Bill OpenAI reasoning tokens as output

[mike1858]

Why

Piebald usage rows can include reasoning tokens alongside output tokens. Splitrail already displays those fields separately, but OpenAI-style model pricing needs reasoning tokens to be billed at the output-token rate. Without that, rows with OpenAI reasoning tokens underreport cost while keeping the token columns otherwise correct.

For example, the same 6/24/2026 row moves from $348.92 before this fix to $360.77 after reasoning tokens are included in billable output:

Date               Cost    Cached Tks       Inp Tks      Outp Tks    Reason Tks   Convs   Tools  Models
Old: 6/24/2026       $348.92       223.34m        43.68m       586.70k       392.69k      29   2.68k  claude-opus-4-8, gpt-5.5
New: 6/24/2026       $360.77       223.34m        43.68m       586.70k       392.69k      29   2.68k  claude-opus-4-8, gpt-5.5

What changed

• Added a shared OpenAI-token-semantics helper for the Piebald analyzer.
• Added a billable-output calculation that includes reasoning tokens for OpenAI-style models before passing token counts into Splitrail pricing.
• Preserved the displayed output_tokens and reasoning_tokens values so the stats table still reports those columns separately.
• Kept Anthropic-style output pricing unchanged.
• Added focused tests covering OpenAI reasoning-token billing and Anthropic output preservation.

Validation

• cargo test --quiet piebald - 16 passed
• cargo build --quiet - passed
• cargo test --quiet - 370 passed; 1 failed: models::tests::repeated_tiered_model_lookups_reuse_the_same_tier_storage
• cargo test --quiet models::tests::repeated_tiered_model_lookups_reuse_the_same_tier_storage - passed on focused rerun
• cargo clippy --quiet -- -D warnings - passed
• cargo doc --quiet - passed
• cargo fmt --all --quiet - passed with a clean worktree afterward

Visual proof is not applicable for this non-visual analyzer pricing change; I reviewed the code path and validated it with focused analyzer tests plus the repository Rust checks above.

Summary by CodeRabbit

• Bug Fixes
  • Updated token counting and billing so costs now better match provider-specific token semantics.
  • Adjusted how cached input reads and reasoning tokens are handled in pricing calculations, improving accuracy for affected models.
  • Preserved existing billing behavior for other providers while aligning output charges with the newer token rules.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 6b54c3fc-03eb-4555-b013-7558f85a6913

📥 Commits

Reviewing files that changed from the base of the PR and between 1c18ee0 and 8c20181.

📒 Files selected for processing (1)

• src/analyzers/piebald.rs

🚧 Files skipped from review as they are similar to previous changes (1)

• src/analyzers/piebald.rs

---

📝 Walkthrough

Walkthrough

Refactors Piebald token billing semantics to count OpenAI reasoning tokens in billed output, reuse the semantic check in input normalization, wire the new output value into cost calculation, and update pricing tests.

Changes

Billable Output Token Semantics

Layer / File(s)	Summary
Token semantics helpers and normalization refactor src/analyzers/piebald.rs	Adds model_uses_openai_token_semantics and billable_output_tokens, and refactors normalize_input_tokens to use the new semantics helper for cached-read subtraction.
Cost computation wiring src/analyzers/piebald.rs	Computes billable_output from billable_output_tokens(...) and passes it to calculate_total_cost_for_service_tier_at instead of raw output_tokens.
Test fixture and assertions update src/analyzers/piebald.rs	Changes the pricing fixture’s reasoning_tokens, updates expected cost, and adds assertions for OpenAI reasoning-token billing and Anthropic output preservation.

Estimated code review effort: 2 (Simple) | ~12 minutes

Possibly related PRs

• Piebald-AI/splitrail#167: Shares the same src/analyzers/piebald.rs token-accounting path, including normalize_input_tokens(...) and cached-read handling.
• Piebald-AI/splitrail#172: Also changes the service-tier cost calculation in the same analyzer flow.
• Piebald-AI/splitrail#193: Uses the same calculate_total_cost_for_service_tier_at call that this PR now feeds with billable output tokens.

Poem

I’m a rabbit with a token chart,
Counting reasoning bits by heart.
Hop-hop — output now includes the glow,
Cached reads trimmed where needed, low and slow.
The billing burrow feels just right today.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: OpenAI reasoning tokens are now billed as output tokens.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch bill-openai-reasoning-tokens-as-output

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}