Production evidence: Anthropic cache bust bug causes $14.82 waste across 3 sessions (22.5% bust rate)

[Arrowyi]

Summary

Production data from OpenCode 1.18.9 shows the Anthropic prompt cache bust bug (Issue #24841) is still active and causing significant financial waste.

Key findings:

• 5 sessions using Opus 5 max cost $38.35
• $24.45 (63.8%) of that was cache write charges
• 9 cache busts in a single 43-message session (22.5% bust rate)
• 5 consecutive cache busts in a row (messages Codex mini #34-38), each costing $1.63
• All 9 cache busts occurred immediately after tool-calls, confirming the root cause identified in Prompt loop DB reload breaks Anthropic cache after tool calls (63% of spend) #24841

Estimated waste: $14.82 (39% of total cost) could have been avoided if caching worked correctly.

---

Evidence

Session 1: ses_03563ecbdffe92x5KFrsYJRWnN ($18.66)

Title: Creating AGENTS.md for repository
Date: 2026-08-04
Messages: 43 (38 tool-calls, 2 stop)
Total tokens: 80 input / 48,028 output / 6,148,780 cache read / 2,301,161 cache write

Cost breakdown:

Component	Tokens	Rate	Cost	%
Input	80	$5.00/M	$0.00	0%
Output	48,028	$25.00/M	$1.20	6.4%
Cache Read	6,148,780	$0.50/M	$3.07	16.5%
Cache Write	2,301,161	$6.25/M	$14.38	77.1%

Cache bust pattern (all after tool-calls):

msg#  cache_read  cache_write  cost    prev_finish  flag
1     0           84,195       $0.55   (start)      CACHE RESET
21    0           233,523      $1.47   tool-calls   CACHE RESET
29    0           255,794      $1.61   tool-calls   CACHE RESET
30    0           256,289      $1.61   tool-calls   CACHE RESET
34    0           259,313      $1.63   tool-calls   CACHE RESET
35    0           259,498      $1.63   tool-calls   CACHE RESET
36    0           259,734      $1.63   tool-calls   CACHE RESET
37    0           260,018      $1.63   tool-calls   CACHE RESET
38    0           260,207      $1.63   tool-calls   CACHE RESET

Consecutive cache busts (#34-38): 5 in a row, wasting $8.15

Expected cost if cache worked: $5.84
Actual cost: $18.66
Waste: $12.82 (69%)

---

Session 2: ses_03861eca0ffeuobD3dLHqYq99p ($11.50)

Title: LLM速度对比与Agent配置优化建议
Date: 2026-08-03
Messages: 73 (67 tool-calls, 3 stop)
Cache busts: 2
Cache write: 523,712 tokens × $6.25/M = $3.27 (28.5% of cost)

---

Session 3: ses_0381fd091ffe4KS4TxsnTcsOFW ($7.51)

Title: Claude Code Max订阅与API token费用对比
Date: 2026-08-03
Messages: 45 (35 tool-calls, 5 stop)
Cache busts: 2
Cache write: 498,810 tokens × $6.25/M = $3.12 (41.5% of cost)

---

Comparison: DeepSeek V4 Pro (no cache write fee)

Session: ses_02b3612c6ffeOFnJw6eQrlTKW0 (2026-08-06)
Cost: $0.03
Cache write: 0 (DeepSeek does not charge for cache writes)
Cache read: grows steadily from 2K to 89K (no busts)

This demonstrates that the issue is specific to Anthropic's cache write pricing combined with the cache bust bug.

---

Root Cause

Matches Issue #24841 exactly:

1. OpenCode's prompt loop reloads all messages from DB via filterCompactedEffect() at the start of every iteration
2. Tool parts transition from pending → completed with output text between API calls
3. toModelMessages() serializes the conversation with different bytes
4. Anthropic's prompt cache sees different content → cache invalidated from that position forward
5. Entire conversation context (~250K tokens) re-written at cache-write pricing ($6.25/M)

Verified: All 9 cache busts in the $18.66 session occurred immediately after tool-calls finish type.

---

New Evidence Not in Previous Reports

1. Consecutive cache busts: 5 in a row (messages Codex mini #34-38), each costing $1.63. This suggests the cache is being invalidated on every single turn during heavy tool-use loops.

2. Cache bust rate: 22.5% (9/40 messages) in a real production session with only 43 messages.

3. Cost per bust: $1.46-$1.63 (250K tokens), lower than the $3.50 (560K) reported in Prompt loop DB reload breaks Anthropic cache after tool calls (63% of spend) #24841, but still significant at scale.

4. Production cost data: Real money wasted ($14.82 across 3 sessions) with exact per-message cost breakdown from OpenCode's SQLite database.

---

Database Queries

Query 1: Session-level cache stats

SELECT 
  id, 
  cost, 
  tokens_input, 
  tokens_output, 
  tokens_cache_read, 
  tokens_cache_write 
FROM session 
WHERE id = 'ses_03563ecbdffe92x5KFrsYJRWnN';

Query 2: Message-level cache pattern

SELECT 
  json_extract(data, '$.finish') as finish,
  json_extract(data, '$.tokens.cache.read') as cache_r,
  json_extract(data, '$.tokens.cache.write') as cache_w,
  json_extract(data, '$.cost') as cost
FROM message 
WHERE session_id = 'ses_03563ecbdffe92x5KFrsYJRWnN' 
  AND json_extract(data, '$.role') = 'assistant'
ORDER BY json_extract(data, '$.time.created');

Query 3: Verify cache busts after tool-calls

WITH msg_list AS (
  SELECT 
    json_extract(data, '$.finish') as finish,
    json_extract(data, '$.tokens.cache.read') as cache_r,
    json_extract(data, '$.tokens.cache.write') as cache_w,
    ROW_NUMBER() OVER (ORDER BY json_extract(data, '$.time.created')) as rn
  FROM message 
  WHERE session_id = 'ses_03563ecbdffe92x5KFrsYJRWnN' 
    AND json_extract(data, '$.role') = 'assistant'
)
SELECT 
  m1.rn,
  m1.cache_r,
  m1.cache_w,
  m2.finish as prev_finish
FROM msg_list m1
LEFT JOIN msg_list m2 ON m1.rn = m2.rn + 1
WHERE m1.cache_r = 0 AND m1.cache_w > 0;

---

Impact

• Financial: $14.82 wasted across 3 sessions (39% of total Opus 5 spend)
• User trust: Users expect prompt caching to work as advertised; silent cache busts at 12.5x read cost are unacceptable
• Model choice: Forces users to avoid Anthropic models on Zen for tool-heavy workloads, limiting model selection

---

Request

1. Reopen Issue Prompt loop DB reload breaks Anthropic cache after tool calls (63% of spend) #24841 or create a new issue with this evidence
2. Prioritize PR fix(session): cache messages across prompt loop to preserve prompt cache byte-identity #36852 (or equivalent fix) to cache messages across prompt loop iterations
3. Add cache bust detection/logging so users can see when this is happening
4. Consider a cost warning when cache bust rate exceeds a threshold

---

Environment

• OpenCode: 1.18.9
• OS: macOS (darwin)
• Provider: OpenCode Zen (opencode/claude-opus-5)
• Variant: max
• Agent: Sisyphus - ultraworker (via oh-my-openagent)

---

Additional Data

Full session data available upon request (can export from ~/.local/share/opencode/opencode.db). Per-message cost breakdown available. Can provide additional sessions if needed.

Related: #24841 #25366 #31525 #20110 #36852

[Arrowyi]

Adding new evidence that was not in the original report and addressing the byte-identity point raised in #31525.

Consecutive cache busts (not previously reported)

The most expensive session (ses_03563ecbdffe92x5KFrsYJRWnN, $18.66, 43 messages) shows 5 consecutive cache busts across messages #34-38. Each turn re-wrote ~260K tokens at cache-write pricing ($6.25/M), costing $1.63 per turn:

msg#  cache_read  cache_write  cost    prev_finish
34    0           259,313      $1.63   tool-calls
35    0           259,498      $1.63   tool-calls
36    0           259,734      $1.63   tool-calls
37    0           260,018      $1.63   tool-calls
38    0           260,207      $1.63   tool-calls

Five consecutive turns, $8.15, zero cache reads. This is not a one-off invalidation; the cache is being re-written on every turn during a tool-use loop.

Full bust pattern (all 9 after tool-calls)

msg#	cache_write	cost	prev_finish
1	84,195	$0.55	(start)
21	233,523	$1.47	tool-calls
29	255,794	$1.61	tool-calls
30	256,289	$1.61	tool-calls
34-38	259-260K	$1.63 x5	tool-calls (x5)

9 of 9 busts immediately follow a tool-calls finish. The correlation is deterministic, not incidental.

On the byte-identity framing (#31525)

I understand the objection that Anthropic caching is keyed on content, not object identity. The data here does not depend on that framing. The observable fact is that cache_read = 0 and cache_write > 0 on turns immediately after tool-calls, and cache_write scales with full context size (~260K tokens), not with the incremental delta. Whatever the exact mechanism, the effect is that the entire conversation is re-serialized into the cache on each tool-turn, and Anthropic charges cache-write rates for it.

Reproducible from the database

No API introspection required. Anyone can verify from ~/.local/share/opencode/opencode.db:

WITH msg_list AS (
  SELECT
    json_extract(data, '$.finish') as finish,
    json_extract(data, '$.tokens.cache.read') as cache_r,
    json_extract(data, '$.tokens.cache.write') as cache_w,
    json_extract(data, '$.cost') as cost,
    ROW_NUMBER() OVER (ORDER BY json_extract(data, '$.time.created')) as rn
  FROM message
  WHERE session_id = 'ses_03563ecbdffe92x5KFrsYJRWnN'
    AND json_extract(data, '$.role') = 'assistant'
)
SELECT m1.rn, m1.cache_r, m1.cache_w, m1.cost, m2.finish as prev_finish
FROM msg_list m1
LEFT JOIN msg_list m2 ON m1.rn = m2.rn + 1
WHERE m1.cache_r = 0 AND m1.cache_w > 0;

Scope and limitations

This is production evidence from OpenCode 1.18.9 with opencode/claude-opus-5 (max), running the Sisyphus ultraworker agent. Total waste across 3 sessions is $14.82 (39% of spend). I have since switched these workloads to DeepSeek, whose sessions show steady cache-read growth and zero cache-write charges, which isolates the issue to Anthropic's cache-write pricing combined with this behavior.

I am not asking for a specific implementation, only that this be looked at with the deterministic 9/9 tool-call correlation and the 5-consecutive-bust sequence on record. If a fix lands in #14743 or elsewhere, happy to re-run the same workload and provide before/after numbers.

[cn4750]

I also see this issue.

[rijnhard]

This was actually included in a published article. https://thenewstack.io/agent-harness-token-costs/

the actual source: https://www.eishanlawrence.com/blog/harness-efficiency-bench

Basically using OpenCode at the moment can have a severe financial impact.