[FEATURE] Add 5-minute and 1-hour cache-write token counts to the OTel api_request event

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

The claude_code.api_request OTel event gives one cache-write count: cache_creation_tokens. The API response gives more data. Its usage.cache_creation object divides cache writes into ephemeral_5m_input_tokens and ephemeral_1h_input_tokens (see 1-hour cache duration).

The two durations have different prices. For Claude Opus 5.5, a 5-minute write costs $5 per million tokens. A 1-hour write costs $8 per million tokens.

Without the split, an OTel consumer cannot calculate the cost of a request from its token counts. When a session uses both durations, one rate for all cache writes gives an incorrect result.

Sessions with both durations are frequent. The agent often starts subagents, and subagents usually write with a different TTL than the main conversation.

The consumer also cannot find the duration from other attributes. The main conversation uses the 1-hour TTL. Subagents request the 5-minute TTL, but #84289 shows that the effective TTL can be different. Thus query_source does not identify the duration.

Proposed Solution

Add two integer attributes to the claude_code.api_request event:

| Attribute | Value |
|---|---|
| cache_creation_5m_tokens | usage.cache_creation.ephemeral_5m_input_tokens from the API response |
| cache_creation_1h_tokens | usage.cache_creation.ephemeral_1h_input_tokens from the API response |

Rules:

• Keep cache_creation_tokens with no change. Existing dashboards and collectors continue to work.
• Send the two attributes only when the API response includes usage.cache_creation. Do not send zero for a count that the API did not report.
• When the event has the two attributes, their sum equals cache_creation_tokens. The API documentation gives the same rule for its fields.
• Add the two attributes to the api_request table on the monitoring page.

Priority

Medium - Would be very helpful

Feature Category

Other

Use Case Example

1. I collect Claude Code OTel events with a local tool. The tool calculates a reference cost from the published API rates.
2. I used claude-opus-5-5 for one session on Claude Code 2.1.290. The session had subagents. The events gave these totals:

| Category | Tokens |
|---|---|
| Input | 4,780 |
| Cache read | 77,498,782 |
| Cache write | 915,534 |
| Output | 238,758 |

1. /cost reported a total of $26.30.
2. My tool uses one rate for all 915,534 cache-write tokens. With the 1-hour rate, the result is $27.62. With the 5-minute rate, the result is $24.87. Neither result agrees with /cost.
3. The main conversation wrote 474,905 of the cache-write tokens. Other requests wrote 440,629. When I price the first group at the 1-hour rate and the second group at the 5-minute rate, the result is $26.30. This agrees with /cost.
4. The transcripts for this project show the same pattern. All main-conversation writes are 1-hour. All subagent writes are 5-minute.
5. With the proposed attributes, the tool can use the correct rate for each request. It does not need an assumption about the duration.

Additional Context

• #68771 asks for the standard gen_ai.usage.* attributes. The OTel GenAI conventions for Anthropic define only gen_ai.usage.cache_write.input_tokens, with no TTL split. Thus #68771 does not supply the split, and this request does not depend on it.
• Rates are the Claude Opus 5.5 rates from the Claude price list, in USD per million tokens: input $4, cache read $0.20, 5-minute cache write $5, 1-hour cache write $8, output $20.
• The two lines below are excerpts from my local Claude Code session transcripts. Both records are from one project and use claude-opus-5-5. I used jq to select two fields from message.usage of each assistant record:

| Line | Transcript file | Claude Code version | Source of request |
|---|---|---|---|
| 1 | ~/.claude/projects/<project>/<session>.jsonl | 2.1.290 | Main conversation |
| 2 | ~/.claude/projects/<project>/<session>/subagents/agent-<id>.jsonl | 2.1.284 | Subagent |

``json
{"cache_creation_input_tokens": 17710, "cache_creation": {"ephemeral_1h_input_tokens": 17710, "ephemeral_5m_input_tokens": 0}}
{"cache_creation_input_tokens": 31523, "cache_creation": {"ephemeral_5m_input_tokens": 31523, "ephemeral_1h_input_tokens": 0}}
``

The api_request event has only cache_creation_tokens. Thus the events for these two records do not show that their cache writes have different prices.