Show cache read and write prices for OpenRouter inference providers

[chrarnoldus]

Related GitHub Issue

Upstreaming some changes from Kilo Code: Kilo-Org/kilocode#1893, Kilo-Org/kilocode#1940

Description

Fixes OpenRouters models that don't list a cache write price, but do support caching showing as "Does not support prompt caching" (e.g. GPT-5).

Fixes all models showing as "Supports prompt caching" when a specific inference provider is selected, even when they don't support prompt caching.

For specific inference providers, show the cache read and write prices from the OpenRouter metadata.

Use the unique tag for inference providers, rather than the non-unique provider_name.

Test Procedure

Go into the API Providers settings, select OpenRouter and verify all displayed info is correct:

• "Supports prompt caching" label
• Cache read/write prices
• Names of inference providers are unique

Pre-Submission Checklist

• Issue Linked: This PR is linked to an approved GitHub Issue (see "Related GitHub Issue" above).
• Scope: My changes are focused on the linked issue (one major feature/fix per PR).
• Self-Review: I have performed a thorough self-review of my code.
• Testing: New and/or updated tests have been added to cover my changes (if applicable).
• Documentation Impact: I have considered if my changes require documentation updates (see "Documentation Updates" section below).
• Contribution Guidelines: I have read and agree to the Contributor Guidelines.

Screenshots / Videos

Before

GPT-5 shown as not supporting prompt caching even though it does:

Qwen3 Coder with DeepInfra inference provider showing it does support prompt caching even though it doesn't:

Inference provider list for Sonnet 4 being incomplete, because the names are not unique:

After

GPT-5 shown as supporting prompt caching:

Qwen3 Coder with DeepInfra inference provider shown as not supporting prompt caching:

GLM 4.5 metadata without specific inference provider selected:

GLM 4.5 metadata with Z.AI selected as inference provider, which supports prompt caching:

Entire list of Sonnet 4 providers shown with unique tags:

Additional Notes

Some of the fields in the OpenRouter metadata are not documented. I have pinged our OpenRouter reps to update the docs.

Get in Touch

Christiaan in shared Slack

---

Important

Fixes cache support and pricing display for OpenRouter models, using unique tags for provider identification.

• Behavior:
  • Fixes incorrect cache support display for models without cache write price in openrouter.ts.
  • Corrects cache support display for specific inference providers in useOpenRouterModelProviders.ts.
  • Displays cache read/write prices from OpenRouter metadata in useOpenRouterModelProviders.ts.
• Schema and Parsing:
  • Adds tag field to openRouterModelEndpointSchema in openrouter.ts and useOpenRouterModelProviders.ts.
  • Updates parseOpenRouterModel to use tag for provider identification in openrouter.ts.
• Tests:
  • Updates tests in openrouter.spec.ts to reflect changes in cache support and provider identification.

^{This description was created by for 4733f15. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Thank you for your contribution! I've reviewed the changes and they look good overall. The implementation correctly addresses the cache support detection issues and improves provider identification. I have a few minor suggestions for improvement.

[roomote]

Could we make this comment more descriptive? Consider:

Suggested change

	const supportsPromptCache = typeof cacheReadsPrice !== "undefined" // some models support caching but don't charge a cacheWritesPrice, e.g. GPT-5
	const supportsPromptCache = typeof cacheReadsPrice !== "undefined" // OpenRouter reports cache support based on read price only, as some models support caching without charging for cache writes (e.g. GPT-5)

[roomote]

Would it be helpful to expand this comment to explain the cache_control limitation more clearly? Something like:

Suggested change

	.filter(([id, _]) => id.startsWith("anthropic/claude") || id.startsWith("google/gemini")) // only these support cache_control breakpoints (https://openrouter.ai/docs/features/prompt-caching)
	.filter(([id, _]) => id.startsWith("anthropic/claude") || id.startsWith("google/gemini")) // Only Anthropic Claude and Google Gemini models support cache_control breakpoints for explicit cache management (https://openrouter.ai/docs/features/prompt-caching)

[roomote]

For consistency with the backend implementation, consider adding a comment explaining the fallback logic:

Suggested change

	const providerName = endpoint.tag ?? endpoint.name
	const providerName = endpoint.tag ?? endpoint.name // Use unique tag when available, fallback to name for backward compatibility

[roomote]

Is this fallback to endpoint.provider_name intentional for backward compatibility? If the OpenRouter API doesn't always provide tags, this is a good defensive approach.