Missing cache_write cost on non-Anthropic models causes inconsistent cost tracking

[s2x]

Description

In models.config.mjs, Claude models define all four cost fields (input, output, cache_read, cache_write), but GPT, Gemini, and Kimi models omit cache_write:

// Claude — has cache_write ✓
"Claude Opus 4.6": {
  cost: { input: 5.5, output: 27.5, cache_read: 0.55, cache_write: 6.75 },
},

// GPT — missing cache_write ✗
"GPT 5.2": {
  cost: { input: 1.75, output: 14.0, cache_read: 0.175 },
},

// Gemini — missing cache_write ✗
"Gemini 2.5 Pro": {
  cost: { input: 1.25, output: 10.0, cache_read: 0.125 },
},

// Kimi — missing cache_write ✗
"Kimi K2.5": {
  cost: { input: 0.6, output: 3.0, cache_read: 0.1 },
},

Impact

When the downstream consumer (opencode) reads the generated opencode.json and attempts to calculate cache write costs:

const cacheWriteCost = model.cost.cache_write * tokensWrittenToCache;
// → NaN for GPT, Gemini, Kimi (because undefined * number = NaN)

This leads to:

• NaN cost calculations displayed to the user.
• Incorrect total cost summaries.
• Potential runtime errors if the consumer expects a number.

Meanwhile, DEFAULT_FALLBACK_COSTS (line 107) does include cache_write: 6.25, so unsupported/unknown models paradoxically get a cache_write value while explicitly supported models like GPT 5.2 do not.

Proposed fix

For models where the provider does not charge for cache writes, explicitly set cache_write: 0:

"GPT 5.2": {
  cost: { input: 1.75, output: 14.0, cache_read: 0.175, cache_write: 0 },
},

Alternatively, if cache write pricing is unknown, document it clearly and ensure downstream consumers handle undefined gracefully.

Acceptance criteria

• All models in SUPPORTED_MODELS have a consistent cost schema: all four fields (input, output, cache_read, cache_write) are present.
• Models where the provider does not charge for cache writes have cache_write: 0 (not omitted).
• DEFAULT_FALLBACK_COSTS schema matches SUPPORTED_MODELS cost schema.

Test cases

test("all supported models should have consistent cost fields", () => {
  const requiredFields = ["input", "output", "cache_read", "cache_write"];
  
  for (const [modelName, config] of Object.entries(SUPPORTED_MODELS)) {
    if (!config.cost) continue;
    for (const field of requiredFields) {
      expect(config.cost[field]).toBeDefined();
      expect(typeof config.cost[field]).toBe("number");
    }
  }
});

test("generated config should have numeric cache_write for all models", () => {
  const modelsList = [
    { id: "gpt-5", name: "GPT 5.2" },
    { id: "claude", name: "Claude Opus 4.6" },
    { id: "gemini", name: "Gemini 2.5 Pro" },
  ];
  
  const { models } = processModels(modelsList, {}, true);
  
  for (const [name, model] of Object.entries(models)) {
    expect(typeof model.cost.cache_write).toBe("number");
  }
});

🤖 Generated with Claude Code