Add support for priority tiers & better long-context surcharge

[mike1858]

Summary

This PR fixes and extends OpenAI GPT pricing support in Splitrail, with Piebald now using service-tier-aware pricing when its database records a service tier.

Changes

• Corrected long-context pricing for gpt-5.4, gpt-5.4-pro, and gpt-5.5.
• Changed those long-context models to use full-session bracket pricing instead of progressive tier pricing.
• Added ServiceTier support to the model pricing layer:
  • Standard
  • Priority
  • Flex
  • Batch
• Added per-service-tier pricing overrides to ModelInfo.
• Added service-tier-aware cost helpers:
  • calculate_input_cost_for_service_tier
  • calculate_output_cost_for_service_tier
  • calculate_cache_cost_for_service_tier
  • calculate_total_cost_for_service_tier
• Added Priority pricing for:
  • gpt-5.5
  • gpt-5.4
  • gpt-5.4-mini
• Added Flex and Batch pricing for:
  • gpt-5.5
  • gpt-5.5-pro
  • gpt-5.4
  • gpt-5.4-mini
  • gpt-5.4-nano
  • gpt-5.4-pro
• Updated the Piebald analyzer to read OpenAI service tiers from its generation config override tables:
  • override_gen_cfg_data_openai_responses.service_tier
  • override_gen_cfg_data_openai_completions.service_tier
• Piebald now maps:
  • priority → ServiceTier::Priority
  • flex → ServiceTier::Flex
  • batch → ServiceTier::Batch
  • unknown/null values → ServiceTier::Standard
• Piebald now calculates estimated costs with calculate_total_cost_for_service_tier(...).

Why

OpenAI’s long-context GPT pricing says prompts above the 272k input-token threshold are priced at the long-context rate for the full session. Splitrail previously represented some of these prices as progressive tiers, which undercounted long-context sessions.

Additionally, OpenAI exposes different pricing for service tiers such as Priority, Flex, and Batch. Splitrail now has a pricing model that can represent these tiers, and Piebald can use that information when it is present in its database.

Notes

• Existing cost helpers still default to Standard pricing, so existing analyzers continue to work unchanged.
• Piebald is the only analyzer wired to detect service tiers in this PR.
• Unknown service-tier values fall back to Standard pricing. This handles unsupported values like scale without inventing pricing behavior.
• Piebald still normalizes OpenAI cached-input token reporting the same way as before; this PR only changes which pricing table is selected.

Verification

Ran:

cargo test --quiet analyzers::piebald::tests
cargo test --quiet models::tests::gpt
cargo build --quiet
cargo clippy --quiet -- -D warnings
cargo fmt --all --quiet

Summary by CodeRabbit

• New Features
  • Added support for multiple service tiers (Standard, Priority, Flex, Batch) with corresponding pricing configurations.
  • Service tier information is now automatically extracted and integrated into cost calculations.
  • Updated supported models with service-tier-specific pricing options for improved cost accuracy.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: e22d8c3a-13c8-4e04-82e8-ebdcda615636

📥 Commits

Reviewing files that changed from the base of the PR and between 838bdf9 and 49c02ce.

📒 Files selected for processing (2)

• src/analyzers/piebald.rs
• src/models.rs

---

📝 Walkthrough

Walkthrough

This PR adds service-tier support to model pricing and cost calculation. It introduces new ServiceTier and ServiceTierPricing types, extends model configuration to register per-tier pricing overrides, refactors cost APIs to select pricing by tier, and integrates service tier extraction in the piebald analyzer to compute tiered costs from database metadata.

Changes

Service-Tier Cost Model

Layer / File(s)	Summary
Service Tier Type Definitions src/models.rs	ServiceTier enum (Standard/Priority/Flex/Batch) and ServiceTierPricing struct introduced; ModelInfo extended with service_tiers: HashMap<ServiceTier, ServiceTierPricing>.
Model Configuration & Validation for Service Tiers src/models.rs	validate_model_info validates service-tier entries; populate_defaults initializes service_tiers and registers GPT tier overrides; bracket_pricing changed to true for tiered pricing/caching; free tier model updated.
Service-Tier-Aware Cost Calculation APIs src/models.rs	pricing_for_service_tier helper selects override or base pricing; new functions calculate_*_for_service_tier compute costs from tier-selected pricing; existing calculate_* reworked as Standard-tier wrappers.
Model Test Coverage for Service Tiers src/models.rs	Test imports and fixtures updated; comprehensive coverage for tiered pricing, fallback behavior, and tier-missing cases.
Piebald Analyzer Service Tier Extraction & Costing src/analyzers/piebald.rs, src/models.rs	PiebaldMessage gains optional service_tier field; SQL query joins override tables and selects tier via COALESCE; parse_service_tier helper normalizes strings to ServiceTier variants; message conversion calls calculate_total_cost_for_service_tier.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related PRs

• Piebald-AI/splitrail#167: Both PRs modify convert_messages in the piebald analyzer to affect cost computation (service-tier-driven calculation vs. cached-read token normalization).
• Piebald-AI/splitrail#125: Both PRs extend cost/pricing APIs in src/models.rs with tiering support (service_tiers vs. bracket-based pricing).
• Piebald-AI/splitrail#149: Both PRs modify GPT-5.4/5.5 default pricing and registry entries in src/models.rs.

Suggested reviewers

• bl-ue

Poem

🐰 Tiers upon tiers, the pricing flows,
Standard, Priority—the model knows,
Flex and Batch join the cost brigade,
Service-tier logic, now properly made!
From database rows to calculations bright, ✨

✨ 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 pb-analyzer

---

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 and usage tips.}