1-hour cache writes priced at 5-minute rate

[marcin-kurdziel]

Bug

CodeBurn applies the 5-minute cache-write rate to all cache writes, including 1-hour ones. Anthropic charges:

• 5m cache write: 1.25× input (e.g. $6.25/MTok on Opus 4.7)
• 1h cache write: 2× input (e.g. $10.00/MTok on Opus 4.7)

Source: https://docs.anthropic.com/en/docs/about-claude/pricing

Root cause is that FALLBACK_PRICING (and the LiteLLM fields it mirrors) only carries one cache_creation_input_token_cost, which corresponds to the 5m rate.

Real-world impact on a Claude Code Opus 4.7 day: ~14% under-report. Plan-mode and long agent sessions use 1h cache heavily, so the gap scales with that share of the workload.

Data is already there

Claude Code session JSONL splits the two:

"usage": {
  "cache_creation_input_tokens": 60120,
  "cache_creation": {
    "ephemeral_5m_input_tokens": 0,
    "ephemeral_1h_input_tokens": 60120
  }
}

So no new data source needed — only parser + pricer change.

Suggested fix

Read usage.cache_creation.ephemeral_5m_input_tokens and usage.cache_creation.ephemeral_1h_input_tokens separately. Price the 5m portion at the existing rate; price the 1h portion at 1.6 × cache_creation_input_token_cost (since 2.0 / 1.25 = 1.6, constant across current Anthropic models). Fall back to current behavior when only the legacy cache_creation_input_tokens field is present.

A parallel issue against BerriAI/litellm to add an explicit cache_creation_input_1h_token_cost field would let downstream consumers drop the hardcoded 1.6×

[ozymandiashh]

Implemented this in PR #304: #304

What changed:

• Claude JSONL parsing now reads the duration split under usage.cache_creation:
  • ephemeral_5m_input_tokens
  • ephemeral_1h_input_tokens
• The existing aggregate field, cache_creation_input_tokens, is still used as the report/export cache-write token total.
• 5-minute cache writes continue to use the existing LiteLLM cache_creation_input_token_cost field.
• 1-hour cache writes are priced at 1.6x that 5-minute cache-write rate. That matches Anthropic pricing because 5m cache writes are 1.25x base input and 1h cache writes are 2.0x base input, so 2.0 / 1.25 = 1.6.
• Legacy Claude JSONL without usage.cache_creation keeps the old behavior: all cache_creation_input_tokens are priced at the 5-minute cache-write rate.

Example behavior:

"usage": {
  "input_tokens": 0,
  "output_tokens": 0,
  "cache_creation_input_tokens": 60120,
  "cache_creation": {
    "ephemeral_5m_input_tokens": 0,
    "ephemeral_1h_input_tokens": 60120
  }
}

For claude-opus-4-7, that now reports the same 60,120 cache-write tokens but costs 60,120 * $10/MTok = $0.6012 instead of 60,120 * $6.25/MTok = $0.37575.

I also bumped the durable daily cache to v6. The cache stores derived cost rollups, not raw per-call split data, so v5 cached historical days could not be safely migrated; they are backed up and recomputed from raw sessions instead.

Validation:

• npm run build
• npx vitest run tests/parser-claude-cwd.test.ts tests/models.test.ts tests/models-hoist.test.ts tests/daily-cache.test.ts (4 files / 103 tests)
• npx vitest run (47 files / 665 tests)
• git diff --check
• Argus local reviewer: PASS
• Claude Opus 4.7 effort max: PASS
• Gemini 3.1 Pro Preview: PASS
• PR checks: check, assess, semgrep all PASS

Source checked: Anthropic pricing docs: https://platform.claude.com/docs/en/about-claude/pricing