fix(pricing): sync flex/priority tier keys to dated OpenAI snapshot variants

[mateo-berri]

TLDR

Problem this solves:

• Dated OpenAI snapshots lack the flex/priority pricing their base alias has
• Pinning a snapshot silently bills service-tier traffic at standard rates

How it solves it:

• Copies base-alias tier keys onto 9 dated snapshot entries
• Adds a schema test that fails on any future tier-key drift

User Flow

Before: pinning a dated snapshot silently bills service-tier traffic at standard rates, undercharging priority and overcharging flex

1. A platform engineer pins gpt-4.1-2025-04-14 for reproducible evals and sends POST https://litellm-domain/v1/chat/completions with "service_tier": "priority" under their virtual key
2. The response body confirms OpenAI served the call at "service_tier": "priority"
3. The x-litellm-response-cost header reads the standard rate, roughly half of what OpenAI invoices for priority processing
4. A teammate running overnight evals on o4-mini-2025-04-16 with "service_tier": "flex" is billed double what OpenAI charges, since the 50% flex discount never applies
5. Key budgets and the spend page at https://litellm-domain/ui/?page=usage drift further from the provider invoice with every tiered call

After: the same pinned snapshots bill the tier's real price

1. A platform engineer pins gpt-4.1-2025-04-14 for reproducible evals and sends POST https://litellm-domain/v1/chat/completions with "service_tier": "priority" under their virtual key
2. The response body confirms OpenAI served the call at "service_tier": "priority"
3. The x-litellm-response-cost header reads the snapshot's priority rate, matching what the base alias gpt-4.1 charges and what OpenAI invoices
4. The teammate's flex eval run is billed at half the standard rate, matching OpenAI's flex discount
5. Key budgets and the spend page track the provider invoice for tiered traffic

Relevant issues

Supersedes #22086

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Both legs hit the real OpenAI API through a local proxy booted with LITELLM_LOCAL_MODEL_COST_MAP=True and this config:

model_list:
  - model_name: gpt-4.1-2025-04-14
    litellm_params:
      model: openai/gpt-4.1-2025-04-14
      api_key: os.environ/OPENAI_API_KEY
  - model_name: o4-mini-2025-04-16
    litellm_params:
      model: openai/o4-mini-2025-04-16
      api_key: os.environ/OPENAI_API_KEY

Priority tier on a pinned gpt-4.1 snapshot (same request against each proxy):

curl -sD - http://localhost:<port>/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"gpt-4.1-2025-04-14","service_tier":"priority","messages":[{"role":"user","content":"Reply with exactly: pong"}]}'

Before (bcce83a, base branch): OpenAI confirms "service_tier": "priority" with usage 12 prompt + 1 completion tokens, yet the proxy bills standard rates

x-litellm-response-cost: 3.2e-05   (12 x 2e-06 + 1 x 8e-06, the standard rate)

After (629c228, this PR): identical request, identical usage (12 + 1), now billed at the snapshot's synced priority keys

x-litellm-response-cost: 5.6e-05   (12 x 3.5e-06 + 1 x 1.4e-05, the priority rate)

Flex tier on a pinned o4-mini snapshot, same curl with "model":"o4-mini-2025-04-16","service_tier":"flex". Both legs returned "service_tier": "flex" with usage 11 prompt + 83 completion (64 reasoning) tokens

Before (bcce83a17e): x-litellm-response-cost: 0.0003773    (standard: 11 x 1.1e-06 + 83 x 4.4e-06)
After  (629c228b40): x-litellm-response-cost: 0.00018865   (flex, exactly 50%: 11 x 5.5e-07 + 83 x 2.2e-06)

Type

🐛 Bug Fix

Changes

OpenAI publishes flex and priority prices per model family, and LiteLLM carries them on the base aliases (gpt-4.1, gpt-4o, o3, o4-mini, gpt-5-nano, and friends), but most dated snapshot entries never received the keys. Anyone pinning a snapshot, which is the recommended practice for reproducible evals and batch jobs, got service-tier traffic billed at standard rates: undercharged for priority, overcharged for flex

This PR syncs the missing *_flex and *_priority keys from each base alias to its dated variants in model_prices_and_context_window.json and the backup copy: gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, gpt-4.1-nano-2025-04-14, gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-5-nano-2025-08-07, o3-2025-04-16, and o4-mini-2025-04-16. A key is copied only when the variant's corresponding standard-rate key exists and exactly matches the base alias, so snapshots with genuinely different pricing are left alone (gpt-4o-2024-05-13 is intentionally skipped for this reason)

tests/test_litellm/test_model_prices_schema.py gains test_dated_variants_carry_base_alias_service_tier_pricing, which walks every <name>-YYYY-MM-DD entry and fails whenever a base alias gains a tier key its price-matched snapshot lacks, so the drift cannot reappear as new tier prices land

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Greptile Summary

The PR synchronizes flex and priority pricing metadata from OpenAI base aliases to price-matched dated snapshots and adds a regression test for future drift.

• Updates the canonical and backup pricing maps consistently.
• Adds schema coverage comparing dated snapshots with their corresponding base aliases.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds service-tier pricing fields to the canonical entries for nine dated OpenAI snapshots.
litellm/model_prices_and_context_window_backup.json	Mirrors the canonical service-tier pricing additions in the backup pricing map.
tests/test_litellm/test_model_prices_schema.py	Adds a regression test that checks price-matched dated variants for base-alias flex and priority fields.

_{Reviews (2): Last reviewed commit: "fix(pricing): sync flex/priority tier ke..." | Re-trigger Greptile}

[veria-ai]

Medium: Priority pricing bypasses budget reservation

These _priority rates are used during final cost calculation, while budget_reservation.py estimates requests exclusively from input_cost_per_token and output_cost_per_token. A user can send priority requests against these models that pass admission using the lower standard estimate but consume up to roughly twice the reserved key or team budget. Update the reservation estimator to select the service-tier-specific input, output, cache, and reasoning rates from request_body["service_tier"], with the same fallback behavior as final cost calculation.

[mateo-berri]

Reservation already ignored service_tier for the base aliases' existing priority keys; this sync doesn't widen that pre-existing gap: tier-aware reservation deserves its own PR

[veria-ai]

PR overview

This pull request updates OpenAI pricing metadata so flex and priority service-tier keys align with dated model snapshot variants.

One security issue remains open: priority-tier requests are admitted using lower standard-rate budget estimates but charged at higher priority rates, allowing users to exceed reserved key or team budgets by roughly twofold. The reservation logic still needs to account for service-tier-specific pricing, and no issues have yet been addressed.

Open issues (1)

• Medium: Priority pricing bypasses budget reservation — model_prices_and_context_window.json:22256

Fixed/addressed: 0 · PR risk: 5/10

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_dated_variant_tier_pricing_sync (629c228) with litellm_internal_staging (732bba0)}

[mateo-berri]

@greptileai