides/sessions-usage) for daily filtering and
drill-down.

Source: https://docs.ochub.org/advanced/pricing/index.mdx