Image generation models (gemini-3-pro-image-preview) billing calculation is incorrect

[sususu98]

Summary

The current billing logic does not correctly handle image generation models like gemini-3-pro-image-preview. The system treats IMAGE output tokens the same as TEXT tokens, resulting in ~7.6x undercharging.

Price Configuration

[models."gemini-3-pro-image-preview"]
mode = "image_generation"
input_cost_per_token = 0.000002
output_cost_per_token = 0.000012
output_cost_per_image = 0.134
output_cost_per_image_token = 0.00012

Example usageMetadata

{
  "promptTokenCount": 326,
  "candidatesTokenCount": 2340,
  "totalTokenCount": 3003,
  "promptTokensDetails": [
    { "modality": "TEXT", "tokenCount": 326 }
  ],
  "candidatesTokensDetails": [
    { "modality": "IMAGE", "tokenCount": 2000 }
  ],
  "thoughtsTokenCount": 337
}

Current vs Expected Billing

Item	Current Calculation	Expected Calculation
Input TEXT	326 × $0.000002 = $0.000652	326 × $0.000002 = $0.000652
Output IMAGE	2000 × $0.000012 = $0.024	2000 × $0.00012 = $0.24
Output TEXT	(included above)	340 × $0.000012 = $0.00408
Thoughts	337 × $0.000012 = $0.00404	337 × $0.000012 = $0.00404
Total	$0.0328	$0.2488

Difference: ~7.6x undercharging

Root Cause

1. response-handler.ts: extractUsageMetrics does not parse candidatesTokensDetails to distinguish IMAGE vs TEXT modality
2. cost-calculation.ts: calculateRequestCost does not use output_cost_per_image_token field for IMAGE tokens

Suggested Fix

1. Extend UsageMetrics type to include output_image_tokens
2. Parse candidatesTokensDetails in extractUsageMetrics to extract IMAGE token count
3. Update calculateRequestCost to apply output_cost_per_image_token for image output tokens

[sususu98]

补充：Google 官方定价文档对照

官方定价（Gemini 3 Pro Image Preview）

类型	官方价格	说明
输入 text/image	$2.00 / 1M tokens	等同于 $0.0011 per image (560 tokens)
输出 text/thinking	$12.00 / 1M tokens	文本和思考 token
输出 image	$120.00 / 1M tokens	图片输出 token
1K/2K 图片	1120 tokens = $0.134	1024x1024 ~ 2048x2048 分辨率
4K 图片	2000 tokens = $0.24	最高 4096x4096 分辨率

来源：Google AI Pricing

---

TOML 价格配置验证

当前 LiteLLM 价格表中的配置：

[models."gemini-3-pro-image-preview"]
mode = "image_generation"
input_cost_per_token = 0.000002         # $2.00 / 1M ✓
output_cost_per_token = 0.000012        # $12.00 / 1M (text/thinking) ✓
input_cost_per_image = 0.0011           # 560 tokens = $0.0011 ✓
output_cost_per_image = 0.134           # 1K/2K 图片价格 ✓
output_cost_per_image_token = 0.00012   # $120.00 / 1M (image tokens) ✓

价格配置与官方文档完全一致，问题仅在于代码未正确使用这些字段。

---

实际案例分析

用户生成了一张 4K 图片，返回的 usageMetadata：

{
  "promptTokenCount": 326,
  "candidatesTokenCount": 2340,
  "promptTokensDetails": [{ "modality": "TEXT", "tokenCount": 326 }],
  "candidatesTokensDetails": [{ "modality": "IMAGE", "tokenCount": 2000 }],
  "thoughtsTokenCount": 337
}

计费对比

费用项	当前系统（错误）	正确计费	差异
输入 TEXT	326 × $0.000002 = $0.000652	326 × $0.000002 = $0.000652	-
输出 IMAGE	2000 × $0.000012 = $0.024	2000 × $0.00012 = $0.24	10x
输出 TEXT	(混在一起)	340 × $0.000012 = $0.00408	-
思考 token	337 × $0.000012 = $0.00404	337 × $0.000012 = $0.00404	-
总计	$0.0328	$0.2488	7.6x

---

需要修改的文件

1. src/types/model-price.ts

   • 已有 output_cost_per_image 字段定义
   • 需新增 output_cost_per_image_token、input_cost_per_image 字段

2. src/app/v1/_lib/proxy/response-handler.ts

   • extractUsageMetrics() 需解析 candidatesTokensDetails 和 promptTokensDetails
   • 区分 TEXT 和 IMAGE modality 的 token 数量

3. src/lib/utils/cost-calculation.ts

   • calculateRequestCost() 需支持：
     • input_cost_per_image 或按 image token 计费
     • output_cost_per_image_token 用于 IMAGE 输出 token

4. UsageMetrics 类型扩展

   • 新增 input_image_tokens?: number
   • 新增 output_image_tokens?: number

---

影响范围

此问题影响所有 mode = "image_generation" 的模型，包括：

• gemini-3-pro-image-preview
• 未来可能新增的图片生成模型

如果不修复，图片生成请求的成本统计将严重偏低，影响：

• 用户成本限额检查（可能超额使用）
• 供应商成本统计
• 排行榜数据准确性

[github-actions]

Thank you for reporting this issue with detailed analysis.

Analysis:

Based on my investigation of the codebase, I can confirm your findings are accurate:

1. Current Implementation:

The extractUsageMetrics function in src/app/v1/_lib/proxy/response-handler.ts (lines 1257-1399) currently handles Gemini usage metadata as follows:

• Line 1275-1284: Extracts promptTokenCount and candidatesTokenCount from usageMetadata
• Line 1281-1284: Maps candidatesTokenCount directly to output_tokens without distinguishing modality
• Line 1300-1303: Adds thoughtsTokenCount to output_tokens

The function does NOT parse candidatesTokensDetails array to distinguish between IMAGE and TEXT modalities.

2. Price Configuration:

The ModelPriceData type in src/types/model-price.ts (lines 1-54) includes:

• output_cost_per_image (line 22) - for per-image pricing
• But NO output_cost_per_image_token field

The calculateRequestCost function in src/lib/utils/cost-calculation.ts (lines 106-293) does not have any logic to handle image token pricing separately from text tokens.

3. Root Cause Confirmation:

Your analysis is correct:

1. extractUsageMetrics treats all candidatesTokenCount as regular output tokens
2. The system applies output_cost_per_token (/bin/bash.000012) instead of output_cost_per_image_token (/bin/bash.00012) for IMAGE modality tokens
3. This results in 10x undercharging for image tokens (0.000012 vs 0.00012)

4. Affected Code Locations:

• src/app/v1/_lib/proxy/response-handler.ts:1257-1399 - extractUsageMetrics function
• src/lib/utils/cost-calculation.ts:106-293 - calculateRequestCost function
• src/types/model-price.ts:1-54 - ModelPriceData type definition

Implementation Considerations:

To fix this issue, the following changes would be needed:

1. Extend UsageMetrics type to include separate fields for image tokens:

   • Add output_image_tokens?: number field

2. Update extractUsageMetrics to parse candidatesTokensDetails:

   • Iterate through the array and sum tokens by modality
   • Separate IMAGE tokens from TEXT tokens

3. Extend ModelPriceData to include:

   • Add output_cost_per_image_token?: number field

4. Update calculateRequestCost to handle image tokens:

   • Apply output_cost_per_image_token rate for image tokens
   • Keep existing output_cost_per_token for text tokens

Related Files:

The test file tests/unit/proxy/extract-usage-metrics.test.ts has comprehensive coverage for Gemini usage parsing (lines 311-344) but does not test candidatesTokensDetails parsing, which would need to be added.

---

Automated response from Claude AI