Releases·Tool Release
We captured raw Claude Code API traffic: where tokens and budget actually go (breakdown of 5 request-level components)
Developers captured raw Claude Code API traffic via local proxy to analyze where tokens and cost go in multi-turn sessions. They identified five cost drivers, including MCP tool schemas consuming 15k-25k input tokens per turn and prompt cache misses that cut the 90% discount. They released an open-source tool, cost-xray, that attributes tokens and costs to individual sources for Claude Code and Codex.
2 engagement1 sourceFri, Sep 11, 2026, 03:57 PM
AI Overview
- Reddit user's proxy capture found 4-5 MCP servers add 15k-25k input tokens per turn in tool schemas; released open-source cost-xray tool
- Prompt cache misses on tool-definition changes drop the 90% discount, jumping a $0.03 turn to $0.35
- Extended thinking blocks can exceed code output 4x on complex refactors, driving over half of per-turn dollar cost
- Tool outputs like 500-line grep logs stay in context every subsequent turn until compaction runs
- Compaction's summarization turn itself requires sending the entire bloated context to the model at once
Related
Social·Community
I traced where my tokens actually go
A developer running Codex and Claude Code traced where tokens go per turn, finding the full conversation resends every turn, so early junk compounds cost. The breakdown quantifies per-item token costs like system prompts, CLAUDE.md, screenshots, and MCP connectors.
A developer running Codex and Claude Code traced where tokens go per turn, finding the full conversation resends every turn, so early junk compounds cost. The breakdown quantifies per-item token costs like system prompts, CLAUDE.md, screenshots, and MCP connectors.
12 engagement1 sourceredditSun, Sep 20, 2026, 02:34 AM
Social·Community
PSA: Claude Code Prompt Suggestions trigger full context Cache Reads, costing up to 10% of weekly limits
A user instrumenting Claude Code found that Prompt Suggestions — the greyed-out autocomplete text in the input line — trigger a Cache Read of the entire context and can cost up to 10% of a weekly limit. The poster advises turning the feature off. Suggestions cost a median 91% of the corresponding normal prompt's cost in their logs.
A user instrumenting Claude Code found that Prompt Suggestions — the greyed-out autocomplete text in the input line — trigger a Cache Read of the entire context and can cost up to 10% of a weekly limit. The poster advises turning the feature off. Suggestions cost a median 91% of the corresponding normal prompt's cost in their logs.
601 engagement1 sourceredditMon, Sep 21, 2026, 09:50 AM
Releases·Tool Release
Developer builds open-source profiler for Claude Code, finds $15k in wasted tokens from re-reads and failed commands
A developer analyzed 60 sessions of Claude Code transcripts and built a free open-source profiler to identify token waste. Key findings: a 5,649-line progress doc was read 118 times, and 455 failed tool calls cost ~$201 in retry billing. Caching saved an estimated $15k in API costs.
A developer analyzed 60 sessions of Claude Code transcripts and built a free open-source profiler to identify token waste. Key findings: a 5,649-line progress doc was read 118 times, and 455 failed tool calls cost ~$201 in retry billing. Caching saved an estimated $15k in API costs.
21029 engagement3 sourcesredditThu, Jul 23, 2026, 10:04 AM
Releases·Tool Release
Claude Code transcript audit finds 4 token leaks; fixes cut cost per call 40%
A developer analyzed Claude Code's per-call token logs and found most cost came from context bloat, mis-inherited model assignments, and repeated file reads — not output verbosity. After four fixes, cost per call fell 40% and subagent spend dropped 77%. The audit and fixes are packaged as a local-only plugin, /scrooge.
A developer analyzed Claude Code's per-call token logs and found most cost came from context bloat, mis-inherited model assignments, and repeated file reads — not output verbosity. After four fixes, cost per call fell 40% and subagent spend dropped 77%. The audit and fixes are packaged as a local-only plugin, /scrooge.
9 engagement1 sourceredditSun, Oct 4, 2026, 05:10 AM
Releases·Tool Release
Developers report Claude Code burns ~70k tokens per session on context rebuilds
Developers analyzing their own Claude Code transcripts found sessions incur roughly 70,000 tokens of overhead before any user input, driven by system prompts, tool schemas, and accumulated rule files. One test logged 33k tokens before the prompt, with 24k from tool schemas alone, while open-source alternatives like OpenCode start closer to 7k. The findings highlight a hidden cost for heavy users and explain the '70% wall' where projects slow down as codebases grow.
Developers analyzing their own Claude Code transcripts found sessions incur roughly 70,000 tokens of overhead before any user input, driven by system prompts, tool schemas, and accumulated rule files. One test logged 33k tokens before the prompt, with 24k from tool schemas alone, while open-source alternatives like OpenCode start closer to 7k. The findings highlight a hidden cost for heavy users and explain the '70% wall' where projects slow down as codebases grow.
02135 engagement3 sourcesredditFri, Jul 31, 2026, 01:58 AM