fix(server): stop overpricing cached Claude tokens

[SunkenInTime]

The Usage page can overprice cached Claude tokens when LiteLLM publishes canonical and provider-qualified entries with the same bare model name. The parser stripped provider prefixes before inserting rates, so a later entry without the Anthropic cache discount could overwrite the canonical rate and price cache reads as ordinary input.

This builds on the exact-key approach in #8529 by @none23. @mackinleysmith identified one important edge case during that review: multiple qualified entries are safe to alias when their effective rates are identical. Treating those entries as ambiguous leaves models such as Claude Opus 4.6 and Claude Haiku 4.5 unpriced.

What changed

• Preserve each LiteLLM rate under its full normalized key.
• Keep canonical bare entries authoritative.
• Add a bare alias when every qualified entry has the same effective rate.
• Leave genuinely conflicting bare names unpriced instead of guessing.

Verification

• vp test run apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageAggregation.test.ts apps/server/src/usage/usageScanCache.test.ts
• vp lint apps/server/src/usage/usagePricing.ts apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageScanCache.ts apps/server/src/usage/usageScanCache.test.ts
• vp run --filter t3 typecheck
• Replayed the parser against the current LiteLLM table and confirmed Claude Opus 5, Claude Fable 5, Claude Opus 4.6, Claude Haiku 4.5, and GPT-5.2 Chat all retain prices.

Closes#5797.

Built with GPT-5.6 Sol in the Codex harness via T3 Code.

Note

Fix cached Claude token overpricing by preserving provider-qualified keys in lookupRate

• Changes normalizeRateKey to trim and lowercase keys without stripping provider prefixes, so parseRateTable stores full qualified keys like anthropic/claude-3.5-sonnet
• lookupRate now matches the exact normalized key; it no longer falls back to the bare model name automatically, which previously caused a cheaper cache entry to match a more expensive non-cache rate
• Adds bare-name aliases (e.g. claude-3.5-sonnet) only when no canonical bare entry exists and all provider-qualified entries for that name share the same rate; conflicting rates leave the bare name unpriced
• Adds sameRate and bareModelName helpers; refactors normalizeModelName to delegate to normalizeRateKey then bareModelName
• Risk: any rate table relying on bare-name fallback for a model with provider-specific rates will now return unpriced instead of guessing; check rate table configs in apps/server/src/usage/usagePricing.ts and any callers of lookupRate that expect bare-name matching

^{Macroscope summarized 63bf1f8.}

---

Note

Medium Risk
Changes how usage costs are computed for models with multiple LiteLLM keys; incorrect aliasing could misprice or leave models unpriced, but behavior is covered by new tests and avoids silent overwrites.

Overview
Fixes cached Claude token overpricing when LiteLLM publishes both a canonical bare model entry (with cache-read discounts) and provider-qualified entries that share the same trailing name. Rates are no longer collapsed to bare names at insert time, so a later DeepInfra-style row cannot overwrite the canonical cache-read rate.

parseRateTable now stores each entry under its full normalized key (trim + lowercase). It may add a bare-name alias only when no canonical bare key exists and every qualified variant agrees on input, output, and cache rates; conflicting bare names stay unpriced (lookupRate returns null).

lookupRate matches the normalized transcript model string against the table directly (still blocking ambiguous family names like opus/sonnet). normalizeModelName keeps the same bare-name contract for callers that strip provider prefixes.

Adds usagePricing.test.ts for normalization, Fable vs DeepInfra ordering, unanimous aliasing, and ambiguous bare names. Scan-cache comments were updated to reference lookupRate instead of normalizeModelName.

^{Reviewed by Cursor Bugbot for commit 63bf1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1fbff086-af1d-4c82-a012-eecf26e7cb3a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Warning

Your free Security trial is over. An organization admin can activate Security or dismiss this notice.

---

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This focused production fix changes token-cost and cache-savings calculations in existing usage summaries. Because those calculations have direct billing and metering implications, the change warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[SunkenInTime]

The Usage page can overprice cached Claude tokens when LiteLLM publishes canonical and provider-qualified entries with the same bare model name. The parser stripped provider prefixes before inserting rates, so a later entry without the Anthropic cache discount could overwrite the canonical rate and price cache reads as ordinary input.

This builds on the exact-key approach in #8529 by @none23. @mackinleysmith identified one important edge case during that review: multiple qualified entries are safe to alias when their effective rates are identical. Treating those entries as ambiguous leaves models such as Claude Opus 4.6 and Claude Haiku 4.5 unpriced.

What changed

• Preserve each LiteLLM rate under its full normalized key.
• Keep canonical bare entries authoritative.
• Add a bare alias when every qualified entry has the same effective rate.
• Leave genuinely conflicting bare names unpriced instead of guessing.

Verification

• vp test run apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageAggregation.test.ts apps/server/src/usage/usageScanCache.test.ts
• vp lint apps/server/src/usage/usagePricing.ts apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageScanCache.ts apps/server/src/usage/usageScanCache.test.ts
• vp run --filter t3 typecheck
• Replayed the parser against the current LiteLLM table and confirmed Claude Opus 5, Claude Fable 5, Claude Opus 4.6, Claude Haiku 4.5, and GPT-5.2 Chat all retain prices.

Closes#5797.

Built with GPT-5.6 Sol in the Codex harness via T3 Code.

Note

Fix cached Claude token overpricing by preserving provider-qualified keys in lookupRate

• Changes normalizeRateKey to trim and lowercase keys without stripping provider prefixes, so parseRateTable stores full qualified keys like anthropic/claude-3.5-sonnet
• lookupRate now matches the exact normalized key; it no longer falls back to the bare model name automatically, which previously caused a cheaper cache entry to match a more expensive non-cache rate
• Adds bare-name aliases (e.g. claude-3.5-sonnet) only when no canonical bare entry exists and all provider-qualified entries for that name share the same rate; conflicting rates leave the bare name unpriced
• Adds sameRate and bareModelName helpers; refactors normalizeModelName to delegate to normalizeRateKey then bareModelName
• Risk: any rate table relying on bare-name fallback for a model with provider-specific rates will now return unpriced instead of guessing; check rate table configs in apps/server/src/usage/usagePricing.ts and any callers of lookupRate that expect bare-name matching

^{Macroscope summarized 63bf1f8.}

---

Note

Medium Risk
Changes how usage costs are computed for models with multiple LiteLLM keys; incorrect aliasing could misprice or leave models unpriced, but behavior is covered by new tests and avoids silent overwrites.

Overview
Fixes cached Claude token overpricing when LiteLLM publishes both a canonical bare model entry (with cache-read discounts) and provider-qualified entries that share the same trailing name. Rates are no longer collapsed to bare names at insert time, so a later DeepInfra-style row cannot overwrite the canonical cache-read rate.

parseRateTable now stores each entry under its full normalized key (trim + lowercase). It may add a bare-name alias only when no canonical bare key exists and every qualified variant agrees on input, output, and cache rates; conflicting bare names stay unpriced (lookupRate returns null).

lookupRate matches the normalized transcript model string against the table directly (still blocking ambiguous family names like opus/sonnet). normalizeModelName keeps the same bare-name contract for callers that strip provider prefixes.

Adds usagePricing.test.ts for normalization, Fable vs DeepInfra ordering, unanimous aliasing, and ambiguous bare names. Scan-cache comments were updated to reference lookupRate instead of normalizeModelName.

^{Reviewed by Cursor Bugbot for commit 63bf1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1fbff086-af1d-4c82-a012-eecf26e7cb3a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Warning

Your free Security trial is over. An organization admin can activate Security or dismiss this notice.

---

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This focused production fix changes token-cost and cache-savings calculations in existing usage summaries. Because those calculations have direct billing and metering implications, the change warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[SunkenInTime]

The Usage page can overprice cached Claude tokens when LiteLLM publishes canonical and provider-qualified entries with the same bare model name. The parser stripped provider prefixes before inserting rates, so a later entry without the Anthropic cache discount could overwrite the canonical rate and price cache reads as ordinary input.

This builds on the exact-key approach in #8529 by @none23. @mackinleysmith identified one important edge case during that review: multiple qualified entries are safe to alias when their effective rates are identical. Treating those entries as ambiguous leaves models such as Claude Opus 4.6 and Claude Haiku 4.5 unpriced.

What changed

• Preserve each LiteLLM rate under its full normalized key.
• Keep canonical bare entries authoritative.
• Add a bare alias when every qualified entry has the same effective rate.
• Leave genuinely conflicting bare names unpriced instead of guessing.

Verification

• vp test run apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageAggregation.test.ts apps/server/src/usage/usageScanCache.test.ts
• vp lint apps/server/src/usage/usagePricing.ts apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageScanCache.ts apps/server/src/usage/usageScanCache.test.ts
• vp run --filter t3 typecheck
• Replayed the parser against the current LiteLLM table and confirmed Claude Opus 5, Claude Fable 5, Claude Opus 4.6, Claude Haiku 4.5, and GPT-5.2 Chat all retain prices.

Closes#5797.

Built with GPT-5.6 Sol in the Codex harness via T3 Code.

Note

Fix cached Claude token overpricing by preserving provider-qualified keys in lookupRate

• Changes normalizeRateKey to trim and lowercase keys without stripping provider prefixes, so parseRateTable stores full qualified keys like anthropic/claude-3.5-sonnet
• lookupRate now matches the exact normalized key; it no longer falls back to the bare model name automatically, which previously caused a cheaper cache entry to match a more expensive non-cache rate
• Adds bare-name aliases (e.g. claude-3.5-sonnet) only when no canonical bare entry exists and all provider-qualified entries for that name share the same rate; conflicting rates leave the bare name unpriced
• Adds sameRate and bareModelName helpers; refactors normalizeModelName to delegate to normalizeRateKey then bareModelName
• Risk: any rate table relying on bare-name fallback for a model with provider-specific rates will now return unpriced instead of guessing; check rate table configs in apps/server/src/usage/usagePricing.ts and any callers of lookupRate that expect bare-name matching

^{Macroscope summarized 63bf1f8.}

---

Note

Medium Risk
Changes how usage costs are computed for models with multiple LiteLLM keys; incorrect aliasing could misprice or leave models unpriced, but behavior is covered by new tests and avoids silent overwrites.

Overview
Fixes cached Claude token overpricing when LiteLLM publishes both a canonical bare model entry (with cache-read discounts) and provider-qualified entries that share the same trailing name. Rates are no longer collapsed to bare names at insert time, so a later DeepInfra-style row cannot overwrite the canonical cache-read rate.

parseRateTable now stores each entry under its full normalized key (trim + lowercase). It may add a bare-name alias only when no canonical bare key exists and every qualified variant agrees on input, output, and cache rates; conflicting bare names stay unpriced (lookupRate returns null).

lookupRate matches the normalized transcript model string against the table directly (still blocking ambiguous family names like opus/sonnet). normalizeModelName keeps the same bare-name contract for callers that strip provider prefixes.

Adds usagePricing.test.ts for normalization, Fable vs DeepInfra ordering, unanimous aliasing, and ambiguous bare names. Scan-cache comments were updated to reference lookupRate instead of normalizeModelName.

^{Reviewed by Cursor Bugbot for commit 63bf1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1fbff086-af1d-4c82-a012-eecf26e7cb3a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Warning

Your free Security trial is over. An organization admin can activate Security or dismiss this notice.

---

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This focused production fix changes token-cost and cache-savings calculations in existing usage summaries. Because those calculations have direct billing and metering implications, the change warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[SunkenInTime]

The Usage page can overprice cached Claude tokens when LiteLLM publishes canonical and provider-qualified entries with the same bare model name. The parser stripped provider prefixes before inserting rates, so a later entry without the Anthropic cache discount could overwrite the canonical rate and price cache reads as ordinary input.

This builds on the exact-key approach in #8529 by @none23. @mackinleysmith identified one important edge case during that review: multiple qualified entries are safe to alias when their effective rates are identical. Treating those entries as ambiguous leaves models such as Claude Opus 4.6 and Claude Haiku 4.5 unpriced.

What changed

• Preserve each LiteLLM rate under its full normalized key.
• Keep canonical bare entries authoritative.
• Add a bare alias when every qualified entry has the same effective rate.
• Leave genuinely conflicting bare names unpriced instead of guessing.

Verification

• vp test run apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageAggregation.test.ts apps/server/src/usage/usageScanCache.test.ts
• vp lint apps/server/src/usage/usagePricing.ts apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageScanCache.ts apps/server/src/usage/usageScanCache.test.ts
• vp run --filter t3 typecheck
• Replayed the parser against the current LiteLLM table and confirmed Claude Opus 5, Claude Fable 5, Claude Opus 4.6, Claude Haiku 4.5, and GPT-5.2 Chat all retain prices.

Closes#5797.

Built with GPT-5.6 Sol in the Codex harness via T3 Code.

Note

Fix cached Claude token overpricing by preserving provider-qualified keys in lookupRate

• Changes normalizeRateKey to trim and lowercase keys without stripping provider prefixes, so parseRateTable stores full qualified keys like anthropic/claude-3.5-sonnet
• lookupRate now matches the exact normalized key; it no longer falls back to the bare model name automatically, which previously caused a cheaper cache entry to match a more expensive non-cache rate
• Adds bare-name aliases (e.g. claude-3.5-sonnet) only when no canonical bare entry exists and all provider-qualified entries for that name share the same rate; conflicting rates leave the bare name unpriced
• Adds sameRate and bareModelName helpers; refactors normalizeModelName to delegate to normalizeRateKey then bareModelName
• Risk: any rate table relying on bare-name fallback for a model with provider-specific rates will now return unpriced instead of guessing; check rate table configs in apps/server/src/usage/usagePricing.ts and any callers of lookupRate that expect bare-name matching

^{Macroscope summarized 63bf1f8.}

---

Note

Medium Risk
Changes how usage costs are computed for models with multiple LiteLLM keys; incorrect aliasing could misprice or leave models unpriced, but behavior is covered by new tests and avoids silent overwrites.

Overview
Fixes cached Claude token overpricing when LiteLLM publishes both a canonical bare model entry (with cache-read discounts) and provider-qualified entries that share the same trailing name. Rates are no longer collapsed to bare names at insert time, so a later DeepInfra-style row cannot overwrite the canonical cache-read rate.

parseRateTable now stores each entry under its full normalized key (trim + lowercase). It may add a bare-name alias only when no canonical bare key exists and every qualified variant agrees on input, output, and cache rates; conflicting bare names stay unpriced (lookupRate returns null).

lookupRate matches the normalized transcript model string against the table directly (still blocking ambiguous family names like opus/sonnet). normalizeModelName keeps the same bare-name contract for callers that strip provider prefixes.

Adds usagePricing.test.ts for normalization, Fable vs DeepInfra ordering, unanimous aliasing, and ambiguous bare names. Scan-cache comments were updated to reference lookupRate instead of normalizeModelName.

^{Reviewed by Cursor Bugbot for commit 63bf1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1fbff086-af1d-4c82-a012-eecf26e7cb3a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Warning

Your free Security trial is over. An organization admin can activate Security or dismiss this notice.

---

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This focused production fix changes token-cost and cache-savings calculations in existing usage summaries. Because those calculations have direct billing and metering implications, the change warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[SunkenInTime]

The Usage page can overprice cached Claude tokens when LiteLLM publishes canonical and provider-qualified entries with the same bare model name. The parser stripped provider prefixes before inserting rates, so a later entry without the Anthropic cache discount could overwrite the canonical rate and price cache reads as ordinary input.

This builds on the exact-key approach in #8529 by @none23. @mackinleysmith identified one important edge case during that review: multiple qualified entries are safe to alias when their effective rates are identical. Treating those entries as ambiguous leaves models such as Claude Opus 4.6 and Claude Haiku 4.5 unpriced.

What changed

• Preserve each LiteLLM rate under its full normalized key.
• Keep canonical bare entries authoritative.
• Add a bare alias when every qualified entry has the same effective rate.
• Leave genuinely conflicting bare names unpriced instead of guessing.

Verification

• vp test run apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageAggregation.test.ts apps/server/src/usage/usageScanCache.test.ts
• vp lint apps/server/src/usage/usagePricing.ts apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageScanCache.ts apps/server/src/usage/usageScanCache.test.ts
• vp run --filter t3 typecheck
• Replayed the parser against the current LiteLLM table and confirmed Claude Opus 5, Claude Fable 5, Claude Opus 4.6, Claude Haiku 4.5, and GPT-5.2 Chat all retain prices.

Closes#5797.

Built with GPT-5.6 Sol in the Codex harness via T3 Code.

Note

Fix cached Claude token overpricing by preserving provider-qualified keys in lookupRate

• Changes normalizeRateKey to trim and lowercase keys without stripping provider prefixes, so parseRateTable stores full qualified keys like anthropic/claude-3.5-sonnet
• lookupRate now matches the exact normalized key; it no longer falls back to the bare model name automatically, which previously caused a cheaper cache entry to match a more expensive non-cache rate
• Adds bare-name aliases (e.g. claude-3.5-sonnet) only when no canonical bare entry exists and all provider-qualified entries for that name share the same rate; conflicting rates leave the bare name unpriced
• Adds sameRate and bareModelName helpers; refactors normalizeModelName to delegate to normalizeRateKey then bareModelName
• Risk: any rate table relying on bare-name fallback for a model with provider-specific rates will now return unpriced instead of guessing; check rate table configs in apps/server/src/usage/usagePricing.ts and any callers of lookupRate that expect bare-name matching

^{Macroscope summarized 63bf1f8.}

---

Note

Medium Risk
Changes how usage costs are computed for models with multiple LiteLLM keys; incorrect aliasing could misprice or leave models unpriced, but behavior is covered by new tests and avoids silent overwrites.

Overview
Fixes cached Claude token overpricing when LiteLLM publishes both a canonical bare model entry (with cache-read discounts) and provider-qualified entries that share the same trailing name. Rates are no longer collapsed to bare names at insert time, so a later DeepInfra-style row cannot overwrite the canonical cache-read rate.

parseRateTable now stores each entry under its full normalized key (trim + lowercase). It may add a bare-name alias only when no canonical bare key exists and every qualified variant agrees on input, output, and cache rates; conflicting bare names stay unpriced (lookupRate returns null).

lookupRate matches the normalized transcript model string against the table directly (still blocking ambiguous family names like opus/sonnet). normalizeModelName keeps the same bare-name contract for callers that strip provider prefixes.

Adds usagePricing.test.ts for normalization, Fable vs DeepInfra ordering, unanimous aliasing, and ambiguous bare names. Scan-cache comments were updated to reference lookupRate instead of normalizeModelName.

^{Reviewed by Cursor Bugbot for commit 63bf1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1fbff086-af1d-4c82-a012-eecf26e7cb3a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Warning

Your free Security trial is over. An organization admin can activate Security or dismiss this notice.

---

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This focused production fix changes token-cost and cache-savings calculations in existing usage summaries. Because those calculations have direct billing and metering implications, the change warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[SunkenInTime]

The Usage page can overprice cached Claude tokens when LiteLLM publishes canonical and provider-qualified entries with the same bare model name. The parser stripped provider prefixes before inserting rates, so a later entry without the Anthropic cache discount could overwrite the canonical rate and price cache reads as ordinary input.

This builds on the exact-key approach in #8529 by @none23. @mackinleysmith identified one important edge case during that review: multiple qualified entries are safe to alias when their effective rates are identical. Treating those entries as ambiguous leaves models such as Claude Opus 4.6 and Claude Haiku 4.5 unpriced.

What changed

• Preserve each LiteLLM rate under its full normalized key.
• Keep canonical bare entries authoritative.
• Add a bare alias when every qualified entry has the same effective rate.
• Leave genuinely conflicting bare names unpriced instead of guessing.

Verification

• vp test run apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageAggregation.test.ts apps/server/src/usage/usageScanCache.test.ts
• vp lint apps/server/src/usage/usagePricing.ts apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageScanCache.ts apps/server/src/usage/usageScanCache.test.ts
• vp run --filter t3 typecheck
• Replayed the parser against the current LiteLLM table and confirmed Claude Opus 5, Claude Fable 5, Claude Opus 4.6, Claude Haiku 4.5, and GPT-5.2 Chat all retain prices.

Closes#5797.

Built with GPT-5.6 Sol in the Codex harness via T3 Code.

Note

Fix cached Claude token overpricing by preserving provider-qualified keys in lookupRate

• Changes normalizeRateKey to trim and lowercase keys without stripping provider prefixes, so parseRateTable stores full qualified keys like anthropic/claude-3.5-sonnet
• lookupRate now matches the exact normalized key; it no longer falls back to the bare model name automatically, which previously caused a cheaper cache entry to match a more expensive non-cache rate
• Adds bare-name aliases (e.g. claude-3.5-sonnet) only when no canonical bare entry exists and all provider-qualified entries for that name share the same rate; conflicting rates leave the bare name unpriced
• Adds sameRate and bareModelName helpers; refactors normalizeModelName to delegate to normalizeRateKey then bareModelName
• Risk: any rate table relying on bare-name fallback for a model with provider-specific rates will now return unpriced instead of guessing; check rate table configs in apps/server/src/usage/usagePricing.ts and any callers of lookupRate that expect bare-name matching

^{Macroscope summarized 63bf1f8.}

---

Note

Medium Risk
Changes how usage costs are computed for models with multiple LiteLLM keys; incorrect aliasing could misprice or leave models unpriced, but behavior is covered by new tests and avoids silent overwrites.

Overview
Fixes cached Claude token overpricing when LiteLLM publishes both a canonical bare model entry (with cache-read discounts) and provider-qualified entries that share the same trailing name. Rates are no longer collapsed to bare names at insert time, so a later DeepInfra-style row cannot overwrite the canonical cache-read rate.

parseRateTable now stores each entry under its full normalized key (trim + lowercase). It may add a bare-name alias only when no canonical bare key exists and every qualified variant agrees on input, output, and cache rates; conflicting bare names stay unpriced (lookupRate returns null).

lookupRate matches the normalized transcript model string against the table directly (still blocking ambiguous family names like opus/sonnet). normalizeModelName keeps the same bare-name contract for callers that strip provider prefixes.

Adds usagePricing.test.ts for normalization, Fable vs DeepInfra ordering, unanimous aliasing, and ambiguous bare names. Scan-cache comments were updated to reference lookupRate instead of normalizeModelName.

^{Reviewed by Cursor Bugbot for commit 63bf1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1fbff086-af1d-4c82-a012-eecf26e7cb3a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Warning

Your free Security trial is over. An organization admin can activate Security or dismiss this notice.

---

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This focused production fix changes token-cost and cache-savings calculations in existing usage summaries. Because those calculations have direct billing and metering implications, the change warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[SunkenInTime]

The Usage page can overprice cached Claude tokens when LiteLLM publishes canonical and provider-qualified entries with the same bare model name. The parser stripped provider prefixes before inserting rates, so a later entry without the Anthropic cache discount could overwrite the canonical rate and price cache reads as ordinary input.

This builds on the exact-key approach in #8529 by @none23. @mackinleysmith identified one important edge case during that review: multiple qualified entries are safe to alias when their effective rates are identical. Treating those entries as ambiguous leaves models such as Claude Opus 4.6 and Claude Haiku 4.5 unpriced.

What changed

• Preserve each LiteLLM rate under its full normalized key.
• Keep canonical bare entries authoritative.
• Add a bare alias when every qualified entry has the same effective rate.
• Leave genuinely conflicting bare names unpriced instead of guessing.

Verification

• vp test run apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageAggregation.test.ts apps/server/src/usage/usageScanCache.test.ts
• vp lint apps/server/src/usage/usagePricing.ts apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageScanCache.ts apps/server/src/usage/usageScanCache.test.ts
• vp run --filter t3 typecheck
• Replayed the parser against the current LiteLLM table and confirmed Claude Opus 5, Claude Fable 5, Claude Opus 4.6, Claude Haiku 4.5, and GPT-5.2 Chat all retain prices.

Closes#5797.

Built with GPT-5.6 Sol in the Codex harness via T3 Code.

Note

Fix cached Claude token overpricing by preserving provider-qualified keys in lookupRate

• Changes normalizeRateKey to trim and lowercase keys without stripping provider prefixes, so parseRateTable stores full qualified keys like anthropic/claude-3.5-sonnet
• lookupRate now matches the exact normalized key; it no longer falls back to the bare model name automatically, which previously caused a cheaper cache entry to match a more expensive non-cache rate
• Adds bare-name aliases (e.g. claude-3.5-sonnet) only when no canonical bare entry exists and all provider-qualified entries for that name share the same rate; conflicting rates leave the bare name unpriced
• Adds sameRate and bareModelName helpers; refactors normalizeModelName to delegate to normalizeRateKey then bareModelName
• Risk: any rate table relying on bare-name fallback for a model with provider-specific rates will now return unpriced instead of guessing; check rate table configs in apps/server/src/usage/usagePricing.ts and any callers of lookupRate that expect bare-name matching

^{Macroscope summarized 63bf1f8.}

---

Note

Medium Risk
Changes how usage costs are computed for models with multiple LiteLLM keys; incorrect aliasing could misprice or leave models unpriced, but behavior is covered by new tests and avoids silent overwrites.

Overview
Fixes cached Claude token overpricing when LiteLLM publishes both a canonical bare model entry (with cache-read discounts) and provider-qualified entries that share the same trailing name. Rates are no longer collapsed to bare names at insert time, so a later DeepInfra-style row cannot overwrite the canonical cache-read rate.

parseRateTable now stores each entry under its full normalized key (trim + lowercase). It may add a bare-name alias only when no canonical bare key exists and every qualified variant agrees on input, output, and cache rates; conflicting bare names stay unpriced (lookupRate returns null).

lookupRate matches the normalized transcript model string against the table directly (still blocking ambiguous family names like opus/sonnet). normalizeModelName keeps the same bare-name contract for callers that strip provider prefixes.

Adds usagePricing.test.ts for normalization, Fable vs DeepInfra ordering, unanimous aliasing, and ambiguous bare names. Scan-cache comments were updated to reference lookupRate instead of normalizeModelName.

^{Reviewed by Cursor Bugbot for commit 63bf1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1fbff086-af1d-4c82-a012-eecf26e7cb3a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Warning

Your free Security trial is over. An organization admin can activate Security or dismiss this notice.

---

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This focused production fix changes token-cost and cache-savings calculations in existing usage summaries. Because those calculations have direct billing and metering implications, the change warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[SunkenInTime]

The Usage page can overprice cached Claude tokens when LiteLLM publishes canonical and provider-qualified entries with the same bare model name. The parser stripped provider prefixes before inserting rates, so a later entry without the Anthropic cache discount could overwrite the canonical rate and price cache reads as ordinary input.

This builds on the exact-key approach in #8529 by @none23. @mackinleysmith identified one important edge case during that review: multiple qualified entries are safe to alias when their effective rates are identical. Treating those entries as ambiguous leaves models such as Claude Opus 4.6 and Claude Haiku 4.5 unpriced.

What changed

• Preserve each LiteLLM rate under its full normalized key.
• Keep canonical bare entries authoritative.
• Add a bare alias when every qualified entry has the same effective rate.
• Leave genuinely conflicting bare names unpriced instead of guessing.

Verification

• vp test run apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageAggregation.test.ts apps/server/src/usage/usageScanCache.test.ts
• vp lint apps/server/src/usage/usagePricing.ts apps/server/src/usage/usagePricing.test.ts apps/server/src/usage/usageScanCache.ts apps/server/src/usage/usageScanCache.test.ts
• vp run --filter t3 typecheck
• Replayed the parser against the current LiteLLM table and confirmed Claude Opus 5, Claude Fable 5, Claude Opus 4.6, Claude Haiku 4.5, and GPT-5.2 Chat all retain prices.

Closes#5797.

Built with GPT-5.6 Sol in the Codex harness via T3 Code.

Note

Fix cached Claude token overpricing by preserving provider-qualified keys in lookupRate

• Changes normalizeRateKey to trim and lowercase keys without stripping provider prefixes, so parseRateTable stores full qualified keys like anthropic/claude-3.5-sonnet
• lookupRate now matches the exact normalized key; it no longer falls back to the bare model name automatically, which previously caused a cheaper cache entry to match a more expensive non-cache rate
• Adds bare-name aliases (e.g. claude-3.5-sonnet) only when no canonical bare entry exists and all provider-qualified entries for that name share the same rate; conflicting rates leave the bare name unpriced
• Adds sameRate and bareModelName helpers; refactors normalizeModelName to delegate to normalizeRateKey then bareModelName
• Risk: any rate table relying on bare-name fallback for a model with provider-specific rates will now return unpriced instead of guessing; check rate table configs in apps/server/src/usage/usagePricing.ts and any callers of lookupRate that expect bare-name matching

^{Macroscope summarized 63bf1f8.}

---

Note

Medium Risk
Changes how usage costs are computed for models with multiple LiteLLM keys; incorrect aliasing could misprice or leave models unpriced, but behavior is covered by new tests and avoids silent overwrites.

Overview
Fixes cached Claude token overpricing when LiteLLM publishes both a canonical bare model entry (with cache-read discounts) and provider-qualified entries that share the same trailing name. Rates are no longer collapsed to bare names at insert time, so a later DeepInfra-style row cannot overwrite the canonical cache-read rate.

parseRateTable now stores each entry under its full normalized key (trim + lowercase). It may add a bare-name alias only when no canonical bare key exists and every qualified variant agrees on input, output, and cache rates; conflicting bare names stay unpriced (lookupRate returns null).

lookupRate matches the normalized transcript model string against the table directly (still blocking ambiguous family names like opus/sonnet). normalizeModelName keeps the same bare-name contract for callers that strip provider prefixes.

Adds usagePricing.test.ts for normalization, Fable vs DeepInfra ordering, unanimous aliasing, and ambiguous bare names. Scan-cache comments were updated to reference lookupRate instead of normalizeModelName.

^{Reviewed by Cursor Bugbot for commit 63bf1f8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1fbff086-af1d-4c82-a012-eecf26e7cb3a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Warning

Your free Security trial is over. An organization admin can activate Security or dismiss this notice.

---

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

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This focused production fix changes token-cost and cache-savings calculations in existing usage summaries. Because those calculations have direct billing and metering implications, the change warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}