Bug: Nested pricing.anthropic above_200k fields cause 2x overcharge for Claude models with >200K context

[sususu98]

Bug Description

Claude models (opus-4-6, sonnet-4-6, sonnet-4-5, etc.) are being charged at ~2x the correct rate when the request's total input context exceeds 200K tokens. This affects all token types: input, output, cache_read, and cache_creation.

Root Cause

The TOML price table (prices-base.toml) contains nested [models."claude-xxx".pricing.anthropic] sections that include input_cost_per_token_above_200k_tokens and similar fields with doubled rates (e.g., input: $10/MTok vs standard $5/MTok for opus-4-6).

The pricing resolution chain in pricing-resolution.ts uses mergePriceData() which spreads the nested provider pricing onto the base price data:

// mergePriceData: {...base, ...pricingNode}
// This adds above_200k fields from pricing.anthropic to the resolved price data

Then in cost-calculation.ts, the cost calculation hits this priority chain:

if (longContextPricing && ...) {           // P1: explicit long_context_pricing (usually null)
} else if (inputAboveThreshold != null) {  // P2: above_200k fields ← BUG HERE
} else if (context1mApplied) {             // P3: hardcoded 2x/1.5x fallback
} else {                                   // P4: standard rate
}

Since modern Claude models (opus-4-6, sonnet-4-6) have flat pricing regardless of context length, the above_200k fields should not exist. But because they're present in the nested TOML pricing sections, they get merged into the resolved price data and trigger the P2 path.

Impact

• Affected models: claude-opus-4-6, claude-sonnet-4-6, claude-sonnet-4-5, claude-sonnet-4-5-20250929, and all regional variants
• Affected requests: Any request where input_tokens + cache_creation_input_tokens + cache_read_input_tokens > 200,000
• Overcharge rate: ~96-100% (nearly 2x the correct cost)
• Additional issue: Even if above_200k fields are removed, the context_1m_applied fallback path (P3) still applies hardcoded 2x input / 1.5x output multipliers for Codex provider type, causing ~5-50% overcharge

Example (claude-opus-4-6)

Token Type	Standard Rate	above_200k Rate (from TOML)	Overcharge
input	$5/MTok	$10/MTok	2x
output	$25/MTok	$37.5/MTok	1.5x
cache_read	$0.5/MTok	$1/MTok	2x
cache_creation_5m	$6.25/MTok	$12.5/MTok	2x

TOML Source

The problematic fields are in nested pricing sections like:

[models."claude-opus-4-6".pricing."anthropic"]
input_cost_per_token = 0.000005
output_cost_per_token = 0.000025
input_cost_per_token_above_200k_tokens = 0.00001      # ← should be removed
output_cost_per_token_above_200k_tokens = 0.0000375    # ← should be removed
cache_read_input_token_cost_above_200k_tokens = 0.000001
cache_creation_input_token_cost_above_200k_tokens = 0.0000125

Suggested Fix

Option A: Remove above_200k from TOML (data fix)

Remove *_above_200k_tokens fields from all pricing.anthropic / pricing.opencode / pricing.vertex_ai sections for Claude models that have flat pricing.

Option B: Code-level guard (defensive fix)

In cost-calculation.ts, skip the above_200k pricing path when the above_200k rate equals the standard rate, or add a model family check.

Option C: Use long_context_pricing with multiplier 1.0

Add explicit long_context_pricing config with multiplier 1.0 for flat-pricing models, so the P1 path is hit with no surcharge.

Workaround

Create source='manual' price records with above_200k fields set equal to standard rates. Manual records have highest priority in findLatestPriceByModel and won't be overwritten by TOML sync.

[coderabbitai]

🔗 Related PRs

#545 - docs: fix Codex model config in usage-doc [merged]
#664 - fix(billing): 修复 Gemini 图片生成模型的 IMAGE modality token 计费问题 [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[github-actions]

Thank you for this detailed and well-researched bug report.

Analysis:

I have verified the issue through codebase analysis. Your root cause analysis is correct.

Confirmed findings:

1. Pricing resolution merges above_200k fields: The mergePriceData() function in src/lib/utils/pricing-resolution.ts (lines 187-207) spreads provider-specific pricing nodes onto base price data:

   return {
     ...base,
     ...pricingNode,  // Includes above_200k fields
     pricing: base.pricing,
     selected_pricing_provider: pricingProviderKey,
   };

2. Cost calculation priority chain: The calculateRequestCostBreakdown() function in src/lib/utils/cost-calculation.ts (lines 527-553 for input tokens) uses this priority:

   • P1: longContextPricing (explicit long_context_pricing config)
   • P2: inputAboveThreshold (above_200k/above_272k fields) - BUG: triggers incorrectly
   • P3: Hardcoded 2x/1.5x multipliers (only when context1mApplied=true)
   • P4: Standard rate

3. Constants confirm the behavior: src/lib/special-attributes/index.ts (lines 15-23) defines:

   • CONTEXT_1M_TOKEN_THRESHOLD = 200000
   • CONTEXT_1M_INPUT_PREMIUM_MULTIPLIER = 2.0
   • CONTEXT_1M_OUTPUT_PREMIUM_MULTIPLIER = 1.5

4. Price data source: The cloud price table is fetched from https://claude-code-hub.app/config/prices-base.toml via src/lib/price-sync/cloud-price-table.ts (line 4).

Affected code paths:

• src/lib/utils/cost-calculation.ts:527-553 - Input token cost calculation
• src/lib/utils/cost-calculation.ts:557-583 - Output token cost calculation
• src/lib/utils/cost-calculation.ts:588-617 - Cache creation 5m cost calculation
• src/lib/utils/cost-calculation.ts:652-673 - Cache read cost calculation
• src/lib/utils/pricing-resolution.ts:187-207 - Price data merging

Verification needed:

To confirm the TOML pricing data contains the problematic above_200k fields for flat-pricing Claude models, could you provide:

• The specific model entries from the TOML showing the above_200k fields
• Any log output showing the resolved price data for affected requests

---

Automated response from Claude AI

[github-actions]

This issue has been automatically marked as stale because it has not had any activity in the last 30 days.

If this issue is still relevant:

• Please comment to keep it open
• Add any new information that might help resolve it

This issue will be automatically closed in 14 days if there is no further activity.

[github-actions]

This issue has been automatically closed due to inactivity.

If you believe this issue is still relevant, please feel free to reopen it with additional information.