Claude 1h cache writes priced at the 5m rate

[machow]

agentsview version

agentsview v0.41.1 (commit a902515, built 2026-08-18T13:35:15Z)

Install method

Built from source

OS / platform

macOS 26.5 (arm64)

Which agent and version

Claude Code 2.1.231, subscription auth

Which model(s)

claude models with 1hr cache

What happened, and what did you expect

Claude Code transcripts declare the cache-write TTL split per request:

"cache_creation_input_tokens": 8989,
"cache_creation": { "ephemeral_1h_input_tokens": 8989,
                    "ephemeral_5m_input_tokens": 0 }

However, the 5 minute rate gets used when calculating the cost above, instead of the 1 hour rate.

Here is the current sequence behind calculating:

1. The Claude parser stores message.usage verbatim, so the TTL
   breakdown is already in every messages.token_usage blob.
2. At query time the usage scanner reads five flat keys only and skips
   the nested cache_creation object
   (parseUsageTokenCountersWithReasoning, internal/db/usage.go:1310).
3. The catalog keeps only cache_creation_input_token_cost (the 5m
   rate); LiteLLM's cache_creation_input_token_cost_above_1hr is
   dropped at parse (internal/pricing/catalog/litellm.go:129).
4. Cost = flat cache-write total x 5m rate, so 1h writes (billed 2x
   input by Anthropic) are priced at 1.25x.

Sample session file or snippet

{"type":"user","cwd":"/home/user/demo","sessionId":"aaaaaaaa-bbbb-cccc-dddd-eeeeeeee0001","version":"2.1.231","uuid":"11111111-1111-1111-1111-111111111101","timestamp":"2026-08-13T12:00:00.000Z","isSidechain":false,"message":{"role":"user","content":"Reply with exactly: hello"}}
{"type":"assistant","cwd":"/home/user/demo","sessionId":"aaaaaaaa-bbbb-cccc-dddd-eeeeeeee0001","version":"2.1.231","uuid":"11111111-1111-1111-1111-111111111102","timestamp":"2026-08-13T12:00:05.000Z","isSidechain":false,"requestId":"req_demo000000000000000000001","message":{"id":"msg_demo000000000000000000001","type":"message","role":"assistant","model":"claude-fable-5","content":[{"type":"text","text":"hello"}],"stop_reason":"end_turn","usage":{"input_tokens":2,"output_tokens":62,"cache_creation_input_tokens":8989,"cache_read_input_tokens":15892,"cache_creation":{"ephemeral_1h_input_tokens":8989,"ephemeral_5m_input_tokens":0},"service_tier":"standard"}}}
{"type":"assistant","cwd":"/home/user/demo","sessionId":"aaaaaaaa-bbbb-cccc-dddd-eeeeeeee0001","version":"2.1.231","uuid":"11111111-1111-1111-1111-111111111103","timestamp":"2026-08-13T12:01:00.000Z","isSidechain":false,"requestId":"req_demo000000000000000000002","message":{"id":"msg_demo000000000000000000002","type":"message","role":"assistant","model":"claude-fable-5","content":[{"type":"text","text":"goodbye"}],"stop_reason":"end_turn","usage":{"input_tokens":2,"output_tokens":6,"cache_creation_input_tokens":77,"cache_read_input_tokens":24881,"cache_creation":{"ephemeral_1h_input_tokens":77,"ephemeral_5m_input_tokens":0},"service_tier":"standard"}}}

Steps to reproduce

Create a session, and use Claude Code's total_cost_usd as
ground truth. E.g. claude -p "Reply with exactly: hello" --output-format json.

I confirmed by checking the output of command above and viewing the cost estimated in agentsview ui.

Here's a table summarizing various ways of calculating the cost (by fable):

Source	Cost	Formula
Claude Code total_cost_usd	$0.225533	cache writes at $20/MTok (1h)
Hand math, 1h rates	$0.225533	exact
agentsview session usage	$0.157539	cache writes at $12.50/MTok (5m)
Hand math, 5m rates	$0.157538	matches agentsview to 1 microdollar

First request, spelled out (fable-5: input $10, output $50, cache read
$1, cache write $12.50 5m / $20 1h per MTok):

2x10 + 62x50 + 8989x20   + 15892x1  = $0.198792  (actual, matches meter)
2x10 + 62x50 + 8989x12.5 + 15892x1  = $0.131375  (agentsview)

Checklist

• I searched existing issues
• I removed secrets and private data from any attached session files