From the first cross-provider field eval (claude-code f8e49d74 vs codex 019fa810, same workspace, 2026-07-28).
Symptom. A 36-minute Codex investigation session traces at $101.99 — more than a full day of Claude work ($93.35) in the same workspace. The header reads 6.6m in/59k out, cache 6.3m r.
Cause. Two subset relationships in Codex usage that ccx treats as disjoint categories:
input_tokens includes cached_input_tokens. Upstream codex-rs (protocol/src/protocol.rs): non_cached_input = input_tokens - cached_input(). ccx stores the raw values (internal/provider/codex/backend.go, token_count handling) and ComputeCost (internal/parser/pricing.go) bills InputTokens at the input rate plus CacheReadTokens at the cache rate — cached tokens are billed twice, once at full input rate.reasoning_output_tokens is a subset of output_tokens (OpenAI output_tokens_details.reasoning_tokens). ComputeCost adds ReasoningTokens * OutputPer1M on top of OutputTokens * OutputPer1M — reasoning billed twice.
Verification against the session: turn 2 shows 2.9m in, 31k out, cache 2.8m r → ccx math 2.9m*$10 + 31k*$80 + 2.8m*$5 = $45.5 ≈ the displayed $45.30. Honest math (subset semantics): ~100k*$10 + 31k*$80 + 2.8m*$5 ≈ $17.5. Session total honest ≈ $39 vs displayed $101.99 — ~2.6x overstated.
Also display semantics. Claude's in excludes cache reads (Anthropic usage semantics); Codex's in includes them. The same header line means different things per provider — 6.6m in on a session whose uncached input is ~300k.
Fix. Normalize at the Codex parse layer to the canonical exclusive semantics (Anthropic-style, which the rest of ccx already assumes):
InputTokens = input_tokens - cached_input_tokens (clamped), for both the session aggregate and per-message deltasOutputTokens unchanged (keeps reasoning inside); ReasoningTokens stays informational- drop the reasoning term from
ComputeCost (same rate as output, already counted)
- document the contract on
MessageUsage
Priority: P1 — cost is a headline number and it is wrong for every Codex session with cache hits, i.e. all of them.
From the first cross-provider field eval (claude-code f8e49d74 vs codex 019fa810, same workspace, 2026-07-28).
Symptom. A 36-minute Codex investigation session traces at $101.99 — more than a full day of Claude work ($93.35) in the same workspace. The header reads
6.6m in/59k out, cache 6.3m r.Cause. Two subset relationships in Codex usage that ccx treats as disjoint categories:
input_tokensincludescached_input_tokens. Upstream codex-rs (protocol/src/protocol.rs):non_cached_input = input_tokens - cached_input(). ccx stores the raw values (internal/provider/codex/backend.go, token_count handling) andComputeCost(internal/parser/pricing.go) billsInputTokensat the input rate plusCacheReadTokensat the cache rate — cached tokens are billed twice, once at full input rate.reasoning_output_tokensis a subset ofoutput_tokens(OpenAIoutput_tokens_details.reasoning_tokens).ComputeCostaddsReasoningTokens * OutputPer1Mon top ofOutputTokens * OutputPer1M— reasoning billed twice.Verification against the session: turn 2 shows
2.9m in, 31k out, cache 2.8m r→ ccx math2.9m*$10 + 31k*$80 + 2.8m*$5 = $45.5≈ the displayed $45.30. Honest math (subset semantics):~100k*$10 + 31k*$80 + 2.8m*$5 ≈ $17.5. Session total honest ≈ $39 vs displayed $101.99 — ~2.6x overstated.Also display semantics. Claude's
inexcludes cache reads (Anthropic usage semantics); Codex'sinincludes them. The same header line means different things per provider —6.6m inon a session whose uncached input is ~300k.Fix. Normalize at the Codex parse layer to the canonical exclusive semantics (Anthropic-style, which the rest of ccx already assumes):
InputTokens=input_tokens - cached_input_tokens(clamped), for both the session aggregate and per-message deltasOutputTokensunchanged (keeps reasoning inside);ReasoningTokensstays informationalComputeCost(same rate as output, already counted)MessageUsagePriority: P1 — cost is a headline number and it is wrong for every Codex session with cache hits, i.e. all of them.