🪙 feat: Add Token Pricing for Devstral, Voxtral, Holo2, and Mistral Medium

[JumpLink]

Summary

Adds token pricing for four open-weight models that are currently missing from tokenValues in packages/data-schemas/src/methods/tx.ts: Devstral, Voxtral, Holo2, and Mistral Medium.

Because they are absent, getValueKey/getMultiplier fall through to defaultRate (6 = $6 per 1M tokens) for balance checks and transaction accounting whenever these models are used on a custom OpenAI-compatible endpoint. For inexpensive open-weight models this massively over-charges credits. These models are, for example, served on Scaleway's Generative APIs (OpenAI-compatible), where they are an order of magnitude cheaper than the $6 default.

Changes

Added representative USD-per-1M-token rates:

key	prompt	completion	source
devstral	0.4	2.0	Mistral API pricing (Devstral 2)
mistral-medium	1.5	7.5	Mistral API pricing (Mistral Medium 3.5)
voxtral / voxtral-small	0.1	0.4	Mistral API pricing (text)
holo2	0.3	0.7	Scaleway Generative APIs public pricing

Generic keys are used intentionally so versioned ids match via the existing longest-substring lookup in findMatchingPattern, e.g.:

• devstral-2-123b-instruct-2512 → devstral
• mistral-medium-3.5-128b → mistral-medium
• voxtral-small-24b-2507 → voxtral-small
• holo2-30b-a3b → holo2

None of these keys collide with existing entries.

Notes

• Sources: Mistral pricing (https://mistral.ai/pricing) for Devstral/Voxtral/Mistral Medium; Scaleway Generative APIs pricing for Holo2 (no first-party hosted price published). Exact figures are best-effort representative rates - happy to adjust to whatever convention you prefer.
• No behaviour change beyond pricing; purely additive entries in the existing map style.

[danny-avila]

Thanks @JumpLink, and the diagnosis is correct: absent keys fall through to defaultRate ($6/1M), which over-charges cheap open-weight models.

I will accept the PR but worth flagging before growing the global table even more in the future: this is already doable per-endpoint via tokenConfig, no code change needed. It takes precedence over the built-in map (getMultiplier checks the endpoint config first), and it's scoped to your endpoint, so it reflects the price you actually pay instead of a global guess applied to every provider serving that model:

endpoints:
  custom:
    - name: 'Scaleway'
      apiKey: '${SCALEWAY_API_KEY}'
      baseURL: 'https://api.scaleway.ai/v1'
      models:
        default: ['devstral-2-123b-instruct-2512', 'mistral-medium-3.5-128b', 'voxtral-small-24b-2507', 'holo2-30b-a3b']
        fetch: true
      # USD per 1M tokens; `context` is the model's context window (tokens)
      tokenConfig:
        devstral-2-123b-instruct-2512: { prompt: 0.4, completion: 2.0, context: 256000 }
        mistral-medium-3.5-128b:       { prompt: 1.5, completion: 7.5, context: 128000 }
        voxtral-small-24b-2507:        { prompt: 0.1, completion: 0.4, context: 32000 }
        holo2-30b-a3b:                 { prompt: 0.3, completion: 0.7, context: 32000 }
        # cacheRead / cacheWrite are also supported (optional)

Keys must match the exact model ids your endpoint serves (this path is exact-match, not substring); prompt, completion, and context are required.

That keeps provider-specific rates out of the shared default table, since pricing varies by provider. If you'd still prefer representative defaults baked in for the out-of-box experience, happy to take it. Just note voxtral-small duplicates voxtral (same rate, and voxtral already matches the versioned id), so that line can be dropped.

https://www.librechat.ai/docs/configuration/token_usage