fix(plugin): map real pricing from /v1/model/info into cost field

CHANGELOG.md

@@ -27,6 +27,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Fixed
+- **Discovered models now carry real pricing instead of always showing
+  `$0.00`.** `toConfigModel()` never read `input_cost_per_token` /
+  `output_cost_per_token` from `/v1/model/info`, so every model —
+  including ones LiteLLM has genuine pricing for — landed in
+  `opencode.json` without a `cost` field, and OpenCode's own default
+  (`0`) made every call look free regardless of actual spend. The
+  plugin now maps those fields into OpenCode's `cost` block (USD per
+  token → USD per million tokens, OpenCode's convention), verified live
+  against `x-litellm-response-cost` on a real completion. Models
+  LiteLLM has no price anchor for (e.g. rerank) are left without a
+  `cost` field, same as before — this fix reports real prices, it
+  doesn't invent ones.
 - **Embedding / image / audio models no longer appear in the OpenCode
   model picker.** The non-chat filter in `toConfigModel()` was a dead
   code path that returned the model entry either way, so models like

README.md

@@ -74,6 +74,7 @@ opencode
 | 📡 **Dynamic discovery** | Queries `/v1/models` so your OpenCode model picker always reflects your live `model_list`. |
 | 🏷️ **Smart formatting** | Turns `anthropic/claude-3-5-sonnet` into `Claude 3 5 Sonnet` in the picker — handles versions, sizes, quantizations, and brand-cased names like `gpt-4o`. |
 | 🧠 **Modality-aware** | Enriches `/v1/models` entries with `/v1/model/info` (`mode`, token limits, capability flags) and hides embedding / image / audio models from the picker. |
+| 💵 **Real pricing** | Maps `input_cost_per_token` / `output_cost_per_token` from `/v1/model/info` into OpenCode's `cost` field, so the picker and `/cost` show what the proxy actually bills instead of `$0.00`. Models LiteLLM has no price for are left unpriced, not falsely marked free. |
 | 🧪 **Reasoning-aware routing** | Auto-routes `gpt-5*` / `o1`/`o3`/`o4*` models through a sibling `litellm-responses` provider that uses `/v1/responses`, so tools + `reasoning_effort` actually work. Override per model via `responsesApiModels` / `chatApiModels`. |
 | 🏢 **Provider extraction** | Pulls `litellm_provider` (or the `provider/model` prefix) into `organizationOwner` so models group correctly in the UI. |
 | 🔐 **Auth-aware** | Honours `LITELLM_API_KEY` / `LITELLM_MASTER_KEY` env vars or `provider.litellm.options.apiKey`. |
@@ -292,7 +293,7 @@ sequenceDiagram
 1. On OpenCode startup the `config` lifecycle hook fires.
 2. If `provider.litellm` exists, its `baseURL` is used. Otherwise common ports are probed.
 3. A health check (`GET /v1/models`) verifies the proxy is reachable and authorized.
-4. Models from the response are enriched with `/v1/model/info` metadata (`mode`, token limits, capability flags — `/v1/models` omits these for database-defined models) and converted into OpenCode model entries keyed by `id`, with formatted `name`, `organizationOwner`, and inferred `modalities`. Non-chat models (embedding / image / audio) are excluded from the picker.
+4. Models from the response are enriched with `/v1/model/info` metadata (`mode`, token limits, capability flags, and per-token pricing — `/v1/models` omits these for database-defined models) and converted into OpenCode model entries keyed by `id`, with formatted `name`, `organizationOwner`, inferred `modalities`, and `cost` (USD/1M tokens, converted from LiteLLM's USD/token). Non-chat models (embedding / image / audio) are excluded from the picker.
 5. Each model is bucketed by transport — reasoning-tier models (`gpt-5*`, `o1`/`o3`/`o4*`, or anything with `mode === 'responses'`) go into the `litellm-responses` provider; everything else goes into `litellm`. Per-model overrides via `responsesApiModels` / `chatApiModels` win.
 6. Discovered models are merged on top of any user-defined ones — never overwriting them. A model is skipped if its key already exists under **either** provider.
 7. The whole flow is wrapped in a `Promise.race` against a 20 s timeout so a slow proxy never blocks boot.

src/plugin/index.ts

@@ -77,9 +77,19 @@ function enrichModel(model: LiteLLMModel, info: LiteLLMModelInfo): LiteLLMModel
     supports_reasoning: model.supports_reasoning ?? info.supports_reasoning,
     supports_pdf_input: model.supports_pdf_input ?? info.supports_pdf_input,
     supports_audio_input: model.supports_audio_input ?? info.supports_audio_input,
+    input_cost_per_token: model.input_cost_per_token ?? info.input_cost_per_token,
+    output_cost_per_token: model.output_cost_per_token ?? info.output_cost_per_token,
   }
 }
 
+/**
+ * OpenCode's `cost` config field is USD per **million** tokens (the
+ * models.dev convention); LiteLLM's `/v1/model/info` reports USD per
+ * single token. `1e6` bridges the two — verified against a live
+ * `x-litellm-response-cost` header, not just the unit names.
+ */
+const USD_PER_TOKEN_TO_PER_MILLION = 1_000_000
+
 /**
  * Convert a discovered LiteLLM model into an OpenCode config-level
  * model entry (the shape used in `provider.*.models` inside
@@ -110,6 +120,16 @@ function toConfigModel(model: LiteLLMModel): Record<string, unknown> | null {
   if (model.supports_vision) {
     entry.attachment = true
   }
+  // Only emit `cost` when LiteLLM actually reported a price. Omitting
+  // it (rather than defaulting to 0) lets OpenCode/models.dev fall back
+  // to their own default instead of us asserting "this model is free"
+  // for something LiteLLM simply has no price anchor for (e.g. rerank).
+  if (model.input_cost_per_token != null || model.output_cost_per_token != null) {
+    entry.cost = {
+      input: (model.input_cost_per_token ?? 0) * USD_PER_TOKEN_TO_PER_MILLION,
+      output: (model.output_cost_per_token ?? 0) * USD_PER_TOKEN_TO_PER_MILLION,
+    }
+  }
   const input: Array<'text' | 'image' | 'pdf' | 'audio'> = ['text']
   if (model.supports_vision) input.push('image')
   if (model.supports_pdf_input) input.push('pdf')

src/types/index.ts

@@ -32,6 +32,14 @@ export interface LiteLLMModel {
   supports_reasoning?: boolean
   supports_pdf_input?: boolean
   supports_audio_input?: boolean
+  /**
+   * USD price per input/output token, reliably available via
+   * `/v1/model/info` (`/v1/models` omits pricing). Absent for models
+   * LiteLLM has no price anchor for (e.g. rerank) — treat as "unknown",
+   * not "free".
+   */
+  input_cost_per_token?: number
+  output_cost_per_token?: number
 }
 
 export interface LiteLLMModelsResponse {
@@ -58,6 +66,8 @@ export interface LiteLLMModelInfo {
   supports_reasoning?: boolean
   supports_pdf_input?: boolean
   supports_audio_input?: boolean
+  input_cost_per_token?: number
+  output_cost_per_token?: number
 }
 
 /** A single entry returned by LiteLLM's `/v1/model/info` endpoint. */